Patent Application
20140011861
This document relates to materials and methods for determining gene expression in cells, and for diagnosing prostate cancer and assessing prognosis of prostate cancer patients.
Prostate cancer is the most common malignancy in men and is the cause of considerable morbidity and mortality (Howe et al. (2001) J. Natl. Cancer Inst. 93:824-842). It may be useful to identify genes that could be reliable early diagnostic and prognostic markers and therapeutic targets for prostate cancer, as well as other diseases and disorders.
This document is based in part on the discovery that RNA expression changes can be identified that can distinguish normal prostate stroma from tumor-adjacent stroma in the absence of tumor cells, and that such expression changes can be used to signal the “presence of tumor.” A linear regression method for the identification of cell-type specific expression of RNA from array data of prostate tumor-enriched samples was previously developed and validated (see, U.S. Publication No. 20060292572 and Stuart et al. (2004) Proc. Natl. Acad. Sci. USA 101:615-620, both incorporated herein by reference in their entirety). As described herein, the approach was extended to evaluate differential expression data obtained from normal volunteer prostate biopsy samples with tumor-adjacent stroma. Over a thousand gene expression changes were observed. A subset of stroma-specific genes were used to derive a classifier of 131 probe sets that accurately identified tumor or nontumor status of a large number of independent test cases. These observations indicate that tumor-adjacent stroma exhibits a larger number of gene expression changes and that subset may be selected to reliably identify tumor in the absence of tumor cells. The classifier may be useful in the diagnosis of stroma-rich biopsies of clinical cases with equivocal pathology readings.
The present disclosure includes, inter alia, the following: (1) extensive cross-validation of RNA biomarkers for prostate cancer relapse, across multiple datasets; (2) a “bi-modal” method for generating classifiers and testing them on samples that have mixed tissue; and (3) two methods for identifying genes in “reactive-stroma” that can be used as markers for the presence of cancer even when the sample does not include tumor but instead has regions of reactive stroma, near tumor.
In one aspect, this document features an in vitro method for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring the level of expression for prostate cancer signature genes in the sample; (c) comparing the measured expression levels to reference expression levels for the prostate cancer signature genes; and (d) if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having prostate cancer, and if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as not having prostate cancer. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein. The method can include determining whether measured expression levels for ten or more prostate cancer signature genes are significantly greater or less than reference expression levels for the ten or more prostate cancer signature genes, and classifying the subject as having prostate cancer that is likely to relapse if the measured expression levels are significantly greater or less than the reference expression levels, or classifying the subject as having prostate cancer not likely to relapse if the measured expression levels are not significantly greater or less than the reference expression levels. The ten or more prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein. The method can include determining whether measured expression levels for twenty or more prostate cancer signature genes are significantly greater or less than reference expression levels for the twenty or more prostate cancer signature genes, and classifying the subject as having prostate cancer that is likely to relapse if the measured expression levels are significantly greater or less than the reference expression levels, or classifying the subject as having prostate cancer not likely to relapse if the measured expression levels are not significantly greater or less than the reference expression levels. The twenty or more prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein.
In another aspect, this document features a method for determining the prognosis of a subject diagnosed as having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring the level of expression for prostate cancer signature genes in the sample; (c) comparing the measured expression levels to reference expression levels for the prostate cancer signature genes; and (d) if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as having a relatively better prognosis than if the measured expression levels are significantly greater or less than the reference expression levels, or if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having a relatively worse prognosis than if the measured expression levels are not significantly greater or less than the reference expression levels. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in Table 8A or 8B herein.
In another aspect, this document features a method for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject, wherein the sample comprises prostate stromal cells; (b) measuring expression levels for one or more genes in the stromal cells, wherein the one or more genes are prostate cancer signature genes; (c) comparing the measured expression levels to reference expression levels for the one or more genes, wherein the reference expression levels are determined in stromal cells from non-cancerous prostate tissue; and (d) if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having prostate cancer, and if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as not having prostate cancer. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein.
In another aspect, this document features a method for determining a prognosis for a subject diagnosed as having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject, wherein the sample comprises prostate stromal cells; (b) measuring expression levels for one or more genes in the stromal cells, wherein the one or more genes are prostate cancer signature genes; (c) comparing the measured expression levels to reference expression levels for the one or more genes, wherein the reference expression levels are determined in stromal cells from non-cancerous prostate tissue; and (d) if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as having a relatively better prognosis than if the measured expression levels are significantly greater or less than the reference expression levels, or if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having a relatively worse prognosis than if the measured expression levels are not significantly greater or less than the reference expression levels. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein.
In still another aspect, this document features a method for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring expression levels for one or more prostate cell-type predictor genes in the sample; (c) determining the percentages of tissue types in the sample based on the measured expression levels; (d) measuring expression levels for one more prostate cancer signature genes in the sample; (e) determining a classifier based on the percentages of tissue types and the measured expression levels; and (f) if the classifier falls into a predetermined range of prostate cancer classifiers, identifying the subject as having prostate cancer, or if the classifier does not fall into the predetermined range, identifying the subject as not having prostate cancer. Steps (b) and (d) can be carried out simultaneously.
This document also features a method for determining a prognosis for a subject diagnosed with and treated for prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring expression levels for one or more prostate tissue predictor genes in the sample; (c) determining the percentages of tissue types in the sample based on the measured expression levels; (d) measuring expression levels for one more prostate cancer signature genes in the sample; (e) determining a classifier based on the percentages of tissue types and the measured expression levels; and (f) if the classifier falls into a predetermined range of prostate cancer relapse classifiers, identifying the subject as being likely to relapse, or if the classifier does not fall into the predetermined range, identifying the subject as not being likely to relapse. Steps (b) and (d) are carried out simultaneously.
In yet another aspect, this document features a method for identifying the proportion of two or more tissue types in a tissue sample, comprising: (a) using a set of other samples of known tissue proportions from a similar anatomical location as the tissue sample in an animal or plant, wherein at least two of the other samples do not contain the same relative content of each of the two or more cell types; (b) measuring overall levels of one or more gene expression or protein analytes in each of the other samples; (c) determining the regression relationship between the relative proportion of each tissue type and the measured overall levels of each gene expression or protein analyte in the other samples; (d) selecting one or more analytes that correlate with tissue proportions in the other samples; (e) measuring overall levels of one or more of the analytes in step (d) in the tissue sample; (f) matching the level of each analyte in the tissue sample with the level of the analyte in step (d) to determine the predicted proportion of each tissue type in the tissue sample; and (g) selecting among predicted tissue proportions for the tissue sample obtained in step (f) using either the median or average proportions of all the estimates. The tissue sample can contain cancer cells (e.g., prostate cancer cells).
In another aspect, this document features a method for comparing the levels of two or more analytes predicted by one or more methods to be associated with a change in a biological phenomenon in two sets of data each containing more than one measured sample, comprising: (a) selecting only analytes that are assayed in both sets of data; (b) ranking the analytes in each set of data using a comparative method such as the highest probability or lowest false discovery rate associated with the change in the biological phenomenon; (c) comparing a set of analytes in each ranked list in step (b) with each other, selecting those that occur in both lists, and determining the number of analytes that occur in both lists and show a change in level associated with the biological phenomenon that is in the same direction; and (d) calculating a concordance score based on the probability that the number of comparisons would show the observed number of change in the same direction, at random. In step (a), the length of each list can be varied to determine the maximum concordance score for the two ranked lists.
Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Although methods and materials similar or equivalent to those described herein can be used to practice the invention, suitable methods and materials are described below. All publications, patent applications, patents, and other references mentioned herein are incorporated by reference in their entirety. In case of conflict, the present specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and not intended to be limiting.
The details of one or more embodiments of the invention are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the invention will be apparent from the description and drawings, and from the claims.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as is commonly understood by one of skill in the art to which the invention(s) belong. All patents, patent applications, published applications and publications, GENBANK® sequences, websites and other published materials referred to throughout the entire disclosure herein, unless noted otherwise, are incorporated by reference in their entirety. In the event that there is a plurality of definitions for terms herein, those in this section prevail. Where reference is made to a URL or other such identifier or address, it understood that such identifiers particular information on the internet can change, equivalent information can be found by searching the internet. Reference thereto evidences the availability and public dissemination of such information.
Differential expression includes to both quantitative as well as qualitative differences in the extend of the genes' expression depending on differential development and/or tumor growth. Differentially expressed genes can represent marker genes, and/or target genes. The expression pattern of a differentially expressed gene disclosed herein can be utilized as part of a prognostic or diagnostic evaluation of a subject. The expression pattern of a differentially expressed gene can be used to identify the presence of a particular cell type in a sample. A differentially expressed gene disclosed herein can be used in methods for identifying reagents and compounds and uses of these reagents and compounds for the treatment of a subject as well as methods of treatment. The terms “biological activity,” “bioactivity,” “activity,” and “biological function” can be used interchangeably, and can refer to an effector or antigenic function that is directly or indirectly performed by a polypeptide (whether in its native or denatured conformation), or by any fragment thereof in vivo or in vitro. Biological activities include, without limitation, binding to polypeptides, binding to other proteins or molecules, enzymatic activity, signal transduction, activity as a DNA binding protein, as a transcription regulator, and ability to bind damaged DNA. A bioactivity can be modulated by directly affecting the subject polypeptide. Alternatively, a bioactivity can be altered by modulating the level of the polypeptide, such as by modulating expression of the corresponding gene.
The term “gene expression analyte” refers to a biological molecule whose presence or concentration can be detected and correlated with gene expression. For example, a gene expression analyte can be a mRNA of a particular gene, or a fragment thereof (including, e.g., by-products of mRNA splicing and nucleolytic cleavage fragments), a protein of a particular gene or a fragment thereof (including, e.g., post-translationally modified proteins or by-products therefrom, and proteolytic fragments), and other biological molecules such as a carbohydrate, lipid or small molecule, whose presence or absence corresponds to the expression of a particular gene.
A gene expression level is to the amount of biological macromolecule produced from a gene. For example, expression levels of a particular gene can refer to the amount of protein produced from that particular gene, or can refer to the amount of mRNA produced from that particular gene. Gene expression levels can refer to an absolute (e.g., molar or gram-quantity) levels or relative (e.g., the amount relative to a standard, reference, calibration, or to another gene expression level). Typically, gene expression levels used herein are relative expression levels. As used herein in regard to determining the relationship between cell content and expression levels, gene expression levels can be considered in terms of any manner of describing gene expression known in the art. For example, regression methods that consider gene expression levels can consider the measurement of the level of a gene expression analyte, or the level calculated or estimated according to the measurement of the level of a gene expression analyte.
A marker gene is a differentially expressed gene which expression pattern can serve as part of a phenotype-indicating method, such as a predictive method, prognostic or diagnostic method, or other cell-type distinguishing evaluation, or which, alternatively, can be used in methods for identifying compounds useful for the treatment or prevention of diseases or disorders, or for identifying compounds that modulate the activity of one or more gene products.
A phenotype indicated by methods provided herein can be a diagnostic indication, a prognostic indication, or an indication of the presence of a particular cell type in a subject. Diagnostic indications include indication of a disease or a disorder in the subject, such as presence of tumor or neoplastic disease, inflammatory disease, autoimmune disease, and any other diseases known in the art that can be identified according to the presence or absence of particular cells or by the gene expression of cells. In another embodiment, prognostic indications refers to the likely or expected outcome of a disease or disorder, including, but not limited to, the likelihood of survival of the subject, likelihood of relapse, aggressiveness of the disease or disorder, indolence of the disease or disorder, and likelihood of success of a particular treatment regimen.
The phrase “gene expression levels that correspond to levels of gene expression analytes” refers to the relationship between an analyte that indicates the expression of a gene, and the actual level of expression of the gene. Typically the level of a gene expression analyte is measured in experimental methods used to determine gene expression levels. As understood by one skilled in the art, the measured gene expression levels can represent gene expression at a variety of levels of detail (e.g., the absolute amount of a gene expressed, the relative amount of gene expressed, or an indication of increased or decreased levels of expression). The level of detail at which the levels of gene expression analytes can indicate levels of gene expression can be based on a variety of factors that include the number of controls used, the number of calibration experiments or reference levels determined, and other factors known in the art. In some methods provided herein, increase in the levels of a gene expression analyte can indicate increase in the levels of the gene expressed, and a decrease in the levels of a gene expression analyte can indicate decrease in the levels of the gene expressed.
A regression relationship between relative content of a cell type and measured overall levels of a gene expression analyte is a quantitative relationship between cell type and level of gene expression analyte that is determined according to the methods provided herein based on the amount of cell type present in two or more samples and experimentally measured levels of gene expression analyte. In one embodiment, the regression relationship is determined by determining the regression of overall levels of each gene expression analyte on determined cell proportions. In one embodiment, the regression relationship is determined by linear regression, where the overall expression level or the expression analyte levle is treated as directly proportional to (e.g., linear in) cell percent either for each cell type in turn or all at once and the slopes of these linear relationships can be expressed as beta values.
As used herein, a heterogeneous sample is to a sample that contains more than one cell type. For example, a heterogeneous sample can contain stromal cells and tumor cells. Typically, as used herein, the different cell types present in a sample are present in greater than about 0.1%, 0.2%, 0.3%, 0.5%, 0.7%, 1%, 2%, 3%, 4% or 5% or greater than 0.1%, 0.2%, 0.3%, 0.5%, 0.7%, 1%, 2%, 3%, 4% or 5%. As is understood in the art, cell samples, such as tissue samples from a subject, can contain minute amounts of a variety of cell types (e.g., nerve, blood, vascular cells). However, cell types that are not present in the sample in amounts greater than about 0.1%, 0.2%, 0.3%, 0.5%, 0.7%, 1%, 2%, 3%, 4% or 5% or greater than 0.1%, 0.2%, 0.3%, 0.5%, 0.7%, 1%, 2%, 3%, 4% or 5%, are not typically considered components of the heterogeneous cell sample, as used herein.
Related cell samples can be samples that contain one or more cell types in common. Related cell samples can be samples from the same tissue type or from the same organ. Related cell samples can be from the same or different sources (e.g., same or different individuals or cell cultures, or a combination thereof). As provided herein, in the case of three or more different cell samples, it is not required that all samples contain a common cell type, but if a first sample does not contain any cell types that are present in the other samples, the first sample is not related to the other samples.
Tumor cells are cells with cytological and adherence properties consisting of nuclear and cyoplasmic features and patterns of cell-to-cell association that are known to pathologists skilled in the art as sufficient for the diagnosis as cancers of various types. In some embodiments, tumor cells have abnormal growth properties, such as neoplastic growth properties.
The “cells associated with tumor” refers to cells that, while not necessarily malignant, are present in tumorous tissues or organs or particular locations of tissues or organs, and are not present, or are present at insignificant levels, in normal tissues or organs, or in particular locations of tissues or organs.
Benign prostatic hyperplastic (BPH) cells are cells of the epithelial lining of hyperplastic prostate glands. Dilated cystic glands cells are cells of the epithelial lining of dilated (atrophic) cystic prostate glands.
Stromal cells include connective tissue cells and smooth muscle cells forming the stroma of an organ. Exemplary stromal cells are cells of the stroma of the prostate gland.
A reference refers to a value or set of related values for one or more variables. In one example, a reference gene expression level refers to a gene expression level in a particular cell type. Reference expression levels can be determined according to the methods provided herein, or by determining gene expression levels of a cell type in a homogenous sample. Reference levels can be in absolute or relative amounts, as is known in the art. In certain embodiments, a reference expression level can be indicative of the presence of a particular cell type. For example, in certain embodiments, only one particular cell type may have high levels of expression of a particular gene, and, thus, observation of a cell type with high measured expression levels can match expression levels of that particular cell type, and thereby indicate the presence of that particular cell type in the sample. In another embodiment, a reference expression level can be indicative of the absence of a particular cell type. As provided herein, two or more references can be considered in determining whether or not a particular cell type is present in a sample, and also can be considered in determining the relative amount of a particular cell type that is present in the sample.
A modified t statistic is a numerical representation of the ability of a particular gene product or indicator thereof to indicate the presence or absence of a particular cell type in a sample. A modified t statistic incorporating goodness of fit and effect size can be formulated according to known methods (see, e.g., Tusher (2001) Proc. Natl. Acad. Sci. USA 98:5116-5121), where σβ is the standard error of the coefficient, and k is a small constant, as follows:
t=β/(k+σβ)
The relative content of a cell type or cell proportion is the amount of a cell mixture that is populated by a particular cell type. Typically, heterogeneous cell mixtures contain two or more cell types, and, therefore, no single cell type makes up 100% of the mixture. Relative content can be expressed in any of a variety of forms known in the art; For example, relative content can be expressed as a percentage of the total amount of cells in a mixture, or can be expressed relative to the amount of a particular cell type. As used herein, percent cell or percent cell composition is the percent of all cells that a particular cell type accounts for in a heterologous cell mixture, such as a microscopic section sampling a tissue.
An array or matrix is an arrangement of addressable locations or addresses on a device. The locations can be arranged in two dimensional arrays, three dimensional arrays, or other matrix formats. The number of locations can range from several to at least hundreds of thousands. Most importantly, each location represents a totally independent reaction site. Arrays include but are not limited to nucleic acid arrays, protein arrays and antibody arrays. A nucleic acid array refers to an array containing nucleic acid probes, such as oligonucleotides, polynucleotides or larger portions of genes. The nucleic acid on the array can be single stranded. Arrays wherein the probes are oligonucleotides are referred to as oligonucleotide arrays or oligonucleotide chips. A microarray, herein also refers to a biochip or biological chip, an array of regions having a density of discrete regions of at least about 100/cm2, and can be at least about 1000/cm2. The regions in a microarray have typical dimensions, e.g., diameters, in the range of between about 10-250 μm, and are separated from other regions in the array by about the same distance. A protein array refers to an array containing polypeptide probes or protein probes which can be in native form or denatured. An antibody array refers to an array containing antibodies which include but are not limited to monoclonal antibodies (e.g., from a mouse), chimeric antibodies, humanized antibodies or phage antibodies and single chain antibodies as well as fragments from antibodies.
An agonist is an agent that mimics or upregulates (e.g., potentiates or supplements) the bioactivity of a protein. An agonist can be a wild-type protein or derivative thereof having at least one bioactivity of the wild-type protein. An agonist can also be a compound that upregulates expression of a gene or which increases at least one bioactivity of a protein. An agonist can also be a compound which increases the interaction of a polypeptide with another molecule, e.g., a target peptide or nucleic acid.
The terms “polynucleotide” and “nucleic acid molecule” refer to nucleotides of any length, either ribonucleotides or deoxyribonucleotides. This term refers only to the primary structure of the molecule. Thus, this term includes double- and single-stranded DNA and RNA. It also includes known types of modifications, for example, labels which are known in the art, methylation, caps, substitution of one or more of the naturally occurring nucleotides with an analog, internucleotide modifications such as, for example, those with uncharged linkages (e.g., phosphorothioates and phosphorodithioates), those containing pendant moieties, such as, for example, proteins (including, e.g., nucleases, toxins, antibodies, signal peptides, and poly-L-lysine), those with intercalators (e.g., acridine and psoralen), those containing chelators (e.g., metals and radioactive metals), those containing alkylators, those with modified linkages (e.g., alpha anomeric nucleic acids), and those containing nucleotide analogs (e.g., peptide nucleic acids), as well as unmodified forms of the polynucleotide.
A polynucleotide derived from a designated sequence typically is a polynucleotide sequence which is comprised of a sequence of approximately at least about 6 nucleotides, at least about 8 nucleotides, at least about 10-12 nucleotides, or at least about 15-20 nucleotides corresponding to a region of the designated nucleotide sequence. Corresponding polynucleotides are homologous to or complementary to a designated sequence. Typically, the sequence of the region from which the polynucleotide is derived is homologous to or complementary to a sequence that is unique to a gene provided herein.
Recombinant polypeptides are polypeptides made using recombinant techniques, i.e., through the expression of a recombinant nucleic acid. A recombinant polypeptide can be distinguished from naturally occurring polypeptide by at least one or more characteristics. For example, the polypeptide may be isolated or purified away from some or all of the proteins and compounds with which it is normally associated in its wild type host, and thus may be substantially pure. For example, an isolated polypeptide is unaccompanied by at least some of the material with which it is normally associated in its natural state, constituting at least about 0.5%, or at least about 5% by weight of the total protein in a given sample. A substantially pure polypeptide comprises at least about 50-75% by weight of the total protein, at least about 80%, or at least about 90%. The definition includes the production of a polypeptide from one organism in a different organism or host cell. Alternatively, the polypeptide may be made at a significantly higher concentration than is normally seen, through the use of an inducible promoter or high expression promoter, such that the protein is made at increased concentration levels. Alternatively, the polypeptide may be in a form not normally found in nature, as in the addition of an epitope tag or amino acid substitutions, insertions and deletions, as discussed below.
The terms “disease” and “disorder” refer to a pathological condition in an organism resulting from, e.g., infection or genetic defect, and characterized by identifiable symptoms.
The “percent sequence identity” between a particular nucleic acid or amino acid sequence and a sequence referenced by a particular sequence identification number is determined as follows. First, a nucleic acid or amino acid sequence is compared to the sequence set forth in a particular sequence identification number using the BLAST 2 Sequences (Bl2seq) program from the stand-alone version of BLASTZ containing BLASTN version 2.0.14 and BLASTP version 2.0.14. This stand-alone version of BLASTZ can be obtained from Fish & Richardson's web site (world wide web at fr.com/blast) or the United States government's National Center for Biotechnology Information web site (world wide web at ncbi.nlm.nih.gov). Instructions explaining how to use the Bl2seq program can be found in the readme file accompanying BLASTZ. Bl2seq performs a comparison between two sequences using either the BLASTN or BLASTP algorithm BLASTN is used to compare nucleic acid sequences, while BLASTP is used to compare amino acid sequences. To compare two nucleic acid sequences, the options are set as follows: -i is set to a file containing the first nucleic acid sequence to be compared (e.g., C:\seq1.txt); -j is set to a file containing the second nucleic acid sequence to be compared (e.g., C:\seq2.txt); -p is set to blastn; -o is set to any desired file name (e.g., C:\output.txt); -q is set to −1; -r is set to 2; and all other options are left at their default setting. For example, the following command can be used to generate an output file containing a comparison between two sequences: CABl2seq c:\seq1.txt -j:\seq2.txt-p blastn-o c:\output.txt -q -1-r 2. To compare two amino acid sequences, the options of Bl2seq are set as follows: -i is set to a file containing the first amino acid sequence to be compared (e.g., C:\seq1.txt); -j is set to a file containing the second amino acid sequence to be compared (e.g., C:\seq2.txt); -p is set to blastp; -o is set to any desired file name (e.g., C:\output.txt); and all other options are left at their default setting. For example, the following command can be used to generate an output file containing a comparison between two amino acid sequences: C:\Bl2seq -i c:\seq1.txt -j c:\seq2.txt -p blastp -o c:\output.txt. If the two compared sequences share homology, then the designated output file will present those regions of homology as aligned sequences. If the two compared sequences do not share homology, then the designated output file will not present aligned sequences.
Once aligned, the number of matches is determined by counting the number of positions where an identical nucleotide or amino acid residue is presented in both sequences. The percent sequence identity is determined by dividing the number of matches either by the length of the sequence set forth in the identified sequence, or by an articulated length (e.g., 100 consecutive nucleotides or amino acid residues from a sequence set forth in an identified sequence), followed by multiplying the resulting value by 100. For example, a nucleic acid sequence that has 1166 matches when aligned with a 1200 bp sequence is 97.1 percent identical to the 1200 bp sequence (i.e., 1166÷1200*100=97.1). It is noted that the percent sequence identity value is rounded to the nearest tenth. For example, 75.11, 75.12, 75.13, and 75.14 is rounded down to 75.1, while 75.15, 75.16, 75.17, 75.18, and 75.19 is rounded up to 75.2. It is also noted that the length value will always be an integer. In another example, a target sequence containing a 20-nucleotide region that aligns with 20 consecutive nucleotides from an identified sequence as follows contains a region that shares 75 percent sequence identity to that identified sequence (i.e., 15÷20*100=75).
Polypeptides that at least 90% identical have percent identities from 90 to 100 relative to the reference polypeptides. Identity at a level of 90% or more can be indicative of the fact that, for a polynucleotide length of 100 amino acids no more than 10% (i.e., 10 out of 100) amino acids in the test polypeptide differ from those of the reference polypeptides. Similar comparisons can be made between test and reference polynucleotides. Such differences can be represented as point mutations randomly distributed over the entire length of an amino acid sequence or they can be clustered in one or more locations of varying length up to the maximum allowable, e.g., 10/100 amino acid difference (approximately 90% identity). Differences are defined as nucleic acid or amino acid substitutions, or deletions. At the level of homologies or identities above about 85-90%, the result should be independent of the program and gap parameters set; such high levels of identity can be assessed readily, often without relying on software.
A primer refers to an oligonucleotide containing two or more deoxyribonucleotides or ribonucleotides, typically more than three, from which synthesis of a primer extension product can be initiated. Experimental conditions conducive to synthesis include the presence of nucleoside triphosphates and an agent for polymerization and extension, such as DNA polymerase, and a suitable buffer, temperature and pH.
Animals can include any animal, such as, but are not limited to, goats, cows, deer, sheep, rodents, pigs and humans. Non-human animals, exclude humans as the contemplated animal. The SPs provided herein are from any source, animal, plant, prokaryotic and fungal.
Genetic therapy can involve the transfer of heterologous nucleic acid, such as DNA, into certain cells, target cells, of a mammal, particularly a human, with a disorder or conditions for which such therapy is sought. The nucleic acid, such as DNA, is introduced into the selected target cells in a manner such that the heterologous nucleic acid, such as DNA, is expressed and a therapeutic product encoded thereby is produced. Alternatively, the heterologous nucleic acid, such as DNA, can in some manner mediate expression of DNA that encodes the therapeutic product, or it can encode a product, such as a peptide or RNA that in some manner mediates, directly or indirectly, expression of a therapeutic product. Genetic therapy can also be used to deliver nucleic acid encoding a gene product that replaces a defective gene or supplements a gene product produced by the mammal or the cell in which it is introduced. The introduced nucleic acid can encode a therapeutic compound, such as a growth factor inhibitor thereof, or a tumor necrosis factor or inhibitor thereof, such as a receptor therefor, that is not normally produced in the mammalian host or that is not produced in therapeutically effective amounts or at a therapeutically useful time. The heterologous nucleic acid, such as DNA, encoding the therapeutic product can be modified prior to introduction into the cells of the afflicted host in order to enhance or otherwise alter the product or expression thereof. Genetic therapy can also involve delivery of an inhibitor or repressor or other modulator of gene expression.
A heterologous nucleic acid is nucleic acid that encodes RNA or RNA and proteins that are not normally produced in vivo by the cell in which it is expressed or that mediates or encodes mediators that alter expression of endogenous nucleic acid, such as DNA, by affecting transcription, translation, or other regulatable biochemical processes. Heterologous nucleic acid, such as DNA, can also be referred to as foreign nucleic acid, such as DNA. Any nucleic acid, such as DNA, that one of skill in the art would recognize or consider as heterologous or foreign to the cell in which is expressed is herein encompassed by heterologous nucleic acid; heterologous nucleic acid includes exogenously added nucleic acid that is also expressed endogenously. Examples of heterologous nucleic acid include, but are not limited to, nucleic acid that encodes traceable marker proteins, such as a protein that confers drug resistance, nucleic acid that encodes therapeutically effective substances, such as anti-cancer agents, enzymes and hormones, and nucleic acid, such as DNA, that encodes other types of proteins, such as antibodies. Antibodies that are encoded by heterologous nucleic acid can be secreted or expressed on the surface of the cell in which the heterologous nucleic acid has been introduced. Heterologous nucleic acid is generally not endogenous to the cell into which it is introduced, but has been obtained from another cell or prepared synthetically. Generally, although not necessarily, such nucleic acid encodes RNA and proteins that are not normally produced by the cell in which it is now expressed.
A therapeutically effective product for gene therapy can be a product encoded by heterologous nucleic acid, typically DNA, that, upon introduction of the nucleic acid into a host, a product is expressed that ameliorates or eliminates the symptoms, manifestations of an inherited or acquired disease or that cures the disease. Also included are biologically active nucleic acid molecules, such as RNAi and antisense.
Disease or disorder treatment or compound can include any therapeutic regimen and/or agent that, when used alone or in combination with other treatments or compounds, can alleviate, reduce, ameliorate, prevent, or place or maintain in a state of remission of clinical symptoms or diagnostic markers associated with the disease or disorder.
Nucleic acids include DNA, RNA and analogs thereof, including peptide nucleic acids (PNA) and mixtures thereof. Nucleic acids can be single or double-stranded. When referring to probes or primers, optionally labeled, with a detectable label, such as a fluorescent or radiolabel, single-stranded molecules are contemplated. Such molecules are typically of a length such that their target is statistically unique or of low copy number (typically less than 5, generally less than 3) for probing or priming a library. Generally a probe or primer contains at least 14, 16 or 30 contiguous of sequence complementary to or identical a gene of interest. Probes and primers can be 10, 20, 30, 50, 100 or more nucleic acids long.
Operative linkage of heterologous nucleic acids to regulatory and effector sequences of nucleotides, such as promoters, enhancers, transcriptional and translational stop sites, and other signal sequences refers to the relationship between such nucleic acid, such as DNA, and such sequences of nucleotides. Thus, operatively linked or operationally associated refers to the functional relationship of nucleic acid, such as DNA, with regulatory and effector sequences of nucleotides, such as promoters, enhancers, transcriptional and translational stop sites, and other signal sequences. For example, operative linkage of DNA to a promoter refers to the physical and functional relationship between the DNA and the promoter such that the transcription of such DNA is initiated from the promoter by an RNA polymerase that specifically recognizes, binds to and transcribes the DNA. In order to optimize expression and/or in vitro transcription, it can be necessary to remove, add or alter 5′ untranslated portions of the clones to eliminate extra, potential inappropriate alternative translation initiation (i.e., start) codons or other sequences that can interfere with or reduce expression, either at the level of transcription or translation. Alternatively, consensus ribosome binding sites (see, e.g., Kozak (1991) J. Biol. Chem. 266:19867-19870) can be inserted immediately 5′ of the start codon and can enhance expression. The desirability of (or need for) such modification can be empirically determined.
A sequence complementary to at least a portion of an RNA, with reference to antisense oligonucleotides, means a sequence having sufficient complementarity to be able to hybridize with the RNA, generally under moderate or high stringency conditions, forming a stable duplex; in the case of double-stranded antisense nucleic acids, a single strand of the duplex DNA (or dsRNA) can thus be tested, or triplex formation can be assayed. The ability to hybridize depends on the degree of complementarily and the length of the antisense nucleic acid. Generally, the longer the hybridizing nucleic acid, the more base mismatches with a gene encoding RNA it can contain and still form a stable duplex (or triplex, as the case can be). One skilled in the art can ascertain a tolerable degree of mismatch by use of standard procedures to determine the melting point of the hybridized complex.
Antisense polynucleotides are synthetic sequences of nucleotide bases complementary to mRNA or the sense strand of double-stranded DNA. Admixture of sense and antisense polynucleotides under appropriate conditions leads to the binding of the two molecules, or hybridization. When these polynucleotides bind to (hybridize with) mRNA, inhibition of protein synthesis (translation) occurs. When these polynucleotides bind to double-stranded DNA, inhibition of RNA synthesis (transcription) occurs. The resulting inhibition of translation and/or transcription leads to an inhibition of the synthesis of the protein encoded by the sense strand. Antisense nucleic acid molecules typically contain a sufficient number of nucleotides to specifically bind to a target nucleic acid, generally at least 5 contiguous nucleotides, often at least 14 or 16 or 30 contiguous nucleotides or modified nucleotides complementary to the coding portion of a nucleic acid molecule that encodes a gene of interest.
An antibody is an immunoglobulin, whether natural or partially or wholly synthetically produced, including any derivative thereof that retains the specific binding ability the antibody. Hence antibody includes any protein having a binding domain that is homologous or substantially homologous to an immunoglobulin binding domain. Antibodies include members of any immunoglobulin groups, including, but not limited to, IgG, IgM, IgA, IgD, IgY and IgE.
An antibody fragment is any derivative of an antibody that is less than full-length, retaining at least a portion of the full-length antibody's specific binding ability. Examples of antibody fragments include, but are not limited to, Fab, Fab′, F(ab)2, single-chain Fvs (scFV), FV, dsFV diabody and Fd fragments. The fragment can include multiple chains linked together, such as by disulfide bridges. An antibody fragment generally contains at least about 50 amino acids and typically at least 200 amino acids.
An Fv antibody fragment is composed of one variable heavy domain (VH) and one variable light domain linked by noncovalent interactions. A dsFV is an Fv with an engineered intermolecular disulfide bond, which stabilizes the VH-VL pair. An F(ab)2 fragment is an antibody fragment that results from digestion of an immunoglobulin with pepsin at pH 4.0-4.5; it can be recombinantly expressed to produce the equivalent fragment.
Fab fragments are antibody fragments that result from digestion of an immunoglobulin with papain; they can be recombinantly expressed to produce the equivalent fragment.
scFVs refer to antibody fragments that contain a variable light chain (VL) and variable heavy chain (VH) covalently connected by a polypeptide linker in any order. The linker is of a length such that the two variable domains are bridged without substantial interference. Included linkers are (Gly-Ser)n residues with some Glu or Lys residues dispersed throughout to increase solubility.
Humanized antibodies are antibodies that are modified to include human sequences of amino acids so that administration to a human does not provoke an immune response. Methods for preparation of such antibodies are known. For example, to produce such antibodies, the encoding nucleic acid in the hybridoma or other prokaryotic or eukaryotic cell, such as an E. coli or a CHO cell, that expresses the monoclonal antibody is altered by recombinant nucleic acid techniques to express an antibody in which the amino acid composition of the non-variable region is based on human antibodies. Computer programs have been designed to identify such non-variable regions.
Diabodies are dimeric scFV; diabodies typically have shorter peptide linkers than scFvs, and they generally dimerize.
The phrase “production by recombinant means by using recombinant DNA methods” refers to the use of the well known methods of molecular biology for expressing proteins encoded by cloned DNA.
An “effective amount” of a compound for treating a particular disease is an amount that is sufficient to ameliorate, or in some manner reduce the symptoms associated with the disease. Such amount can be administered as a single dosage or can be administered according to a regimen, whereby it is effective. The amount can cure the disease but, typically, is administered in order to ameliorate the symptoms of the disease. Repeated administration can be required to achieve the desired amelioration of symptoms.
A compound that modulates the activity of a gene product either decreases or increases or otherwise alters the activity of the protein or, in some manner up- or down-regulates or otherwise alters expression of the nucleic acid in a cell.
Pharmaceutically acceptable salts, esters or other derivatives of the conjugates include any salts, esters or derivatives that can be readily prepared by those of skill in this art using known methods for such derivatization and that produce compounds that can be administered to animals or humans without substantial toxic effects and that either are pharmaceutically active or are prodrugs.
A drug or compound identified by the screening methods provided herein refers to any compound that is a candidate for use as a therapeutic or as a lead compound for the design of a therapeutic. Such compounds can be small molecules, including small organic molecules, peptides, peptide mimetics, antisense molecules or dsRNA, such as RNAi, antibodies, fragments of antibodies, recombinant antibodies and other such compounds that can serve as drug candidates or lead compounds.
A non-malignant cell adjacent to a malignant cell in a subject is a cell that has a normal morphology (e.g., is not classified as neoplastic or malignant by a pathologist, cell sorter, or other cell classification method), but, while the cell was present intact in the subject, the cell was adjacent to a malignant cell or malignant cells. As provided herein, cells of a particular type (e.g., stroma) adjacent to a malignant cell or malignant cells can display an expression pattern that differs from cells of the same type that are not adjacent to a malignant cell or malignant cells. In accordance with the methods provided herein, cells that are adjacent to malignant cells can be distinguished from cells of the same type that are adjacent to non-malignant cells, according to their differential gene expression. As used herein regarding the location of cells, adjacent refers to a first cell and a second cell being sufficiently proximal such that the first cell influences the gene expression of the second cell. For example, adjacent cells can include cells that are in direct contact with each other, adjacent cell can include cells within 500 microns, 300 microns, 200 microns 100 microns or 50 microns, of each other.
A tumor is a collection of malignant cells. Malignant as applied to a cell refers to a cell that grows in an uncontrolled fashion. In some embodiments, a malignant cell can be anaplastic. In some embodiments, a malignant cell can be capable of metastasizing.
Hybridization stringency for, which can be used to determine percentage mismatch is as follows:
1) high stringency: 0.1×SSPE, 0.1% SDS, 65° C.
2) medium stringency: 0.2×SSPE, 0.1% SDS, 50° C.
3) low stringency: 1.0×SSPE, 0.1% SDS, 50° C.
A vector (or plasmid) refers to discrete elements that can be used to introduce heterologous nucleic acid into cells for either expression or replication thereof. Vectors typically remain episomal, but can be designed to effect integration of a gene or portion thereof into a chromosome of the genome. Also contemplated are vectors that are artificial chromosomes, such as yeast artificial chromosomes and mammalian artificial chromosomes. Selection and use of such vehicles are well known to those of skill in the art. An expression vector includes vectors capable of expressing DNA that is operatively linked with regulatory sequences, such as promoter regions, that are capable of effecting expression of such DNA fragments. Thus, an expression vector refers to a recombinant DNA or RNA construct, such as a plasmid, a phage, recombinant virus or other vector that, upon introduction into an appropriate host cell, results in expression of the cloned DNA. Appropriate expression vectors are well known to those of skill in the art and include those that are replicable in eukaryotic cells and/or prokaryotic cells and those that remain episomal or those that integrate into the host cell genome.
Disease prognosis refers to a forecast of the probable outcome of a disease or of a probable outcome resultant from a disease. Non-limiting examples of disease prognoses include likely relapse of disease, likely aggressiveness of disease, likely indolence of disease, likelihood of survival of the subject, likelihood of success in treating a disease, condition in which a particular treatment regimen is likely to be more effective than another treatment regimen, and combinations thereof.
Aggressiveness of a tumor or malignant cell is the capacity of one or more cells to attain a position in the body away from the tissue or organ of origin, attach to another portion of the body, and multiply. Experimentally, aggressiveness can be described in one or more manners, including, but not limited to, post-diagnosis survival of subject, relapse of tumor, and metastasis of tumor. Thus, in the disclosures provided herein, data indicative of time length of survival, relapse, non-relapse, time length for metastasis, or non-metastasis, are indicative of the aggressiveness of a tumor or a malignant cell. When survival is considered, one skilled in the art will recognize that aggressiveness is inversely related to the length of time of survival of the subject. When time length for metastasis is considered, one skilled in the art will recognize that aggressiveness is directly related to the length of time of survival of a subject. As used herein, indolence refers to non-aggressiveness of a tumor or malignant cell; thus, the more aggressive a tumor or cell, the less indolent, and vice versa. As an example of a cell attaining a position in the body away from the tissue or organ of origin, a malignant prostate cell can attain an extra-prostatic position, and thus have one characteristic of an aggressive malignant cell. Attachment of cells can be, for example, on the lymph node or bone marrow of a subject, or other sites known in the art.
A composition refers to any mixture. It can be a solution, a suspension, liquid, powder, a paste, aqueous, non-aqueous or any combination thereof.
A fluid is composition that can flow. Fluids thus encompass compositions that are in the form of semi-solids, pastes, solutions, aqueous mixtures, gels, lotions, creams and other such compositions.
Primary tissues are composed of many (e.g., two or more) types of cells. Identification of genes expressed in a specific cell type present within a tissue in other methods can require physical separation of that cell type and the cell type's subsequent assay. Although it is possible to physically separate cells according to type, by methods such as laser capture microdissection, centrifugation, FACS, and the like, this is time consuming and costly and in certain embodiments impractical to perform. Known expression profiling assays (either RNA or protein) of primary tissues or other specimens containing multiple cell types either (1) do not take into account that multiple cell types are present or (2) physically separate the component cell types before performing the assay. Other analyses have been performed without regard to the presence of multiple cell types, thereby identifying markers indicative of a shift in the relative proportion of various cell types present in a sample, but not representative of a specific cell type. Previous analytic approaches cannot discern interactions between different types of cells.
Provided herein are methods, compositions and kits based on the development of a model, where the level of each gene product assayed can be correlated to a specific cell type. This approach for determination of cell-type-specific gene expression obviates the need for physical separation of cells from tissues or other specimens with heterogeneous cell content. Furthermore, this method permits determination of the interaction between the different types of cells contained in such heterogeneous mixtures, which would otherwise have been difficult or impossible had the cells been first physically separated and then assayed. Using the approaches provided herein, a number of biomarkers can be identified related to various diseases and disorders. Exemplified herein is the identification of biomarkers for prostate cancer and benign prostatic hypertophy. Such biomarkers can be used in diagnosis and prognosis and treatment decisions.
The methods, compositions, combinations and kits provided herein employ a regression-based approach for identification of cell-type-specific patterns of gene expression in samples containing more than one type of cell. In one example, the methods, compositions, combinations and kits provided herein employ a regression-based approach for identification of cell-type-specific patterns of gene expression in cancer. These methods, compositions, combinations and kits provided herein can be used in the identification of genes that are differentially expressed in malignant versus non-malignant cells and further identify tumor-dependent changes in gene expression of non-malignant cells associated with malignant cells relative to non-malignant cells not associated with malignant cells. The methods, compositions, combinations and kits provided herein also can be used in correlating a phenotype with gene expression in one or more cell types. For example such a method can include determining the relative content of each cell type in two or more related heterogeneous cell samples, wherein at least two of the samples do not contain the same relative content of each cell type, measuring overall levels of one or more gene expression analytes in each sample, determining the regression relationship between the relative content of each cell type and the measured overall levels, and calculating the level of each of the one or more analytes in each cell type according to the regression relationship, where gene expression levels correspond to the calculated levels of analytes. In another example such a method can include determining the relative content of each cell type in two or more related heterogeneous cell samples, wherein at least two of the samples do not contain the same relative content of each cell type, measuring overall levels of two or more gene expression analytes in each sample, determining the regression relationship between the relative content of each cell type and the measured overall levels, and calculating the level of each of the two or more analytes in each cell type according to the regression relationship, where gene expression levels correspond to the calculated levels of analytes. Such methods can further include identifying genes differentially expressed in at least one cell type relative to at least one other cell type. In such methods, the analyte can be a nucleic acid molecule and a protein.
The methods provided herein can be used for determining cell-type-specific gene expression in any heterogeneous cell population. The methods provided herein can find application in samples known to contain a variety of cell types, such as brain tissue samples and muscle tissue samples. The methods provided herein also can find application in samples in which separation of cell type can represent a tedious or time consuming operation, which is no longer required under the methods provided herein. Samples used in the present methods can be any of a variety of samples, including, but not limited to, blood, cells from blood (including, but not limited to, non-blood cells such as epithelial cells in blood), plasma, serum, spinal fluid, lymph fluid, skin, sputum, alimentary and genitourinary samples (including, but not limited to, urine, semen, seminal fluid, prostate aspirate, prostatic fluid, and fluid from the seminal vesicles), saliva, milk, tissue specimens (including, but not limited to, prostate tissue specimens), tumors, organs, and also samples of in vitro cell culture constituents.
In certain embodiments, the methods provided herein can be used to differentiate true markers of tumor cells, hyperplastic cells, and stromal cells of cancer. As exemplified herein, least squares regression using individual cell-type proportions can be used to produce clear predictions of cell-specific expression for a large number of genes. In an example provided herein applied to prostate cancer, many of these predictions are accepted on the basis of prior knowledge of prostate gene expression and biology, which provide confidence in the method. These are illustrated by numerous genes predicted to be preferentially expressed by stromal cells that are characteristic of connective tissue and only poorly expressed or absent in epithelial cells.
In some embodiments, the methods provided herein allow segregation of molecular tumor and nontumor markers into more discrete and informative groups. Thus, genes identified as tumor-associated can be further categorized into tumor versus stroma (epithelial versus mesenchymal) and tumor versus hyperplastic (perhaps reflecting true differences between the malignant cell and its hyperplastic counterpart). The methods provided herein can be used to distinguish tumor and non-tumor markers in a variety of cancers, including, without limitation, cancers classified by site such as cancer of the oral cavity and pharynx (lip, tongue, salivary gland, floor of mouth, gum and other mouth, nasopharynx, tonsil, oropharynx, hypopharynx, other oral/pharynx); cancers of the digestive system (esophagus; stomach; small intestine; colon and rectum; anus, anal canal, and anorectum; liver; intrahepatic bile duct; gallbladder; other biliary; pancreas; retroperitoneum; peritoneum, omentum, and mesentery; other digestive); cancers of the respiratory system (nasal cavity, middle ear, and sinuses; larynx; lung and bronchus; pleura; trachea, mediastinum, and other respiratory); cancers of the mesothelioma; bones and joints; and soft tissue, including heart; skin cancers, including melanomas and other non-epithelial skin cancers; Kaposi's sarcoma and breast cancer; cancer of the female genital system (cervix uteri; corpus uteri; uterus, nos; ovary; vagina; vulva; and other female genital); cancers of the male genital system (prostate gland; testis; penis; and other male genital); cancers of the urinary system (urinary bladder; kidney and renal pelvis; ureter; and other urinary); cancers of the eye and orbit; cancers of the brain and nervous system (brain; and other nervous system); cancers of the endocrine system (thyroid gland and other endocrine, including thymus); lymphomas (Hodgkin's disease and non-Hodgkin's lymphoma), multiple myeloma, and leukemias (lymphocytic leukemia; myeloid leukemia; monocytic leukemia; and other leukemias); and cancers classified by histological type, such as Neoplasm, malignant; carcinoma, NOS; carcinoma, undifferentiated, NOS; giant and spindle cell carcinoma; small cell carcinoma, NOS; papillary carcinoma, NOS; squamous cell carcinoma, NOS; lymphoepithelial carcinoma; basal cell carcinoma, NOS; pilomatrix carcinoma; transitional cell carcinoma, NOS; papillary transitional cell carcinoma; adenocarcinoma, NOS; gastrinoma, malignant; cholangiocarcinoma; hepatocellular carcinoma, NOS; combined hepatocellular carcinoma and cholangiocarcinoma; trabecular adenocarcinoma; adenoid cystic carcinoma; adenocarcinoma in adenomatous polyp; adenocarcinoma, familial polyposis coli; solid carcinoma, NOS; carcinoid tumor, malignant; bronchiolo-alveolar adenocarcinoma; papillary adenocarcinoma, NOS; ccarcinoma; acidophil carcinoma; oxyphilic adenocarcinoma; basophil carcinoma; clear cell adenocarcinoma, NOS; granular cell carcinoma; follicular adenocarcinoma, NOS; papillary and follicular adenocarcinoma; nonencapsulating sclerosing carcinoma; adrenal cortical carcinoma; endometroid carcinoma; skin appendage carcinoma; apocrine adenocarcinoma; sebaceous adenocarcinoma; ceruminous adenocarcinoma; mucoepidermoid carcinoma; cystadenocarcinoma, NOS; papillary cystadenocarcinoma, NOS; papillary serous cystadenocarcinoma; mucinous cystadenocarcinoma, NOS; mucinous adenocarcinoma; signet ring cell carcinoma; infiltrating duct carcinoma; medullary carcinoma, NOS; lobular carcinoma; inflammatory carcinoma; Paget's disease, mammary; acinar cell carcinoma; adenosquamous carcinoma; adenocarcinoma with squamous metaplasia; thymoma, malignant; ovarian stromal tumor, malignant; thecoma, malignant; granulosa cell tumor, malignant; androblastoma, malignant; Sertoli cell carcinoma; Leydig cell tumor, malignant; lipid cell tumor, malignant; paraganglioma, malignant; extra-mammary paraganglioma, malignant; pheochromocytoma; glomangiosarcoma; malignant melanoma, NOS; amelanotic melanoma; superficial spreading melanoma; malignant melanoma in giant pigmented nevus; epithelioid cell melanoma; blue nevus, malignant; sarcoma, NOS; fibrosarcoma, NOS; fibrous histiocytoma, malignant; myxosarcoma; liposarcoma, NOS; leiomyosarcoma, NOS; rhabdomyosarcoma, NOS; embryonal rhabdomyosarcoma; alveolar rhabdomyosarcoma; stromal sarcoma, NOS; mixed tumor, malignant, NOS; Mullerian mixed tumor; nephroblastoma; hepatoblastoma; carcinosarcoma, NOS; mesenchymoma, malignant; Brenner tumor, malignant; phyllodes tumor, malignant; synovial sarcoma, NOS; mesothelioma, malignant; dysgerminoma; embryonal carcinoma, NOS; teratoma, malignant, NOS; struma ovarii, malignant; choriocarcinoma; mesonephroma, malignant; hemangiosarcoma; hemangioendothelioma, malignant; Kaposi's sarcoma; hemangiopericytoma, malignant; lymphangiosarcoma; osteosarcoma, NOS; juxtacortical osteosarcoma; chondrosarcoma, NOS; chondroblastoma, malignant; mesenchymal chondrosarcoma; giant cell tumor of bone; Ewing's sarcoma; odontogenic tumor, malignant; ameloblastic odontosarcoma; ameloblastoma, malignant; ameloblastic fibrosarcoma; pinealoma, malignant; chordoma; glioma, malignant; ependymoma, NOS; astrocytoma, NOS; protoplasmic astrocytoma; fibrillary astrocytoma; astroblastoma; glioblastoma, NOS; oligodendroglioma, NOS; oligodendroblastoma; primitive neuroectodermal; cerebellar sarcoma, NOS; ganglioneuroblastoma; neuroblastoma, NOS; retinoblastoma, NOS; olfactory neurogenic tumor; meningioma, malignant; neurofibrosarcoma; neurilemmoma, malignant; granular cell tumor, malignant; malignant lymphoma, NOS; Hodgkin's disease, NOS; Hodgkin's; paragranuloma, NOS; malignant lymphoma, small lymphocytic; malignant lymphoma, large cell, diffuse; malignant lymphoma, follicular, NOS; mycosis fungoides; other specified non-Hodgkin's lymphomas; malignant histiocytosis; multiple myeloma; mast cell sarcoma; immunoproliferative small intestinal disease; leukemia, NOS; lymphoid leukemia, NOS; plasma cell leukemia; erythroleukemia; lymphosarcoma cell leukemia; myeloid leukemia, NOS; basophilic leukemia; eosinophilic leukemia; monocytic leukemia, NOS; mast cell leukemia; megakaryoblastic leukemia; myeloid sarcoma; and hairy cell leukemia.
In an example comparing the results of a prostate tissue analysis using the methods provided herein to the results of previous methods, the vast majority of markers associated with normal prostate tissues in previous microarray-based studies relate to cells of the stroma. This result is not surprising given that normal samples can be composed of a relatively greater proportion of stromal cells.
In the example of prostate analysis, the strongest single discriminator between benign prostate hyperplasia (BPH) cells and tumor cells was CK15, a result confirmed by immunohistochemistry. CK15 has previously received little attention in this context, but BPH markers play an important role in the diagnosis of ambiguous clinical cases.
Transcripts whose expression levels have high covariance with cross-products of tissue proportions suggest that expression in one cell type depends on the proportion of another tissue, as would be expected in a paracrine mechanism. The stroma transcript with the highest dependence on tumor percentage was TGF-β2. Another such stroma cell gene for which immunohistochemistry was practical was desmin, which showed altered staining in the tumor-associated stroma. In fact, a large number of typical stroma cell genes displayed dependence on the proportion of tumor, adding evidence to the speculation that tumor-associated stroma differs from non-associated stroma. Tumor-stroma paracrine signaling can be reflected in peritumor halos of altered gene expression that can present a much bigger target for detection than the tumor cells alone.
The methods provided herein provide a straightforward approach using simple and multiple linear regression to identify genes whose expression in tissue is specifically correlated with a specific cell type (e.g., in prostate tissue with either tumor cells, BPH epithelial cells or stromal cells). Context-dependent expression that is not readily attributable to single cell types is also recognized. The investigative approach described here is also applicable to a wide variety of tumor marker discovery investigations in a variety of tissues and organs. The exemplary prostate analysis results presented herein demonstrate the ability to identify a large number of gene candidates as specific products of various cells involved in prostate cancer pathogenesis.
A model for cell-specific gene expression is established by both (1) determination of the proportion of each constituent cell type (e.g., epithelium, stroma, tumor, or other discriminating entity) within a given type of tissue or specimen (e.g., prostate, breast, colon, marrow, and the like) and (2) assay of the expression profile (e.g., RNA or protein) of that same tissue or specimen. In some embodiments, cell type specific expression of a gene can be determined by fitting this model to data from a collection of tissue samples.
The methods provided herein can include a step of determining the relative content of each cell type in a heterogeneous sample. Identification of a cell type in a sample can include identifying cell types that are present in a sample in amounts greater than about 1%, 2%, 3%, 4% or 5% or greater than 1%, 2%, 3%, 4% or 5%. Any of a variety of known methods for cell type identification can be used herein.
For example, cell type can be determined by an individual skilled in the ability to identify cell types, such as a pathologist or a histologist. In another example, cell types can be determined by cell sorting and/or flow cytometry methods known in the art.
The methods provided herein can be used to determine that the nucleotide or proteins are differentially expressed in at least one cell type relative to at least one other cell type. Such genes include those that are up-regulated (i.e., expressed at a higher level), as well as those that are down-regulated (i.e., expressed at a lower level). Such genes also include sequences that have been altered (i.e., truncated sequences or sequences with substitutions, deletions or insertions, including point mutations) and show either the same expression profile or an altered profile. In certain embodiments, the genes can be from humans; however, as will be appreciated by those in the art, genes from other organisms can be useful in animal models of disease and drug evaluation; thus, other genes are provided, from vertebrates, including mammals, including rodents (e.g., rats, mice, hamsters, and guinea pigs), primates, and farm animals (e.g., sheep, goats, pigs, cows, and horses). In some cases, prokaryotic genes can be useful. Gene expression in any of a variety of organisms can be determined by methods provided herein or otherwise known in the art.
Gene products measured according to the methods provided herein can be nucleic acid molecules, including, but not limited to mRNA or an amplicate or complement thereof, polypeptides, or fragments thereof. Methods and compositions for the detection of nucleic acid molecules and proteins are known in the art. For example, oligonucleotide probes and primers can be used in the detection of nucleic acid molecules, and antibodies can be used in the detection of polypeptides.
In the methods provided herein, one or more gene products can be detected. In some embodiments, two or more gene products are detected. In other embodiments, 3 or more, 4 or more, 5 or more, 7 or more, 10 or more 15 or more, 20 or more 25, or more, 35 or more, 50 or more, 75 or more, or 100 or more gene products can be detected in the methods provided herein.
The expression levels of the marker genes in a sample can be determined by any method or composition known in the art. The expression level can be determined by isolating and determining the level (i.e., amount) of nucleic acid transcribed from each marker gene. Alternatively, or additionally, the level of specific proteins translated from mRNA transcribed from a marker gene can be determined.
Determining the level of expression of specific marker genes can be accomplished by determining the amount of mRNA, or polynucleotides derived therefrom, or protein present in a sample. Any method for determining protein or RNA levels can be used. For example, protein or RNA is isolated from a sample and separated by gel electrophoresis. The separated protein or RNA is then transferred to a solid support, such as a filter. Nucleic acid or protein (e.g., antibody) probes representing one or more markers are then hybridized to the filter by hybridization, and the amount of marker-derived protein or RNA is determined. Such determination can be visual, or machine-aided, for example, by use of a densitometer. Another method of determining protein or RNA levels is by use of a dot-blot or a slot-blot. In this method, protein, RNA, or nucleic acid derived therefrom, from a sample is labeled. The protein, RNA or nucleic acid derived therefrom is then hybridized to a filter containing oligonucleotides or antibodies derived from one or more marker genes, wherein the oligonucleotides or antibodies are placed upon the filter at discrete, easily-identifiable locations. Binding, or lack thereof, of the labeled protein or RNA to the filter is determined visually or by densitometer. Proteins or polynucleotides can be labeled using a radiolabel or a fluorescent (i.e., visible) label.
Methods provided herein can be used to detect mRNA or amplicates thereof, and any fragment thereof. In one example, introns of mRNA or amplicate or fragment thereof can be detected. Processing of mRNA can include splicing, in which introns are removed from the transcript. Detection of introns can be used to detect the presence of the entire mRNA, and also can be used to detect processing of the mRNA, for example, when the intron region alone (e.g., intron not attached to any exons) is detected.
In another embodiment, methods provided herein can be used to detect polypeptides and modifications thereof, where a modification of a polypeptide can be a post-translation modification such as lipidylation, glycosylation, activating proteolysis, and others known in the art, or can include degradational modification such as proteolytic fragments and ubiquitinated polypeptides.
These examples are not intended to be limiting; other methods of determining protein or RNA abundance are known in the art.
Alternatively, proteins can be separated by two-dimensional gel electrophoresis systems. Two-dimensional gel electrophoresis is well-known in the art and can involve isoelectric focusing along a first dimension followed by SDS-PAGE electrophoresis along a second dimension. See, e.g., Hames et al. (1990) Gel Electrophoresis of Proteins: A Practical Approach, IRL Press, New York; Shevchenko et al. (1996) Proc. Natl. Acad. Sci. USA 93:1440-1445; Sagliocco et al. (1996) Yeast 12:1519-1533; and Lander (1996) Science 274:536-539. The resulting electropherograms can be analyzed by numerous techniques, including mass spectrometric techniques, western blotting and immunoblot analysis using polyclonal and monoclonal antibodies.
Alternatively, marker-derived protein levels can be determined by constructing an antibody microarray in which binding sites comprise immobilized antibodies, such as monoclonal antibodies, specific to a plurality of protein species encoded by the cell genome. Antibodies can be present for a substantial fraction of the marker-derived proteins of interest. Methods for making monoclonal antibodies are well known (see, e.g., Harlow and Lane (1988) Antibodies: A Laboratory Manual, Cold Spring Harbor, N.Y., which is incorporated in its entirety for all purposes). In one embodiment, monoclonal antibodies are raised against synthetic peptide fragments designed based on genomic sequence of the cell. With such an antibody array, proteins from the cell are contacted to the array, and their binding is assayed with assays known in the art. The expression, and the level of expression, of proteins of diagnostic or prognostic interest can be detected through immunohistochemical staining of tissue slices or sections.
In another embodiment, expression of marker genes in a number of tissue specimens can be characterized using a tissue array (Kononen et al. (1998) Nat. Med. 4:844-847). In a tissue array, multiple tissue samples are assessed on the same microarray. The arrays allow in situ detection of RNA and protein levels; consecutive sections allow the analysis of multiple samples simultaneously.
In some embodiments, polynucleotide microarrays are used to measure expression so that the expression status of each of the markers above is assessed simultaneously. In one embodiment, the microarrays provided herein are oligonucleotide or cDNA arrays comprising probes hybridizable to the genes corresponding to the marker genes described herein. A microarray as provided herein can comprise probes hybridizable to the genes corresponding to markers able to distinguish cells, identify phenotypes, identify a disease or disorder, or provide a prognosis of a disease or disorder (e.g., a classifier as described herein). For example, provided herein are polynucleotide arrays comprising probes to a subset or subsets of at least 2, 5, 10, 15, 20, 30, 40, 50, 75, 100, or more than 100 genetic markers, up to the full set of markers present in a classifier as described in the Examples below. Also provided herein are probes to markers with a modified t statistic greater than or equal to 2.5, 3, 3.5, 4, 4.5 or 5. Also provided herein are probes to markers with a modified t statistic less than or equal to −2.5, −3, −3.5, −4, −4.5 or −5. In specific embodiments, the invention provides combinations such as arrays in which the markers described herein comprise at least 50%, 60%, 70%, 80%, 85%, 90%, 95% or 98% of the probes on the combination or array.
General methods pertaining to the construction of microarrays comprising the marker sets and/or subsets above are known in the art as described herein.
Microarrays can be prepared by selecting probes that comprise a polypeptide or polynucleotide sequence, and then immobilizing such probes to a solid support or surface. For example, the probes can comprise DNA sequences, RNA sequences, or antibodies. The probes can also comprise amino acid, DNA and/or RNA analogues, or combinations thereof. The probes can be prepared by any method known in the art.
The probe or probes used in the methods of the invention can be immobilized to a solid support which can be either porous or non-porous. For example, the probes of the can be attached to a nitrocellulose or nylon membrane or filter. Alternatively, the solid support or surface can be a glass or plastic surface. In another embodiment, hybridization levels are measured to microarrays of probes consisting of a solid phase on the surface of which are immobilized a population of probes. The solid phase can be a nonporous or, optionally, a porous material such as a gel.
In another embodiment, the microarrays are addressable arrays, such as positionally addressable arrays. More specifically, each probe of the array can be located at a known, predetermined position on the solid support such that the identity (i.e., the sequence) of each probe can be determined from its position in the array (i.e., on the support or surface).
A skilled artisan will appreciate that positive control probes, e.g., probes known to be complementary and hybridizable to sequences in target polynucleotide molecules, and negative control probes, e.g., probes known to not be complementary and hybridizable to sequences in target polynucleotide molecules, can be included on the array. In one embodiment, positive controls can be synthesized along the perimeter of the array. In another embodiment, positive controls can be synthesized in diagonal stripes across the array. Other variations are known in the art. Probes can be immobilized on the to solid surface by any of a variety of methods known in the art.
In certain embodiments, this model can be further extended to include sample characteristics, such as cell or organism phenotypes, allowing cell type specific expression to be linked to observable indicia such as clinical indicators and prognosis (e.g., clinical disease progression, response to therapy, and the like). In one embodiment, a model for prostate tissue is provided, resulting in identification of cell-type-specific markers of cancer, epithelial hypertrophy, and disease progression. In another embodiment, a method for studying differential gene expression between subjects with cancers that relapse and those with cancers that do not relapse, is disclosed. Also provided is the framework for studying mixed cell type samples and more flexible models allowing for cross-talk among genes in a sample. Also provided are extensions to defining differences in expression between samples with different characteristics, such as samples from subjects who subsequently relapse versus those who do not.
The methods provided herein include determining the regression relationship between relative cell content and measured expression levels. For example, the regression relationship can be determined by determining the regression of measured expression levels on cell proportions. Statistical methods for determining regression relationships between variables are known in the art. Such general statistical methods can be used in accordance with the teachings provided herein regarding regression of measured expression levels on cell proportions.
The methods provided herein also include calculating the level of analytes in each cell type based on the regression relationship between relative cell content and expression levels. The regression relationship can be determined according to methods provided herein, and, based on the regression relationship, the level of a particular analyte can be calculated for a particular cell type. The methods provided herein can permit the calculation of any of a variety of analyte for particular cell types. For example, the methods provided herein can permit calculation of a single analyte for a single cell type, or can permit calculation of a plurality of analytes for a single cell type, or can permit calculation of a single analyte for a plurality of cell types, or can permit calculation of a plurality of analytes for a plurality of cell types. Thus, the number of analytes whose level can be calculated for a particular cell type can range from a single analyte to the total number of analytes measured (e.g., the total number of analytes measured using a microarray). In another embodiment, the total number of cell types for which analyte levels can be calculated can range from a single cell type, to all cell types present in a sample at sufficient levels. The levels of analyte for a particular cell type can be used to estimate expression levels of the corresponding gene, as provided elsewhere herein.
The methods provided herein also can include identifying genes differentially expressed in a first cell type relative to a second cell type. Expression levels of one or more genes in a particular cell type can be compared to one or more additional cell types. Differences in expression levels can be represented in any of a variety of manners known in the art, including mathematical or statistical representations, as provided herein. For example, differences in expression level can be represented as a modified t statistic, as described elsewhere herein.
The methods provided herein also can serve as the basis for methods of indicating the presence of a particular cell type in a subject. The methods provided herein can be used for identifying the expression levels in particular cell types. Using any of a variety of classifier methods known in the art, such as a naïve Bayes classifier, gene expression levels in cells of a sample from a subject can be compared to reference expression levels to determine the presence of absence, and, optionally, the relative amount, of a particular cell type in the sample. For example, the markers provided herein as associated with prostate tumor, stroma or BPH can be selected in a prostate tumor classifier in accordance with the modified t statistic associated with each marker provided in the Tables herein. Methods for using a modified t statistic in classifier methods are provided herein and also are known in the art. In another embodiment, the methods provided herein can be used in phenotype-indicating methods such as diagnostic or prognostic methods, in which the gene expression levels in a sample from a subject can be compared to references indicative of one or more particular phenotypes.
For purposes of exemplification, and not for purposes of limitation, an exemplary method of determining gene expression levels in one or more cell types in a heterogeneous cell sample is provided as follows. Suppose that there are four cell types: BPH, Tumor, Stroma, fij(y), iε{BPH, Tumor, Stroma, Cystic Atrophy} and Cystic Atrophy. Supposing that each cell type has a (possibly) different distribution for y, the expression level for a gene j, denoted by:
and that sample k has proportions
X
k=(xk,BPH,xk,Tumor,xk,stroma,xk,Cystic Atrophy)
of each cell type is studied. The distribution of the expression level for gene j is then
if the expression levels are additive in the cell proportions as they would be if each cell's expression level depends only on the type of cell (and not, say, on what other types of cells can be present in the sample). In a later section this formulation is extended to cases in which the expression of a given cell type depends on what other types of cells are present.
The average expression level in a sample is then the weighted average of the expectations with weights corresponding to the cell proportions:
This is the known form for a multiple linear regression equation (without specifying an intercept), and when multiple samples are available one can estimate the βij. Once these estimates are in hand, estimates for the differences in gene expression of two cell types are of the form:
{circumflex over (β)}i1j−{circumflex over (β)}i2j
and standard methods for testing linear hypotheses about the coefficients βij can be applied to test whether the average expression levels of cell types i1 and i2 are different. The term ‘expression levels’ as used in this exemplification of the method is used in a generic sense: ‘expression levels’ could be readings of mRNA levels, cRNA levels, protein levels, fluorescent intensity from a feature on an array, the logarithm of that reading, some highly post-processed reading, and the like. Thus, differences in the coefficients can correspond to differences, log ratios, or some other functions of the underlying transcript abundance.
For computational convenience, one may in certain embodiments use Z=XT and γ=T−1β setting up T so that one column of T has all zeroes but for a one in position i1 and a minus one in position i2 such as
The columns of Z that result are the unit vector (all ones), χk,BPH+χk,Tumor, χk,BPH−χk,Tumor, and χk,Stroma. With this setup, twice the coefficient of χk,BPH−χk,Tumor estimates the average difference in expression level of a tumor cell versus a BPH cell. With this parameterization, standard software can be used to provide an estimate and a tesmodified t statistic for the average difference of tumor and BPH cells. Further, this can simplify the specification of restricted models in which two or more of the tissue components have the same average expression level.
The data for a study can contain a large number of samples from a smaller number of different men. It is plausible that the samples from one man may tend to share a common level of expression for a given gene, differences among his cells according to their type notwithstanding. This will tend to lead to positive covariance among the measurements of expression level within men. Ordinary least squares (OLS) estimates are less than fully efficient in such circumstances. One alternative to OLS is to use a weighted least squares approach that treats a collection of samples from a single subject as having a common (non-negative) covariance and identical variances.
The estimating equation for this setup can be solved via iterative methods using software such as the gee library from R (Ihaka and Gentleman (1996) J. Comp. Graph. Stat. 5:299-314). When the estimated covariance is negative—as sometimes happens when there is an extreme outlier in the dataset—it can be fixed at zero. Also the sandwich estimate (Liang and Zeger (1986) Biometrika 73:13-22) of the covariance structure can be used.
The estimating equation approach will provide a tesmodified t statistic for a single transcript. Assessment of differential expression among a group of 12625 transcripts is handled by permutation methods that honor a suitable null model. That null model is obtained by regressing the expression level on all design terms except for the ‘BPH—tumor’ term using the exchangeable, non-negative correlation structure just mentioned. For performing permutation tests, the correlation structure in the residuals can be accounted for. Let κ1 be the set of n1 indexes of samples for subject 1. First, we find yjk−ŷjk=ejk, kεκ1, as the residuals from that fitted null model for subject 1. The inverse square root of the correlation matrix of these residuals is used to transform them, i.e., {tilde over (e)}j=φ−1/2ej., where φ is the (block diagonal) correlation matrix obtained by substituting the estimate of r from gee as the off-diagonal elements of blocks corresponding to measurements for each subject and ej. and {tilde over (e)}j. are the vector of residuals and transformed residuals for all subjects for gene j. Asymptotically, the {tilde over (e)}jk have means and covariances equal to zero. Random permutations of these, {tilde over (e)}j(i), i=1, . . . , M, are obtained and used to form pseudo-observations:
{tilde over (y)}
j.
(i)
=ŷ
j.+φ1/2{tilde over (e)}j.(i)
This permutation scheme preserves the null model and enforces its correlation structure asymptotically.
In certain embodiments, the contribution of each type of cell does not depend on what other cell types are present in the sample. However, there can be instances in which contribution of each type of cell does depend on other cell types present in the sample. It may happen that putatively ‘normal’ cells exhibit genomic features that influence both their expression profiles and their potential to become malignant. Such cells would exhibit the same expression pattern when located in normal tissue, but are more likely to be found in samples that also have tumor cells in them. Another possible effect is that signals generated by tumor cells trigger expression changes in nearby cells that would not be seen if those same cells were located in wholly normal tissue. In either case, the contribution of a cell may be more or less than in another tissue environment leading to a setup in which the contributions of individual cell types to the overall profile depend on the proportions of all types present, viz.
as do the expected proportions
The methods used herein above can still be applied in the context provided some calculable form is given for βij(Xk). One choice is given by
βij(Xk)=(φjR(Xk))i
where Φj is a 4×m matrix of unknown coefficients and R(Xk) is a column vector of m elements. This reduces to the case in which each cell's expression level depends only on the type of cell when Φj is 4×1 matrix and R(Xk) is just ‘1’.
Consider the case:
(and recall that ΣjXk,j=1.) Here the subscript for Tumor has been abbreviated T etc., for brevity. This setup provides that BPH (B), tumor, and cystic atrophy (C) cells have expression profiles that do not depend on the other cell types in the sample. However, the expression levels of stromal cells (S) depend on the proportion of tumor cells as reflected by the coefficient δj. Notice that
is linear in Xk,B, Xk,T, Xk,S, Xk,C, and Xk,SXk,T with the unknown coefficients being
X
kφjR(Xk)=xk,BvBj+xk,TvTj+xk,SvSj+xk,Sxk,xδj+xk,CvCj
multipliers of those terms. So, the unknowns in this case are linear functions of the gene expression levels and can be determined using standard linear models as was done earlier. The only change here is the addition of the product of Xk,S and Xk,T. Such a product, when significant, is termed an “interaction” and refers to the product archiving a significance level owing to a correlation of Xk,S with Xk,T. Thus, it is possible to accommodate variations in gene expression that occur when the level of a transcript in one cell type is influenced by the amount of another cell type in the sample. In one aspect, a setup involving a dependency of tumor on the amount of stroma
the expression for XkΦjR(Xk) is precisely as it was just above.
Accordingly, one can screen for dependencies by including as regressors products of the proportions of cell types. In certain embodiments, it may not be possible to detect interactions if two different cell types experience equal and opposite changes—one type expressing more with increases in the other and the other expressing less with increases in the first. In one embodiment, dependence of gene expression refers to the dependence of gene expression in one cell type on the level of gene expression in another cell type. In another embodiment, dependence of gene expression refers to the dependence of gene expression in one cell type on the amount of another cell type.
The contribution of each type of cell can depend on what other cell types are present in the sample, but also can depend on other characteristics of the sample, such as clinical characteristics of the subject who contributed it. For example, clinical characteristics such as disease symptoms, disease prognosis such as relapse and/or aggressiveness of disease, likelihood of success in treating a disease, likelihood of survival, condition in which a particular treatment regimen is likely to be more effective than another treatment regimen, can be correlated with cell expression. For example, cell type specific gene expression can differ between a subject with a cancer that does not relapse after treatment and a subject with a cancer that does relapse after treatment. In this case, the contribution of a cell type may be more or less than in another subject leading to an instance in which the contributions of individual cell types to the overall profile depend on the characteristics of the subject or sample. Here, the model used earlier is extended to allow for dependence on a vector of sample specific covariates, Zk:
as do the expected proportions:
The methods used herein above can still be applied in this context provided some reasonable form is given for βij(Xk,Zk). One useful choice is given by:
βij(Xk,Zk)=(φjR(Zk))i
Where Φj is a 4×m matrix of unknown coefficients and R(Zk) is a column vector of m elements.
Consider how this would be used to study differences in gene expression among subjects who relapse and those who do not. In this case, Zk is an indicator variable taking the value zero for samples of subjects who do not relapse and one for those who do. Then
and Φ is a four by two matrix of coefficients:
Notice that this leads to
X
kφjR(Zk)=xk,BvBj+xk,TvTj+xk,SvSj+xk,CvCj+xk,BZkδBj+xk,TZkδTj+xk,SZkδSj+xk,CZkδCj
The v coefficients give the average expression of the different cell types in subjects who do not relapse, while the δ coefficients give the difference between the average expression of the different cell types in subjects who do relapse and those who do not. Thus, a non-zero value of δT would indicate that in tumor cells, the average expression level differs for subjects who relapse and those who do not. The above equation is linear in its coefficients, so standard statistical methods can be applied to estimation and inference on the coefficients. Extensions that allow β to depend on both cell proportions and on sample covariates can be determined according to the teachings provided herein or other methods known in the art.
Provided herein are tables and exhibits listing probe sets and genes associated with the probe set, including, for some tables, GENBANK accession number, and/or locus ID. The tables may include modified t statistics for an Affymetrix microarrays, including associated t statistics for BPH, tumor, stroma and cystic atrophy, for example. Probe IDs for the microarray that map to Probe IDs for a different microarray, and the mapping itself, also may be provided, where the mapping can represent Probe IDs of microarrays that can hybridize to the same gene. By virtue of such mapping, Probe IDs can be associated with nucleotide sequences. Tables also may list the top genes identified as up- and down-regulated in prostate tumor cells of relapse patients, calculated by linear regression including all samples with prostate cancer. Genes that have greater than, for example, a 1.5 fold ratio of predicted expression between relapse and non-relapse tissue can be identified, as can an absolute difference in expression that exceeds the expression level reported for most genes queried by the array.
The tables provided herein also may list the top genes identified as up- and down-regulated in tumors and/or prostate stroma of relapse patients, calculated by linear regression including all samples with prostate cancer. Exemplary genes whose expression can be examined in methods for identifying or characterizing a sample may be provided, as well as Probe IDs that can be used for such gene expression identification.
Splice variants of genes also may be useful for determining diagnosis and prognosis of prostate cancer. As will be understood in the art, multiple splicing combinations are provided for some genes. Reference herein to one or more genes (including reference to products of genes) also contemplates reference to spliced gene sequences. Similarly, reference herein to one or more protein gene products also contemplates proteins translated from splice variants.
Exemplary, non-limiting examples of genes whose products can be detected in the methods provided herein include IGF-1, microsimino protein, and MTA-1. In one embodiment detection of the expression of one or more of these genes can be performed in combination with detection of expression of one or more additional genes as listed in the tables herein.
Uses of probes and detection of genes identified in the tables may be described and exemplified herein. It is contemplated herein that uses and methods similar to those exemplified can be applied to the probe and gene nucleotide sequences in accordance with the teachings provided herein.
The isolated nucleic acids can contain least 10 nucleotides, 25 nucleotides, 50 nucleotides, 100 nucleotides, 150 nucleotides, or 200 nucleotides or more, contiguous nucleotides of a gene listed herein. In another embodiment, the nucleic acids are smaller than 35, 200 or 500 nucleotides in length.
Also provided are fragments of the above nucleic acids that can be used as probes or primers and that contain at least about 10 nucleotides, at least about 14 nucleotides, at least about 16 nucleotides, or at least about 30 nucleotides. The length of the probe or primer is a function of the size of the genome probed; the larger the genome, the longer the probe or primer required for specific hybridization to a single site. Those of skill in the art can select appropriately sized probes and primers. Probes and primers as described can be single-stranded. Double stranded probes and primers also can be used, if they are denatured when used. Probes and primers derived from the nucleic acid molecules are provided. Such probes and primers contain at least 8, 14, 16, 30, 100 or more contiguous nucleotides. The probes and primers are optionally labeled with a detectable label, such as a radiolabel or a fluorescent tag, or can be mass differentiated for detection by mass spectrometry or other means. Also provided is an isolated nucleic acid molecule that includes the sequence of molecules that is complementary to a nucleotide. Double-stranded RNA (dsRNA), such as RNAi is also provided.
Plasmids and vectors containing the nucleic acid molecules are also provided. Cells containing the vectors, including cells that express the encoded proteins are provided. The cell can be a bacterial cell, a yeast cell, a fungal cell, a plant cell, an insect cell or an animal cell.
For recombinant expression of one or more genes, the nucleic acid containing all or a portion of the nucleotide sequence encoding the genes can be inserted into an appropriate expression vector, i.e., a vector that contains the elements for the transcription and translation of the inserted protein coding sequence. Transcriptional and translational signals also can be supplied by the native promoter for the genes, and/or their flanking regions.
Also provided are vectors that contain nucleic acid encoding a gene listed herein. Cells containing the vectors are also provided. The cells include eukaryotic and prokaryotic cells, and the vectors are any suitable for use therein.
Prokaryotic and eukaryotic cells containing the vectors are provided. Such cells include bacterial cells, yeast cells, fungal cells, plant cells, insect cells and animal cells. The cells can be used to produce an oligonucleotide or polypeptide gene products by (a) growing the above-described cells under conditions whereby the encoded gene is expressed by the cell, and then (b) recovering the expressed compound.
A variety of host-vector systems can be used to express the protein coding sequence. These include, but are not limited to, mammalian cell systems infected with virus (e.g., vaccinia virus and adenovirus); insect cell systems infected with virus (e.g., baculovirus); microorganisms such as yeast containing yeast vectors; or bacteria transformed with bacteriophage, DNA, plasmid DNA, or cosmid DNA. The expression elements of vectors vary in their strengths and specificities. Depending on the host-vector system used, any one of a number of suitable transcription and translation elements can be used.
Any methods known to those of skill in the art for the insertion of nucleic acid fragments into a vector can be used to construct expression vectors containing a chimeric gene containing appropriate transcriptional/translational control signals and protein coding sequences. These methods can include in vitro recombinant DNA and synthetic techniques and in vivo recombinants (genetic recombination). Expression of nucleic acid sequences encoding polypeptide can be regulated by a second nucleic acid sequence so that the genes or fragments thereof are expressed in a host transformed with the recombinant DNA molecule(s). For example, expression of the proteins can be controlled by any promoter/enhancer known in the art.
Protein products of the genes listed herein, derivatives, and analogs can be produced by various methods known in the art. For example, once a recombinant cell expressing such a polypeptide, or a domain, fragment or derivative thereof, is identified, the individual gene product can be isolated and analyzed. This is achieved by assays based on the physical and/or functional properties of the protein, including, but not limited to, radioactive labeling of the product followed by analysis by gel electrophoresis, immunoassay, cross-linking to marker-labeled product, and assays of protein activity or antibody binding.
Polypeptides can be isolated and purified by standard methods known in the art (either from natural sources or recombinant host cells expressing the complexes or proteins), including but not restricted to column chromatography (e.g., ion exchange, affinity, gel exclusion, reversed-phase high pressure and fast protein liquid), differential centrifugation, differential solubility, or by any other standard technique used for the purification of proteins. Functional properties can be evaluated using any suitable assay known in the art.
Manipulations of polypeptide sequences can be made at the protein level. Also contemplated herein are polypeptide proteins, domains thereof, derivatives or analogs or fragments thereof, which are differentially modified during or after translation, e.g., by glycosylation, acetylation, phosphorylation, amidation, derivatization by known protecting/blocking groups, proteolytic cleavage, linkage to an antibody molecule or other cellular ligand. Any of numerous chemical modifications can be carried out by known techniques, including but not limited to specific chemical cleavage by cyanogen bromide, trypsin, chymotrypsin, papain, V8 protease, NaBH4, acetylation, formulation, oxidation, reduction, metabolic synthesis in the presence of tunicamycin and other such agents.
In addition, domains, analogs and derivatives of a polypeptide provided herein can be chemically synthesized. For example, a peptide corresponding to a portion of a polypeptide provided herein, which includes the desired domain or which mediates the desired activity in vitro can be synthesized by use of a peptide synthesizer. Furthermore, if desired, nonclassical amino acids or chemical amino acid analogs can be introduced as a substitution or addition into the polypeptide sequence. Non-classical amino acids include but are not limited to the D-isomers of the common amino acids, a-amino isobutyric acid, 4-aminobutyric acid, Abu, 2-aminobutyric acid, .epsilon.-Abu, e-Ahx, 6-amino hexanoic acid, Aib, 2-amino isobutyric acid, 3-amino propionoic acid, ornithine, norleucine, norvaline, hydroxyproline, sarcosine, citrulline, cysteic acid, t-butylglycine, t-butylalanine, phenylglycine, cyclohexylalanine, .beta.-alanine, fluoro-amino acids, designer amino acids such as .beta.-methyl amino acids, Ca-methyl amino acids, Na-methyl amino acids, and amino acid analogs in general. Furthermore, the amino acid can be D (dextrorotary) or L (levorotary).
Oligonucleotide or polypeptide gene products can be used in a variety of methods to identify compounds that modulate the activity thereof. Nucleotide sequences and genes can be identified in different cell types and in the same cell type in which subject have different phenotypes. Methods are provided herein for screening compounds can include contacting cells with a compound and measuring gene expression levels, wherein a change in expression levels relative to a reference identifies the compound as a compound that modulates a gene expression.
Also provided herein are methods for identification and isolation of agents, such as compounds that bind to products of the genes listed herein. The assays are designed to identify agents that bind to the RNA or polypeptide gene product. The identified compounds are candidates or leads for identification of compounds for treatments of tumors and other disorders and diseases.
A variety of methods can be used, as known in the art. These methods can be performed in solution or in solid phase reactions.
Methods for identifying an agent, such as a compound, that specifically binds to an oligonucleotide or polypeptide encoded by a gene as listed herein also are provided. The method can be practiced by (a) contacting the gene product with one or a plurality of test agents under conditions conducive to binding between the gene product and an agent; and (b) identifying one or more agents within the one or plurality that specifically binds to the gene product. Compounds or agents to be identified can originate from biological samples or from libraries, including, but are not limited to, combinatorial libraries. Exemplary libraries can be fusion-protein-displayed peptide libraries in which random peptides or proteins are presented on the surface of phage particles or proteins expressed from plasmids; support-bound synthetic chemical libraries in which individual compounds or mixtures of compounds are presented on insoluble matrices, such as resin beads, or other libraries known in the art.
Modulators of the Activity of Gene products
Provided herein are compounds that modulate the activity of a gene product. These compounds can act by directly interacting with the polypeptide or by altering transcription or translation thereof. Such molecules include, but are not limited to, antibodies that specifically bind the polypeptide, antisense nucleic acids or double-stranded RNA (dsRNA) such as RNAi, that alter expression of the polypeptide, antibodies, peptide mimetics and other such compounds.
Antibodies are provided, including polyclonal and monoclonal antibodies that specifically bind to a polypeptide gene product provided herein. An antibody can be a monoclonal antibody, and the antibody can specifically bind to the polypeptide. The polypeptide and domains, fragments, homologs and derivatives thereof can be used as immunogens to generate antibodies that specifically bind such immunogens. Such antibodies include but are not limited to polyclonal, monoclonal, chimeric, single chain, Fab fragments, and an Fab expression library. In a specific embodiment, antibodies to human polypeptides are produced. Methods for monoclonal and polyclonal antibody production are known in the art. Antibody fragments that specifically bind to the polypeptide or epitopes thereof can be generated by techniques known in the art. For example, such fragments include but are not limited to: the F(ab′)2 fragment, which can be produced by pepsin digestion of the antibody molecule; the Fab′ fragments that can be generated by reducing the disulfide bridges of the F(ab′)2 fragment, the Fab fragments that can be generated by treating the antibody molecular with papain and a reducing agent, and Fv fragments.
Peptide analogs are commonly used in the pharmaceutical industry as non-peptide drugs with properties analogous to those of the template peptide. These types of non-peptide compounds are termed peptide mimetics or peptidomimetics (Luthman et al., A Textbook of Drug Design and Development, 14:386-406, 2nd Ed., Harwood Academic Publishers (1996); Joachim Grante (1994) Angew. Chem. Int. Ed. Engl., 33:1699-1720; Fauchere (1986) J. Adv. Drug Res., 15:29; Veber and Freidinger (1985) TINS, p. 392; and Evans et al. (1987) J. Med. Chem. 30:1229). Peptide mimetics that are structurally similar to therapeutically useful peptides can be used to produce an equivalent or enhanced therapeutic or prophylactic effect. Preparation of peptidomimetics and structures thereof are known to those of skill in this art.
Polypeptide products of the coding sequences (e.g., genes) listed herein can be detected in diagnostic methods, such as diagnosis of tumors and other diseases or disorders. Such methods can be used to detect, prognose, diagnose, or monitor various conditions, diseases, and disorders. Exemplary compounds that can be used in such detection methods include polypeptides such as antibodies or fragments thereof that specifically bind to the polypeptides listed herein, and oligonucleotides such as DNA probes or primers that specifically bind oligonucleotides such as RNA encoded by the nucleic acids provided herein.
A set of one or more, or two or more compounds for detection of markers containing a particular nucleotide sequence, complements thereof, fragments thereof, or polypeptides encoded thereby, can be selected for any of a variety of assay methods provided herein. For example, one or more, or two or more such compounds can be selected as diagnostic or prognostic indicators. Methods for selecting such compounds and using such compounds in assay methods such as diagnostic and prognostic indicator applications are known in the art. For example, the Tables provided herein list a modified t statistic associated with each marker, where the modified t statistic indicate the ability of the associated marker to indicate (by presence or absence of the marker, according to the modified t statistic) the presence or absence of a particular cell type in a prostate sample.
In another embodiment, marker selection can be performed by considering both modified t statistics and expected intensity of the signal for a particular marker. For example, markers can be selected that have a strong signal in a cell type whose presence or absence is to be determined, and also have a sufficiently large modified t statistic for gene expression in that cell type. Also, markers can be selected that have little or no signal in a cell type whose presence or absence is to be determined, and also have a sufficiently large negative modified t statistic for gene expression in that cell type.
Exemplary assays include immunoassays such as competitive and non-competitive assay systems using techniques such as western blots, radioimmunoassays, ELISA (enzyme linked immunosorbent assay), sandwich immunoassays, immunoprecipitation assays, precipitin reactions, gel diffusion precipitin reactions, immunodiffusion assays, agglutination assays, complement-fixation assays, immunoradiometric assays, fluorescent immunoassays and protein A immunoassays. Other exemplary assays include hybridization assays which can be carried out by a method by contacting a sample containing nucleic acid with a nucleic acid probe, under conditions such that specific hybridization can occur, and detecting or measuring any resulting hybridization.
Kits for diagnostic use are also provided, that contain in one or more containers an anti-polypeptide antibody, and, optionally, a labeled binding partner to the antibody. A kit is also provided that includes in one or more containers a nucleic acid probe capable of hybridizing to the gene-encoding nucleic acid. In a specific embodiment, a kit can include in one or more containers a pair of primers (e.g., each in the size range of 6-30 nucleotides) that are capable of priming amplification. A kit can optionally further include in a container a predetermined amount of a purified control polypeptide or nucleic acid.
The kits can contain packaging material that is one or more physical structures used to house the contents of the kit, such as invention nucleic acid probes or primers, and the like. The packaging material is constructed by well known methods, and can provide a sterile, contaminant-free environment. The packaging material has a label which indicates that the compounds can be used for detecting a particular oligonucleotide or polypeptide. The packaging materials employed herein in relation to diagnostic systems are those customarily utilized in nucleic acid or protein-based diagnostic systems. A package is to a solid matrix or material such as glass, plastic, paper, foil, and the like, capable of holding within fixed limits an isolated nucleic acid, oligonucleotide, or primer of the present invention. Thus, for example, a package can be a glass vial used to contain milligram quantities of a contemplated nucleic acid, oligonucleotide or primer, or it can be a microtiter plate well to which microgram quantities of a contemplated nucleic acid probe have been operatively affixed. The kits also can include instructions for use, which can include a tangible expression describing the reagent concentration or at least one assay method parameter, such as the relative amounts of reagent and sample to be admixed, maintenance time periods for reagent/sample admixtures, temperature, buffer conditions, and the like.
Pharmaceutical compositions containing the identified compounds that modulate expression of a gene or bind to a gene product are provided herein. Also provided are combinations of such a compound and another treatment or compound for treatment of a disease or disorder, such as a chemotherapeutic compound.
Expression modulator or binding compound and other compounds can be packaged as separate compositions for administration together or sequentially or intermittently. Alternatively, they can be provided as a single composition for administration or as two compositions for administration as a single composition. The combinations can be packaged as kits.
Compounds and compositions provided herein can be formulated as pharmaceutical compositions, for example, for single dosage administration. The concentrations of the compounds in the formulations are effective for delivery of an amount, upon administration, that is effective for the intended treatment. In certain embodiments, the compositions are formulated for single dosage administration. To formulate a composition, the weight fraction of a compound or mixture thereof is dissolved, suspended, dispersed or otherwise mixed in a selected vehicle at an effective concentration such that the treated condition is relieved or ameliorated. Pharmaceutical carriers or vehicles suitable for administration of the compounds provided herein include any such carriers known to those skilled in the art to be suitable for the particular mode of administration.
In addition, the compounds can be formulated as the sole pharmaceutically active ingredient in the composition or can be combined with other active ingredients. The active compound is included in the pharmaceutically acceptable carrier in an amount sufficient to exert a therapeutically useful effect in the absence of undesirable side effects on the subject treated. The therapeutically effective concentration can be determined empirically by testing the compounds in known in vitro and in vivo systems. The concentration of active compound in the drug composition depends on absorption, inactivation and excretion rates of the active compound, the physicochemical characteristics of the compound, the dosage schedule, and amount administered as well as other factors known to those of skill in the art. Pharmaceutically acceptable derivatives include acids, salts, esters, hydrates, solvates and prodrug forms. The derivative can be selected such that its pharmacokinetic properties are superior to the corresponding neutral compound. Compounds are included in an amount effective for ameliorating or treating the disorder for which treatment is contemplated.
Formulations suitable for a variety of administrations such as perenteral, intramuscular, subcutaneous, alimentary, transdermal, inhaling and other known methods of administration, are known in the art. The pharmaceutical compositions can also be administered by controlled release means and/or delivery devices as known in the art. Kits containing the compositions and/or the combinations with instructions for administration thereof are provided. The kit can further include a needle or syringe, which can be packaged in sterile form, for injecting the complex, and/or a packaged alcohol pad. Instructions are optionally included for administration of the active agent by a clinician or by the patient.
The compounds can be packaged as articles of manufacture containing packaging material, a compound or suitable derivative thereof provided herein, which is effective for treatment of a diseases or disorders contemplated herein, within the packaging material, and a label that indicates that the compound or a suitable derivative thereof is for treating the diseases or disorders contemplated herein. The label can optionally include the disorders for which the therapy is warranted.
The compounds provided herein can be used for treating or preventing diseases or disorders in an animal, such as a mammal, including a human. In one embodiment, the method includes administering to a mammal an effective amount of a compound that modulates the expression of a particular gene (e.g., a gene listed herein) or a compound that binds to a product of a gene, whereby the disease or disorder is treated or prevented. Exemplary inhibitors provided herein are those identified by the screening assays. In addition, antibodies and antisense nucleic acids or double-stranded RNA (dsRNA), such as RNAi, are contemplated.
In a specific embodiment, as described hereinabove, gene expression can be inhibited by antisense nucleic acids. The therapeutic or prophylactic use of nucleic acids of at least six nucleotides, up to about 150 nucleotides, that are antisense to a gene or cDNA is provided. The antisense molecule can be complementary to all or a portion of the gene. For example, the oligonucleotide is at least 10 nucleotides, at least 15 nucleotides, at least 100 nucleotides, or at least 125 nucleotides. The oligonucleotides can be DNA or RNA or chimeric mixtures or derivatives or modified versions thereof, single-stranded or double-stranded. The oligonucleotide can be modified at the base moiety, sugar moiety, or phosphate backbone. The oligonucleotide can include other appending groups such as peptides, or agents facilitating transport across the cell membrane, hybridization-triggered cleavage agents or intercalating agents.
RNA interference (RNAi) (see, e.g., Chuang et al. (2000) Proc. Natl. Acad. Sci. U.S.A. 97:4985) can be employed to inhibit the expression of a nucleic acid. Interfering RNA (RNAi) fragments, such as double-stranded (ds) RNAi, can be used to generate loss-of-gene function. Methods relating to the use of RNAi to silence genes in organisms including, mammals, C. elegans, Drosophila and plants, and humans are known. Double-stranded RNA (dsRNA)-expressing constructs are introduced into a host, such as an animal or plant using, a replicable vector that remains episomal or integrates into the genome. By selecting appropriate sequences, expression of dsRNA can interfere with accumulation of endogenous mRNA. RNAi also can be used to inhibit expression in vitro. Regions include at least about 21 (or 21) nucleotides that are selective (i.e., unique) for the selected gene are used to prepare the RNAi. Smaller fragments of about 21 nucleotides can be transformed directly (i.e., in vitro or in vivo) into cells; larger RNAi dsRNA molecules can be introduced using vectors that encode them. dsRNA molecules are at least about 21 bp long or longer, such as 50, 100, 150, 200 and longer. Methods, reagents and protocols for introducing nucleic acid molecules in to cells in vitro and in vivo are known to those of skill in the art.
In an exemplary embodiment, nucleic acids that include a sequence of nucleotides encoding a polypeptide of a gene as listed herein can be administered to promote polypeptide function, by way of gene therapy. Gene therapy refers to therapy performed by administration of a nucleic acid to a subject. In this embodiment, the nucleic acid produces its encoded protein that mediates a therapeutic effect by promoting polypeptide function. Any of the methods for gene therapy available in the art can be used (see, Goldspiel et al., Clinical Pharmacy 12:488-505 (1993); Wu and Wu, Biotherapy 3:87-95 (1991); Tolstoshev, An. Rev. Pharmacol. Toxicol. 32:573-596 (1993); Mulligan, Science 260:926-932 (1993); and Morgan and Anderson, An. Rev. Biochem. 62:191-217 (1993); TIBTECH 11 (5):155-215 (1993).
In some embodiments, vaccines based on the genes and polypeptides provided herein can be developed. For example genes can be administered as DNA vaccines, either single genes or combinations of genes. Naked DNA vaccines are generally known in the art. Methods for the use of genes as DNA vaccines are well known to one of ordinary skill in the art, and include placing a gene or portion of a gene under the control of a promoter for expression in a patient with cancer. The gene used for DNA vaccines can encode full-length proteins, but can encode portions of the proteins including peptides derived from the protein. For example, a patient can be immunized with a DNA vaccine comprising a plurality of nucleotide sequences derived from a particular gene. In another embodiment, it is possible to immunize a patient with a plurality of genes or portions thereof. Without being bound by theory, expression of the polypeptide encoded by the DNA vaccine, cytotoxic T-cells, helper T-cells and antibodies are induced that recognize and destroy or eliminate cells expressing the proteins provided herein.
DNA vaccines can include a gene encoding an adjuvant molecule with the DNA vaccine. Such adjuvant molecules include cytokines that increase the immunogenic response to the polypeptide encoded by the DNA vaccine. Additional or alternative adjuvants are known to those of ordinary skill in the art and find use in the invention.
Also provided herein, the nucleotide the genes, nucleotide molecules and polypeptides disclosed herein find use in generating animal models of cancers, such as lymphomas and carcinomas. As is appreciated by one of ordinary skill in the art, when one of the genes provided herein is repressed or diminished, gene therapy technology wherein antisense RNA directed to the gene will also diminish or repress expression of the gene. An animal generated as such serves as an animal model that finds use in screening bioactive drug candidates. In another embodiment, gene knockout technology, for example as a result of homologous recombination with an appropriate gene targeting vector, will result in the absence of the protein. When desired, tissue-specific expression or knockout of the protein can be accomplished using known methods.
It is also possible that a protein is overexpressed in cancer. As such, transgenic animals can be generated that overexpress the protein. Depending on the desired expression level, promoters of various strengths can be employed to express the transgene. Also, the number of copies of the integrated transgene can be determined and compared for a determination of the expression level of the transgene. Animals generated by such methods find use as animal models and are additionally useful in screening for bioactive molecules to treat cancer.
The various techniques, methods, and aspects of the methods provided herein can be implemented in part or in whole using computer-based systems and methods. In another embodiment, computer-based systems and methods can be used to augment or enhance the functionality described above, increase the speed at which the functions can be performed, and provide additional features and aspects as a part of or in addition to those of the invention described elsewhere in this document. Various computer-based systems, methods and implementations in accordance with the above-described technology are presented below.
A processor-based system can include a main memory, such as random access memory (RAM), and can also include a secondary memory. The secondary memory can include, for example, a hard disk drive and/or a removable storage drive, representing a floppy disk drive, a magnetic tape drive, or an optical disk drive. The removable storage drive reads from and/or writes to a removable storage medium. Removable storage medium refers to a floppy disk, magnetic tape, optical disk, and the like, which is read by and written to by a removable storage drive. As will be appreciated, the removable storage medium can comprise computer software and/or data.
In alternative embodiments, the secondary memory may include other similar means for allowing computer programs or other instructions to be loaded into a computer system. Such means can include, for example, a removable storage unit and an interface. Examples of such can include a program cartridge and cartridge interface (such as the found in video game devices), a movable memory chip (such as an EPROM or PROM) and associated socket, and other removable storage units and interfaces, which allow software and data to be transferred from the removable storage unit to the computer system.
The computer system can also include a communications interface. Communications interfaces allow software and data to be transferred between computer system and external devices. Examples of communications interfaces can include a modem, a network interface (such as, for example, an Ethernet card), a communications port, a PCMCIA slot and card, and the like. Software and data transferred via a communications interface are in the form of signals, which can be electronic, electromagnetic, optical or other signals capable of being received by a communications interface. These signals are provided to communications interface via a channel capable of carrying signals and can be implemented using a wireless medium, wire or cable, fiber optics or other communications medium. Some examples of a channel can include a phone line, a cellular phone link, an RF link, a network interface, and other communications channels.
In this document, the terms computer program medium and computer usable medium are used to refer generally to media such as a removable storage device, a disk capable of installation in a disk drive, and signals on a channel. These computer program products are means for providing software or program instructions to a computer system.
Computer programs (also called computer control logic) are stored in main memory and/or secondary memory. Computer programs can also be received via a communications interface. Such computer programs, when executed, permit the computer system to perform the features of the invention as discussed herein. In particular, the computer programs, when executed, permit the processor to perform the features of the invention. Accordingly, such computer programs represent controllers of the computer system.
In an embodiment where the elements are implemented using software, the software may be stored in, or transmitted via, a computer program product and loaded into a computer system using a removable storage drive, hard drive or communications interface. The control logic (software), when executed by the processor, causes the processor to perform the functions of the invention as described herein.
In another embodiment, the elements are implemented in hardware using, for example, hardware components such as PALs, application specific integrated circuits (ASICs) or other hardware components Implementation of a hardware state machine so as to perform the functions described herein will be apparent to person skilled in the relevant art(s). In yet another embodiment, elements are implanted using a combination of both hardware and software.
In another embodiment, the computer-based methods can be accessed or implemented over the World Wide Web by providing access via a Web Page to the methods of the invention. Accordingly, the Web Page is identified by a Universal Resource Locator (URL). The URL denotes both the server machine and the particular file or page on that machine. In this embodiment, it is envisioned that a consumer or client computer system interacts with a browser to select a particular URL, which in turn causes the browser to send a request for that URL or page to the server identified in the URL. The server can respond to the request by retrieving the requested page and transmitting the data for that page back to the requesting client computer system (the client/server interaction can be performed in accordance with the hypertext transport protocol (HTTP)). The selected page is then displayed to the user on the client's display screen. The client may then cause the server containing a computer program of the invention to launch an application to, for example, perform an analysis according to the methods provided herein.
Provided herein are probe and gene sequences that can be indicative of the presence and/or absence of prostate cancer in a subject. Also provided herein are probe and gene sequences that can be indicative of presence and/or absence of benign prostatic hyperplasia (BPH) in a subject. Also provided herein are probe and gene sequences that can be indicative of a prognosis of prostate cancer, where such a prognosis can include likely relapse of prostate cancer, likely aggressiveness of prostate cancer, likely indolence of prostate cancer, likelihood of survival of the subject, likelihood of success in treating prostate cancer, condition in which a particular treatment regimen is likely to be more effective than another treatment regimen, and combinations thereof. In one embodiment, the probe and gene sequences can be indicative of the likely aggressiveness or indolence of prostate cancer.
As provided in the methods and Tables herein, probes have been identified that hybridize to one or more nucleic acids of a prostate sample at different levels according to the presence or absence of prostate tumor, BPH and stroma in the sample. The probes provided herein are listed in conjunction with modified t statistics that represent the ability of that particular probe to indicate the presence or absence of a particular cell type in a prostate sample. Use of modified t statistics for such a determination is described elsewhere herein, and general use of modified t statistics is known in the art. Accordingly, provided herein are nucleotide sequences of probes that can be indicative of the presence or absence of prostate tumor and/or BPH cells, and also can be indicative of the likelihood of prostate tumor relapse in a subject.
Also provided in the methods and Tables herein are nucleotide and predicted amino acid sequences of genes and gene products associated with the probes provided herein. Accordingly, as provided herein, detection of gene products (e.g., mRNA or protein) or other indicators of gene expression, can be indicative of the presence or absence of prostate tumor and/or BPH cells, and also can be indicative of the likelihood of prostate tumor relapse in a subject. As with the probe sequences, the nucleotide and amino acid sequences of these gene products are listed in conjunction with modified t statistics that represent the ability of that particular gene product or indicator thereof to indicate the presence or absence of a particular cell type in a prostate sample.
Methods for determining the presence of prostate tumor and/or BPH cells, the likelihood of prostate tumor relapse in a subject, the likelihood of survival of prostate cancer, the aggressiveness of prostate tumor, the indolence of prostate tumor, survival, and other prognoses of prostate tumor, can be performed in accordance with the teachings and examples provided herein. Also provided herein, a set of probes or gene products can be selected according to their modified t statistic for use in combination (e.g., for use in a microarray) in methods of determining the presence of prostate tumor and/or BPH cells, and/or the likelihood of prostate tumor relapse in a subject.
Also provided herein, the gene products identified as present at increased levels in prostate cancer or in subjects with likely relapse of cancer, can serve as targets for therapeutic compounds and methods. For example an antibody or siRNA targeted to a gene product present at increased levels in prostate cancer can be administered to a subject to decrease the levels of that gene product and to thereby decrease the malignancy of tumor cells, the aggressiveness of a tumor, indolence of a tumor, survival, or the likelihood of tumor relapse. Methods for providing molecules such as antibodies or siRNA to a subject to decrease the level of gene product in a subject are provided herein or are otherwise known in the art.
In some embodiments, gene products identified as present at decreased levels in prostate cancer or in subjects with likely relapse of cancer, can serve as subjects for therapeutic compounds and methods. For example a nucleic acid molecule, such as a gene expression vector encoding a particular gene, can be administered to a individual with decreased levels of the particular gene product to increase the levels of that gene product and to thereby decrease the malignancy of tumor cells, the aggressiveness of a tumor, indolence of a tumor, likelihood of survival, or the likelihood of tumor relapse. Methods for providing gene expression vectors to a subject to increase the level of gene product in a subject are provided herein or are otherwise known in the art.
As used herein, the term “prostate cancer signature” refers to genes that exhibit altered expression (e.g., increased or decreased expression) with prostate cancer as compared to control levels of expression (e.g., in normal prostate tissue). Genes included in a prostate cancer signature can include any of those listed in the tables presented herein (e.g., Tables 3 and 4). For example, one or more (e.g., two, three, four, five, six, seven, eight nine, ten, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, or more) of the genes listed in Table 3 can be are present in a prostate tissue sample (e.g., a prostate tissue sample containing normal stroma, prostate cancer cells, or both) at a level greater than or less than the level observed in normal, non-cancerous prostate tissue. In some cases, a prostate cancer signature can be a gene expression profile in which at least 10, 20, 30, 40, 50, 60, 70, 80, 90, or 100 percent of the genes listed in a table herein (e.g., Table 3 or Table 4) are expressed at a level greater than or less than their corresponding control levels in non-cancerous tissue.
As used herein, the terms “prostate cell-type predictor” genes and “prostate tissue predictor” genes refer to genes that can, based on their expression levels, serve as indicators as to whether a particular sample of prostate tissue contains particular cell types (e.g., prostate cancer cells, normal stromal cells, epithelial cells of benign prostate hyperplasia, or epithelial cells of dilated cystic glands). Such genes also can indicate the relative amounts of such cell types within the prostate tissue sample.
In some embodiments, this document features methods for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring the level of expression for prostate cancer signature genes in the sample; (c) comparing the measured expression levels to reference expression levels for the prostate cancer signature genes; and (d) if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having prostate cancer, and if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as not having prostate cancer. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in the Tables herein (e.g., in Table 3 or Table 4). The method can include determining whether measured expression levels for ten or more prostate cancer signature genes are significantly greater or less than reference expression levels for the ten or more prostate cancer signature genes, and classifying the subject as having prostate cancer that is likely to relapse if the measured expression levels are significantly greater or less than the reference expression levels, or classifying the subject as having prostate cancer not likely to relapse if the measured expression levels are not significantly greater or less than the reference expression levels. The ten or more prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein, for example. The method can include determining whether measured expression levels for twenty or more prostate cancer signature genes are significantly greater or less than reference expression levels for the twenty or more prostate cancer signature genes, and classifying the subject as having prostate cancer that is likely to relapse if the measured expression levels are significantly greater or less than the reference expression levels, or classifying the subject as having prostate cancer not likely to relapse if the measured expression levels are not significantly greater or less than the reference expression levels. The twenty or more prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein, for example.
This document also features methods for determining the prognosis of a subject diagnosed as having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring the level of expression for prostate cancer signature genes in the sample; (c) comparing the measured expression levels to reference expression levels for the prostate cancer signature genes; and (d) if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as having a relatively better prognosis than if the measured expression levels are significantly greater or less than the reference expression levels, or if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having a relatively worse prognosis than if the measured expression levels are not significantly greater or less than the reference expression levels. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in the Tables herein (e.g., Table 8A or 8B).
In addition, this document provides methods for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject, wherein the sample comprises prostate stromal cells; (b) measuring expression levels for one or more genes in the stromal cells, wherein the one or more genes are prostate cancer signature genes; (c) comparing the measured expression levels to reference expression levels for the one or more genes, wherein the reference expression levels are determined in stromal cells from non-cancerous prostate tissue; and (d) if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having prostate cancer, and if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as not having prostate cancer. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in Table 3 or Table 4 herein, for example.
This document also provides methods for determining a prognosis for a subject diagnosed as having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject, wherein the sample comprises prostate stromal cells; (b) measuring expression levels for one or more genes in the stromal cells, wherein the one or more genes are prostate cancer signature genes; (c) comparing the measured expression levels to reference expression levels for the one or more genes, wherein the reference expression levels are determined in stromal cells from non-cancerous prostate tissue; and (d) if the measured expression levels are not significantly greater or less than the reference expression levels, identifying the subject as having a relatively better prognosis than if the measured expression levels are significantly greater or less than the reference expression levels, or if the measured expression levels are significantly greater or less than the reference expression levels, identifying the subject as having a relatively worse prognosis than if the measured expression levels are not significantly greater or less than the reference expression levels. The prostate tissue sample may not include tumor cells, or the prostate tissue sample may include tumor cells and stromal cells. The prostate cancer signature genes can be selected from the genes listed in the tables herein (e.g., Table 3 or Table 4).
Further, this document features a method for identifying a subject as having or not having prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring expression levels for one or more prostate cell-type predictor genes in the sample; (c) determining the percentages of tissue types in the sample based on the measured expression levels; (d) measuring expression levels for one more prostate cancer signature genes in the sample; (e) determining a classifier based on the percentages of tissue types and the measured expression levels; and (f) if the classifier falls into a predetermined range of prostate cancer classifiers, identifying the subject as having prostate cancer, or if the classifier does not fall into the predetermined range, identifying the subject as not having prostate cancer. Steps (b) and (d) can be carried out simultaneously.
This document also features a method for determining a prognosis for a subject diagnosed with and treated for prostate cancer, comprising: (a) providing a prostate tissue sample from the subject; (b) measuring expression levels for one or more prostate tissue predictor genes in the sample; (c) determining the percentages of tissue types in the sample based on the measured expression levels; (d) measuring expression levels for one more prostate cancer signature genes in the sample; (e) determining a classifier based on the percentages of tissue types and the measured expression levels; and (f) if the classifier falls into a predetermined range of prostate cancer relapse classifiers, identifying the subject as being likely to relapse, or if the classifier does not fall into the predetermined range, identifying the subject as not being likely to relapse. Steps (b) and (d) are carried out simultaneously.
In some embodiments, methods as described herein can be used for identifying the proportion of two or more tissue types in a tissue sample. Such methods can include, for example: (a) using a set of other samples of known tissue proportions from a similar anatomical location as the tissue sample in an animal or plant, wherein at least two of the other samples do not contain the same relative content of each of the two or more cell types; (b) measuring overall levels of one or more gene expression or protein analytes in each of the other samples; (c) determining the regression relationship between the relative proportion of each tissue type and the measured overall levels of each gene expression or protein analyte in the other samples; (d) selecting one or more analytes that correlate with tissue proportions in the other samples; (e) measuring overall levels of one or more of the analytes in step (d) in the tissue sample; (f) matching the level of each analyte in the tissue sample with the level of the analyte in step (d) to determine the predicted proportion of each tissue type in the tissue sample; and (g) selecting among predicted tissue proportions for the tissue sample obtained in step (f) using either the median or average proportions of all the estimates. The tissue sample can contain cancer cells (e.g., prostate cancer cells).
Methods described herein can be used for comparing the levels of two or more analytes predicted by one or more methods to be associated with a change in a biological phenomenon in two sets of data each containing more than one measured sample. Such methods can comprise: (a) selecting only analytes that are assayed in both sets of data; (b) ranking the analytes in each set of data using a comparative method such as the highest probability or lowest false discovery rate associated with the change in the biological phenomenon; (c) comparing a set of analytes in each ranked list in step (b) with each other, selecting those that occur in both lists, and determining the number of analytes that occur in both lists and show a change in level associated with the biological phenomenon that is in the same direction; and (d) calculating a concordance score based on the probability that the number of comparisons would show the observed number of change in the same direction, at random. In step (a), the length of each list can be varied to determine the maximum concordance score for the two ranked lists.
The invention will be further described in the following examples, which do not limit the scope of the invention described in the claims.
Over one million prostate biopsies are performed in the U.S. every year. Pathology examination is not definitive in a significant percentage of cases, however, due to the presence of equivocal structures or continuing clinical suspicion. To investigate gene expression changes in the tumor microenvironment vs. normal stroma, gene expression profiles from 15 volunteer biopsy specimens were compared to profiles from 13 specimens containing largely tumor-adjacent stroma. As described below, more than a thousand significant expression changes were identified and filtered to eliminate possible age-related genes, as well as genes that also are expressed at detectable levels in tumor cells. A stroma-specific classifier was constructed based on the 114 remaining unique candidate genes (131 Affymetrix probe sets). The classifier was tested on 380 independent cases, including 255 tumor-bearing cases and 125 non-tumor cases (normal biopsies, normal autopsies, remote stroma as well as pure tumor adjacent stroma). The classifier predicted the tumor status of patients with an average accuracy of 97.4% (sensitivity=98.0% and specificity=89.7%), whereas a randomly generated and trained classifier had no diagnostic value. These results indicate that the prostate cancer microenvironment exhibits reproducible changes useful for categorizing stroma as “presence of tumor” and “non-presence of tumor.”
Prostate Cancer Patients Samples and Expression Analysis:
Datasets 1 and 2 (Table 1) were obtained using post-prostatectomy frozen tissue samples. All tissues, except where noted, were collected at surgery and escorted to pathology for expedited review, dissection, and snap freezing in liquid nitrogen. RNA for expression analysis was prepared directly from frozen tissue following dissection of OCT (optimum cutting temperature compound) blocks with the aid of a cryostat. For expression analysis, 50 micrograms (10 micrograms for biopsy tissue) of total RNA samples were processed for hybridization to Affymetrix GeneChips.
Dataset 1 consists of 109 post-prostatectomy frozen tissue samples from 87 patients. Twenty-two cases were analyzed twice using one sample from a tumor-enriched specimen and one sample from a non-tumor specimen (more than 1.5 cm away from the tumor), usually the contralateral lobe. In addition, Dataset 1 contains 27 prostate biopsy specimens obtained as fresh snap frozen biopsy cores from 18 normal participants in a clinical trial to evaluate the role of Difluoromethylornithine (DFMO) to decrease the prostate size of normal men (Simoneau et al. (2008) Cancer Epidemiol. Biomarkers Prev. 17:292-299). Finally, Dataset 1 contains 13 cases of normal prostates obtained from the rapid autopsy program of the Sun Health Research Institute, from subjects with an average age of 82 years.
Dataset 2 contains 136 samples from 82 patients, where 54 cases were analyzed as pairs of tumor-enriched samples and, for most cases, non-tumor tissue obtained from the same OCT block as tumor-adjacent tissue. This series includes specimens for which expression coefficients were validated (Stuart et al. (2004) Proc. Natl. Acad. Sci. U.S.A. 101:615-620).
Expression analysis for Datasets 1 and 2 was carried out using Affymetrix U133Plus2 and U133A GeneChips, respectively; the expression data are publicly available at GEO database on the World Wide Web at ncbi.nlm.nih.gov/geo, with accession numbers GSE17951 (Dataset 1) and GSE8218 (Dataset 2). For both datasets, cell type distributions for the four principal cell types (tumor epithelial cells, stroma cells, epithelial cells of BPH, and epithelial cells of dilated cystic glands) were determined from frozen sections prepared immediately before and after the sections pooled for RNA preparation by three (Dataset 1) or four (Dataset 2) pathologists whose estimates were averaged as described (Stuart et al., supra). The distributions of tumor percentage for Dataset 1 and 2 are shown in
Dataset 3 consists of a published series (Stephenson et al. (2005) Cancer 104:290-298) of 79 cases for which expression data were measured with Affymetrix U133A chips. The cell composition was not documented at the time of data collection. Cell composition was estimated using multigene signatures that are invariant with tumor surgical pathology parameters of Gleason and stage by the CellPred program (World Wide Web at webarraydb.org), which confirmed that all 79 samples included tumor cells, with tumor content ranging from 24% to 87% (
Dataset 4 includes 57 samples from 44 patients, including 13 tumor-adjacent stroma samples and 44 tumor-bearing samples. Gene expression in these 57 samples was measured with Affymetrix U133A GeneChips. Tumor percentage (ranging from 0% to 80%,
Dataset 5 consists of 4 pooled normal stromal samples and 12 tumor samples gleaned by Laser Capture Micro dissection (LCM) using frozen tissue samples. Each pooled normal stroma sample was pooled from two LCM captured stroma samples from specimens from which no tumor was recovered in the surgical samples available for the research protocol described herein, whereas tumor samples were LCM-captured prostate cancer cells. Gene expression in these 16 samples (using 10 micrograms of total RNA) was measured using Affymetrix U133Plus2 chips.
Compared to U133A (with ˜22,000 probe sets) used for Datasets 2, 3 and 4, the U133Plus2 platform used for Datasets 1 and 5 had about 30,000 more probe sets. To attain an analysis across multiple datasets, only the probes common to these two platforms were used, i.e., only about 22,000 common probe sets in each Dataset were considered. First, Dataset 1 was quantile-normalized using function ‘normalizeQuantiles( )’ of LIMMA routine (Dalgaard (2002) Statistics and Computing: Introductory Statistics with R, p. 260, Springer-Verlag Inc., New York. Datasets 2-5 were then quantile-normalized by referencing normalized Dataset 1 with a modified function ‘REFnormalizeQuantiles( ),’ which is available from ZJ.
1P, B, A, and L represent patient, normal biopsy, normal rapid autopsy, and LCM, respectively. Datasets 1 and 2 were collected from five participating institutions in San Diego County, CA. Demographic, Pathology, and clinical values are individually recorded (Shadow charts) and maintained in the UCI SPECS consortium database including tracking sheets of elapsed times following surgery during sample handling.
Statistical Tools Implemented in R.:
The Linear Models for Microarray Data (LIMMA package from Bioconductor, on the World Wide Web at bioconductor.org) was used to detect differentially expressed genes. Prediction Analysis of Microarray (PAM, implemented by the PAMR package from Bioconductor) was used to develop an expression-based classifier from training set and then applied to the test sets without any change (Guo et al. (2007) Biostatistics 8:86-100). Fisher's Exact Test was used to demonstrate the efficiency of the classifier when it was tested on remote stroma versus tumor adjacent stroma. Fisher's test was used instead of chi-square because chi-square test is not suitable when the expected values in any of the cells of the table are below 10. All statistical analysis was done using R language (World Wide Web at r-project.org).
Multiple Linear Regression Model:
A multiple linear regression (MLR) model was used to describe the observed Affymetrix intensity of a gene as the summation of the contributions from different types of cells given the pathological cell constitution data:
where g is the expression value for a gene, p is the percentage data determined by the pathologists, and β's are the expression coefficients associated with different cell types. In model (1), C is the number of tissue types under consideration. In the present case, three major tissue types were included, i.e., tumor, stroma, and BPH. βj is the estimate of the relative expression level in cell type j (i.e., the expression coefficient) compared to the overall mean expression level β0. The regression model was applied to the patient cases in Dataset 1 to obtain the model parameters (β's) and their corresponding p-values, which were used to aid subsequent gene screening. The application to prostate cancer expression data and validation by immunohistochemistry and by correlation of derived βj values with LCM-derived samples assayed by qPCR has been described (Stuart et al., supra).
Identification of Stroma-Derived Genes and Development of the Diagnostic Classifier:
It was hypothesized that stroma within and directly adjacent to prostate cancer epithelial cell formations of infiltrating tumors exhibit significant RNA expression changes compared to normal prostate stroma. To obtain an initial comparison of tumor-adjacent stroma to normal stroma, normal fresh frozen biopsy tissue was used as a source of normal stroma. Out of 27 normal biopsy samples, 15 were selected from 15 different participants. The remaining 12 biopsy samples were reserved for testing. Gene expression microarray data were obtained and compared to 13 tumor-bearing patient cases from Dataset 1 selected to tumor (T) greater than 0% but less than 10% tumor cell content (the average stroma content is ˜80%). These criteria ensured that the majority of stroma tissues included were close to tumor, while T<10% ensures that the impact from tumor cells was minimal since the aim was to capture altered expression signals from stroma cells rather than from tumor cells.
As the number of biopsies available was limited, a permutation strategy was adopted to maximize their use. First 13 of the 15 normal biopsy samples were selected and their gene expression was compared to the 13 tumor-adjacent stroma samples using the moderated t-test implemented in the LIMMA package of R (Dalgaard, supra). This comparison yielded 3888 expression changes between these two groups with a p value <0.05.
A substantial difference in age existed between the normal stroma group (average age=51.9 years) and the tumor-adjacent stroma group (average age=60.6 years). The overall gene expression of the 13 normal stroma samples used for training was compared to that of 13 normal prostate specimens obtained from the rapid autopsy program (see above), with an average age of 82 years. The comparison revealed 8898 significant expression changes (p<0.05), of which 2210 also were detected in the comparison of normal stroma samples between tumor-adjacent stroma (
A potential issue related to using patient cases with 10%>T>0% was that the detected expression changes may have included expression changes specific to tumor cells or epithelium cells rather than only to stroma cells. To reduce the possibility that epithelial-cell derived expression changes dominated, a secondary gene screening via MLR analysis was used. MLR was used to determine cell-specific gene expression based on “knowledge” of the percent cell composition of the samples of Dataset 1 as determined by a panel of four pathologists (Stuart et al., supra; the distribution is shown in
Thus from the 1678 genes of the initial analysis, 160 candidate probe sets with three criteria were selected: (1) βs<0, (2) βs<10×βTβS>10×βT, and (3) p (βs)<0.1. When the values of the βs's were compared to the Ns, it became apparent that the expression levels of these 160 probe sets in stroma cells were substantially higher than in tumor cells (
The second step for the permutation analysis was then carried out. The above procedure was repeated using a different selections of 13 biopsy samples of the 15 until all 105 possible combinations of 13 normal biopsy samples drawn from 15 (C1513=105, where Cnm is the number of combinations of m elements chosen from a total of n elements) was complete. A total of 339 probe sets (Table 3) were generated by the 105-fold gene selection procedure with a frequency of selection as summarized in
Probe sets with at least 50 occurrences (about 50%) of the 105-fold permutation were selected for classifier construction. Prediction Analysis for Microarrays (PAM; Tibshirani et al. (2002) Proc. Natl. Acad. Sci. U.S.A. 99:6567-6572) was used to build a diagnostic classifier. The training set (Table 2, line 1) included all 15 normal biopsies and the 13 tumor-adjacent stroma samples that were used for the derivation of significant differences. Of the 146 PAM-input probe sets, 131 were retained following the 10-fold cross validation procedure of PAM, leading to a prediction accuracy of 96.4%. The separation of normal and tumor-adjacent stroma cases of the training set by the Classifier is illustrated into two distinct populations is shown in
Testing with Independent Datasets:
The 131-element classifier was then tested on numerous prostate samples not used for training, including 55 tumor-bearing cases from Dataset 1 and 65 tumor-bearing cases from Dataset 2. Also included were two additional datasets of 79 tumor-bearing cases (Dataset 3) and 44 tumor-bearing cases (Dataset 4), where both the samples and expression analyses were from separate institutes (Table 1). These four test sets were composed entirely of tumor bearing samples (Table 2, lines 2 to 5). In all four tests, almost all samples (n=243) were recognized as “tumor” with high average accuracy ˜99%.
The classifier also was tested using specimens composed mainly of normal prostate stroma and epithelium. First, the classifier was tested on the 12 remaining biopsies from the DMFO study which were separated into two groups. Group 1 (Table 2, line 6) included second biopsies of the same participants whose first biopsy samples were included in the training set, and therefore are not completely independent cases. Group 2 (Table 2, line 7) included the five biopsy samples of cases not used for training. These samples were devoid of tumor but contained normal epithelial components, typically ranging from ˜35% to ˜45%. Microarray data were obtained for these 12 cases and used for testing. The biopsy samples in group 1 were accurately (100%) identified as non-tumor. For group 2, two out of five biopsy samples were categorized as “presence of tumor.” When the histories for these cases were consulted, however, it was found that both had consistently exhibited elevated PSA levels of 6.1, 9.6, and 8 ng/ml (normal values <3 ng/ml), respectively, although no tumor was observed in either of two sets of sextant biopsies obtained from these cases. All other donors of normal biopsies exhibited normal PSA values. The classifier was then tested on 13 specimens obtained by rapid autopsy of individuals dying of unrelated causes (Table 2, line 8). Twelve out of these 13 cases (i.e., 92.3%), were classified as nontumor. Histological examination of all embedded tissue of the two “misclassified” cases revealed multiple foci of small “latent” tumors. The 25 samples which were drawn from normal tissues were correctly classified as having no tumor present, or were classified in accordance with abnormal features that were subsequently uncovered. These results provide further support for the ability of the classifier to discriminate between normal and abnormal prostate tissues in the absence of histologically recognizable tumor cells in the samples studied.
Validation by Manual Microdissection and LCM of Tumor-Adjacent and Remote Stroma:
Based on the strong performance with mixed tissue test samples, experiments were conducted to validate the classifier by developing histologically confirmed pure tumor-adjacent stroma samples. Tumor-bearing tissue mounted in OCT blocks in a cryostat were examined by frozen section to visualize the location of the tumor. The OCT-embedded block was etched with a single straight cut with a scalpel to divide the embedded tissue into a tumor zone and tumor-adjacent stroma. Subsequent cryosections were separated into two halves and used for H and E staining to confirm their composition. For sections of tumor-adjacent stroma with a large area (i.e., ˜10 mm2), multiple frozen sections were pooled and used for RNA preparation and microarray hybridization. A final frozen section was stained and examined to confirm that it was free of tumor cells. For smaller areas of the tumor-adjacent zone, the adjacent tissue was removed as a piece, remounted in reverse orientation and a final frozen section was made to confirm that the piece was free of tumor cells. This tissue was then used for RNA preparation and expression analysis.
Seventy-one tumor-adjacent stroma samples were obtained from the samples of Dataset 2, 13 from the samples of Dataset 4, and 12 from the samples of Dataset 1, using the manual microdissection method. These tumor-adjacent stroma samples were then used for expression analysis. The expression values for the 131 classifier probe sets were tested using the PAM procedure. Accuracies of 97.1%, 100%, and 75% were observed for the classification as “presence of tumor” (Table 2, lines 9-11). These results indicate an overall accuracy of 94.7% for the 96 independent samples.
Finally, examined laser capture microdissected samples were prepared from the samples of Dataset 5. Twelve tumor cell samples were prepared as 100% prostate cancer cells, while four pooled stroma control samples were prepared from cases where no tumor had been recovered in the surgical samples available for the research protocol. These samples were categorized by the classifier as 100% “presence of tumor” and 100% “no presence of tumor,” respectively.
Since several cases (especially from Dataset 1) appeared “misclassified,” it was of interest to know how far from a known tumor site the expression changes characteristic of tumor stroma may extend. There was insufficient tissue for a systematic analysis of samples at various known distances, but 28 cases from Dataset 1 were available that were greater than 1.5 cm from the tumor sites of the same gland and generally were from the contralateral lobe of the donor gland. Array data was collected from all pieces and categorized by the classifier. Only ten of the 28 samples (35.7%) were categorized as tumor-associated stroma. This distribution of classifications was compared to the distribution for the original 12 tumor-adjacent stroma samples manually prepared from samples of Dataset 1 (Table 2, line 11) using the Fisher Exact Test. The distribution for the 28 “remote” samples was significantly different than the category distribution for the 12 authentic tumor-adjacent stroma samples of the same cases as judged by a Fischer Exact test, p=0.038. This result strongly suggests that the expression changes of tumor-adjacent stroma are not inevitable in stroma taken from arbitrary sites of the same tumor-bearing glands, and likely reflect that proximity to tumor affects the expression changes of the genes of the classifier developed here.
Comparison with Random-Gene Classifiers:
To further validate the 131-element diagnostic classifier, 100 randomized experiments were carried out. In each experiment, 1,700 probe sets were randomly selected from the 12,901 probe set basis, which was obtained by subtracting 9376 aging related probe sets from the entire 22277 probe sets, where 9376 aging related expression changes were defined exactly as before. Finally, the sampled probe sets were screened with the same MLR criteria used for development of the 131-element classifier, i.e., (1) βs>0, (2) βs>10×βT, and (3) p (βs<0.1). In each random experiment, the genes that survived the MLR filter were used to develop a classifier with PAM exactly as for the 131-probe set classifier. PAM selected an average of 6.2 probe sets (<<131), and the average performance of these random-gene classifiers based on the tests of other datasets are summarized in Table 5. These random-gene classifiers failed to detect the presence of tumor in most of the test sets. The random classifier was particularly poor, however, in defining a normal distribution for Dataset 1, leading an 8.7% (Table 5, line 2) sensitivity suggesting a bias toward “no presence of tumor.” This correlated with the second lack of normal distribution due to a similar bias toward “no presence of tumor,” but this time affecting the normal tissues and thereby giving rise to the appearance of accuracy with an average of 82.3% (Table 5, average lines 6-9 and 13). In general, however, the random model tended to be a normal distribution with poor accuracies in the range of 12.9% to 19.2%, indicating that the results obtained with the developed 131-probe set classifier cannot be attributed to chance.
laevis)
cerevisiae)
laevis)
elegans)
1logFC is the logarithm Fold Change as tumorous stroma being compared to normal stroma.
Three methods utilized in the development of predictive gene signature of prostate cancer are described in this example. First, an analytical method based on a linear combination model for the determination of the percent cell composition of the tumor epithelial cells and the stoma cells from array data of mixed cell type prostate tissue is described. The method utilizes fixed expression coefficients of a small (<100) genes that with expression characteristics that are distinct for tumor epithelial and stroma cells.
Second, a new method for the determination of tumor cell specific biomarkers for the prediction of relapse of prostate cancer using an extended linear combination model is described and validated. A gene profile based on the expression of RNA of prostate cancer epithelial cells that predicts the differential gene expression of relapse (aggressive) vs. non relapse (indolent) prostate cancer is derived. These genes are validated by their identification in independent sets of prostate cancer patients (technical retrospective validation) is described. This method may be used to identify aggressive prostate cancer from data obtained at the time of diagnosis. The method and profiles are novel.
Third, an analogous new method for the determination of stroma cell specific biomarkers for the prediction of relapse of prostate cancer is described. Thus the predictions are based on non tumor cell types. A gene profile based on the expression of RNA of stroma cells of tumor-bearing prostate tissue that predicts the differential gene expression of relapse (aggressive) vs. non relapse (indolent) prostate cancer that is validated by prediction of differences of an independent set of prostate cancer patients (technical retrospective validation) is described. These methods and profiles may be used to identify aggressive prostate cancer from data obtained at the time of diagnosis. The results further indicate that the microenvironment of tumor foci of prostate cancer exhibit altered gene expression at the time of diagnosis which is distinct in non relapse and relapsed prostate cancer.
Datasets:
The goals of this study were to continue development of predicative biomarkers of prostate cancer. In particular the goal of this study is to use independent datasets to validate genes deduced as predictive based on studies of dataset 1 (infra vide). Here “dataset” refers to the array-based RNA expression data of all cases of a given set together with the clinical data defining whether a given case relapsed (recurred cancer) or remained disease free, a censored quantity. Only the categorical value, relapsed or non relapsed, is used in the analyses described here.
The three datasets used for this study included 1) 148 Affymetrix U133A array data acquired from 91 patients (publicly available in the GEO database as accession no. GSE8218) which is the principal dataset utilized in previous studies; 2) Illumina (of Illumina Inc., San Diego) beads arrays data from 103 patients as analyzed on 115 arrays, a published dataset (Bibilova et al. (2007) Genomics 89:666-672); and 3) Affymetrix U133A array data from 79 patients, also a published dataset (Stephenson et al., supra). These are referred to in this example as datasets 1, 2, and 3 respectively.
For the purposes herein, relapsed prostate cancer is taken as a surrogate of aggressive disease, while non-relapse is taken as indolent disease with a variable degree of indolence that is directly proportional to the disease-free survival time. Dataset 1 contains 40 non-relapse patients and 47 relapse patients; dataset 2 contains 75 non-relapse patients and 22 relapse patients, and dataset 3 contains 42 non-relapse patients and 37 relapse patients. The first two datasets samples have various amount of different tissue and cell types, including tumor cells, stroma cells (a collective term for fibroblasts, myofibroblasts, smooth muscle, and small amounts of nerve and vascular elements), BPH (epithelial cells of benign prostate hypertrophy) and dilated cystic glands (AKA “atrophic” cystic glands), as estimated by four pathologists (Stuart et al., supra) for dataset 1 and one pathologist for dataset 2. Dataset 3 samples were tumor-enriched samples. In this study, published datasets 2 and 3 were used for the purpose of validation only. A major goal of this study was to use “external” published datasets to validate the properties deduced for genes based on analysis of the dataset 1.
Determination of Cell Specific Gene Expression in Prostate Cancer:
Using linear models applied to microarray data from prostate tissues with various amounts of different cell types as estimated by a team of four pathologists, identified genes were identified as being specifically expressed in different cell types (tumor, stroma, BPH and dilated cystic glands) of prostate tissue following published methods (Stuart et al., supra). Thus, the following linear models were applied for generating tissue specific genes.
Model 1
For any gene i, the hybridization intensity, G, from an Affymetrix GeneChip is due to the sum of the cell contributions to the total mRNA:
G
i=(βtumorPtumor+βstroma·Pstroma+βBPH·PBPH+βBPH dilated cystic·Pgland dilated cystic gland)i
Where a “cell contribution” is the amount of the cellular component, Pcell type, multiplied times the characteristic expression level of gene i by that cell type, β. Only the β values are unknown and are determined by simple or multiple linear regressions. Note that in general a minimum of four estimates of Gi (i.e. four cases) are required to estimate four unknown β whereas in practice many dozens of cases are available so that the unknown coefficients are “over determined”.
Model 2
Since the epithelia of dilated cystic glands were not a major component of prostate tissue, it may be removed from the linear model to simplify the model.
G
i=(βtumor·Ptumor+βstroma·Pstroma+βBPH·PBPH)i
Models 3˜6
To further simplify the model, cell composition also can be considered as two different cell types, usually one specific cell type and all the other cell types were grouped together.
G
i=(βtumor·Ptumor+βnon-tumor·Pnon-tumor)i
G
i=(βstroma·Pstroma+βnon-stroma·Pnon-stroma)i
G
i=(βBPH·PBPH+βnon-BPH·Pnon-BPH)i
G
i=(βdilated cystic gland·Pdilated cystic gland+βnon-dilated cystic gland·Pnon-dilated cystic gland)i
The gene lists (with p<0.001) developed from models 3 and 4 using dataset 1 are listed in Table 6.
A New Method for Determination of Cell Type Composition Prediction Using Gene Expression Profiles:
Using linear models based on a small list of cell specific genes, i.e., genes from Table 6, the approximate percentage of cell types in samples hybridized to the array may be estimated using only the microarray data utilizing model 3. Potentially all of the genes in Table 6 can be used for cell percent composition prediction. For each individual gene, a new sample's gene expression value from microarray data can be fitted to models 3˜6, for a prediction of corresponding cell type percentage. Each gene employed in model 3 provides an estimate of percent tumor cell composition. The median of the predictions based on multiple genes was used to generate a more reliable result estimate of tumor cell content. These prediction genes can be selected/ranked by either their correlation coefficient (for correlation between gene expression level and cell type percentage) or by combination of genes with the best prediction power. In the present case, only a very limited number of genes (8-52 genes) were used for such a prediction. Even fewer genes might be sufficient.
To validate the method of tumor or stroma percent composition determination, the known percent composition figures of dataset 1 were used to predict the tumor cell and stroma cell compositions for dataset 2 with known cell composition. For example, the number of genes used for cell type (tumor epithelial cells or stroma cells) prediction between dataset 1 and dataset 2 ranges from 8 to 52 genes, which are listed in Table 7A. The Pearson correlation coefficient between predicted cell type percentage (tumor epithelial cells or stroma cells) and pathologist estimated percentage ranged from 0.7 to 0.87. Tissue (tumor or stroma) specific genes identified from dataset 2 and used for prediction are listed in Table 7B.
Since dataset 1 and dataset 2 data were based on different array platforms, the cross-platform normalization were applied using median rank scores (MRS) method (Warnat et al. (2005) BMC Bioinformatics 6:265).
A New Method for Determination of Cell Specific Relapse Related Genes of Prostate Cancer:
Using dataset 1, the genes correlating with patient relapse status were estimated using the following linear models.
Model 7
G
i=β′tumor,iPtumor+β′stroma,iPstroma+β′BPH,iPBPH+β′dilated cystic gland,iPdilated cystic gland+rs(γtumor,iPtumor+γstroma,iPstroma+γBPH,iPBPH+γdilated cystic gland,iPdilated cystic gland)
For any gene i, Gi (the array reported gene intensity)=the sum of 4 cell type contributions for non relapsed cases (βcell type,i×Percentcell type)+Sum of 4 cell type contributions for relapsed cases (γcell type,i×Percentcell type)+error term. RS may be either 0 or 1 where 0 is utilized for all non relapse cases and RS=0 is utilized for relapse cases. Thus when RS=0 the expression coefficients β′ for non relapse cases are determined while when RS=1 the coefficients (β′+γ) are determined. Coefficients are numerically determined by multiple linear regression using least squares determination of best fit coefficients±error. The differences in expression between non relapse (β′) and relapse (β′+γ) is just γ and the significance γ may be estimated by T-test and other standard statistical methods.
Model 8˜11
The following models also were implemented to simplify the models:
G
i=β′tumor,iPtumor+β′relapse status,iRS+β′interaction,iPtumor:RS
G
i=β′stroma,iPstroma+β′relapse status,iRS+β′interaction,iPstroma:RS
G
i=β′Btumor,iPtumor+β′relapse status,iRS+β′intreaction,iPtumor:RS
G
i=β′dilated cystic gland,iPtumor+β′relapse status,iRS+β′interaction,iPdilated cystic gland:RS
Only the samples with >0% tumor epithelial cells were used for the above analysis to remove those far-stroma samples (i.e., non-tumor cell bearing samples). This exclusion of “far-stroma” accommodates the possibility that stroma may contain expression changes characteristic of prostates with cancer, but that these changes might be confined to stroma regions near tumor cells. Because multiple samples are used from some subjects, the estimating equations approach implemented in the “gee” library for R (i.e., the open source R bioinformatics analysis package) was used (Zeger and Liang (1986) Biometrics 42:121-130). Cell type (tumor epithelial cells or stroma cells) specific genes showed significant (p<0.005) expression level changes between relapse and non-relapse samples using model 8-9, are listed in Tables 8A and 8B.
The gene list was then validated using independent dataset 3 to test whether any of the same genes were independently identified. Since dataset 3 has unknown tumor/stroma content, the method was first used for predicting tumor/stroma percentage (
An analogous analysis was carried for the determination of stroma cell specific genes (
An analogous analysis was carried out using datasets 1 and 2 with a significance cut off of 0.2 for dataset 2 (Table 9). Thirteen coincident genes were identified at this threshold even though the array of dataset three is relatively small (˜500 genes). Ten of these 13 genes had the same direction in relapse in both datasets (p<0.011), as shown in
A similar analysis for stroma-specifically expressed genes revealed BTG2 as a stroma specific relapse gene (Affymetrix ID: 201235_s_at) as a common gene in dataset 1 and 2 that exhibited up-regulation in both datasets.
These results indicate that three sets of validated genes with significant differential expression may be extracted once tumor percentage is taken into account, which may be useful in the prediction of relapse by analysis of expression data obtained at the time of diagnosis.
203160_s_at
212486_s_at
205111_s_at
209831_x_at
215311_at
52975_at
205447_s_at
212818_s_at
206637_at
204636_at
210140_at
204502_at
205543_at
219838_at
219801_at
210408_s_at
211871_x_at
219815_at
214078_at
204221_x_at
209827_s_at
217965_s_at
207375_s_at
213804_at
207436_x_at
212550_at
219821_s_at
209716_at
213533_at
219970_at
209603_at
53991_at
202744_at
203217_s_at
205192_at
207614_s_at
207457_s_at
221886_at
204437_s_at
203010_at
203187_at
217452_s_at
220452_x_at
214617_at
64942_at
202663_at
203734_at
211256_x_at
204879_at
213906_at
219390_at
220246_at
214033_at
204982_at
215506_s_at
218029_at
208213_s_at
204504_s_at
212823_s_at
221832_s_at
205112_at
219738_s_at
203598_s_at
219464_at
35846_at
209243_s_at
211843_x_at
206403_at
202530_at
200015_s_at
204552_at
206009_at
205121_at
206178_at
210692_s_at
203798_s_at
200066_at
203741_s_at
218805_at
211072_x_at
219213_at
221753_at
212639_x_at
213509_x_at
204513_s_at
211194_s_at
205255_x_at
212130_x_at
218266_s_at
216017_s_at
206050_s_at
203348_s_at
218997_at
212227_x_at
201515_s_at
209789_at
212926_at
217914_at
204642_at
40472_at
213030_s_at
37152_at
213066_at
217721_at
203045_at
209940_at
214118_x_at
210882_s_at
205760_s_at
220027_s_at
214285_at
204043_at
203167_at
217220_at
204038_s_at
211330_s_at
218677_at
52837_at
202410_x_at
221044_s_at
40560_at
221656_s_at
218950_at
211809_x_at
205240_at
214995_s_at
211780_x_at
211325_x_at
213932_x_at
219114_at
219529_at
203197_s_at
213922_at
210079_x_at
203456_at
212079_s_at
219616_at
37384_at
221779_at
221552_at
214853_s_at
207053_at
208325_s_at
212134_at
219195_at
221699_s_at
203069_at
220016_at
215439_x_at
206191_at
202092_s_at
210794_s_at
206087_x_at
219768_at
204627_s_at
52651_at
200886_s_at
221551_x_at
205159_at
218775_s_at
209688_s_at
36829_at
203592_s_at
210347_s_at
213644_at
211058_x_at
203047_at
209877_at
218807_at
213885_at
220937_s_at
205405_at
50965_at
207747_s_at
203757_s_at
209171_at
209320_at
207984_s_at
212280_x_at
202098_s_at
204047_s_at
209618_at
203530_s_at
204428_s_at
221052_at
204747_at
217312_s_at
215734_at
201934_at
202652_at
204234_s_at
209721_s_at
218802_at
208842_s_at
218310_at
212695_at
219148_at
217608_at
206033_s_at
205429_s_at
213269_at
204044_at
214806_at
31845_at
222217_s_at
203046_s_at
208103_s_at
202590_s_at
207654_x_at
213270_at
220142_at
221036_s_at
217993_s_at
213646_x_at
218766_s_at
217904_s_at
220399_at
204763_s_at
211801_x_at
207988_s_at
220144_s_at
219767_s_at
208393_s_at
211892_s_at
206688_s_at
213100_at
202059_s_at
213630_at
213679_at
219684_at
201977_s_at
211401_s_at
207018_s_at
212076_at
212479_s_at
211668_s_at
209910_at
204174_at
201420_s_at
207971_s_at
212790_x_at
204589_at
219238_at
213467_at
34221_at
203666_at
217910_x_at
205104_at
217598_at
202191_s_at
209145_s_at
221234_s_at
219154_at
205528_s_at
205243_at
205008_s_at
210410_s_at
204177_s_at
212436_at
215767_at
209745_at
201294_s_at
204883_s_at
208018_s_at
208903_at
209257_s_at
213685_at
210702_s_at
214210_at
61734_at
212719_at
210736_x_at
213608_s_at
201090_x_at
220661_s_at
212360_at
43977_at
209841_s_at
217930_s_at
209534_x_at
202945_at
204633_s_at
218868_at
212803_at
205909_at
216187_x_at
207396_s_at
205786_s_at
209672_s_at
209308_s_at
205850_s_at
209867_s_at
221550_at
204556_s_at
218558_s_at
220071_x_at
213393_at
206122_at
213237_at
218424_s_at
205432_at
201183_s_at
202791_s_at
40446_at
218953_s_at
219134_at
221818_at
221885_at
221738_at
204736_s_at
219538_at
212373_at
207059_at
210785_s_at
203208_s_at
214036_at
211720_x_at
219628_at
218874_s_at
212427_at
218159_at
205902_at
208009_s_at
214909_s_at
219635_at
203278_s_at
204809_at
219602_s_at
213115_at
202831_at
214481_at
40837_at
218146_at
53720_at
209195_s_at
212235_at
219723_x_at
213260_at
212395_s_at
215493_x_at
208648_at
215411_s_at
213063_at
214436_at
208569_at
221795_at
208955_at
209866_s_at
33307_at
200813_s_at
218562_s_at
211366_x_at
204402_at
219243_at
204476_s_at
212299_at
222018_at
203879_at
213223_at
218373_at
218598_at
203944_x_at
204798_at
220634_at
213601_at
219563_at
213009_s_at
203586_s_at
204903_x_at
212706_at
219209_at
200697_at
201033_x_at
202646_s_at
208856_x_at
205632_s_at
203947_at
206032_at
217740_x_at
212468_at
216652_s_at
204882_at
203790_s_at
204062_s_at
219033_at
209726_at
208923_at
205453_at
202632_at
203369_x_at
211378_x_at
202783_at
44065_at
220818_s_at
204003_s_at
208158_s_at
209188_x_at
211006_s_at
221018_s_at
202022_at
221508_at
205325_at
39966_at
204063_s_at
220773_s_at
211316_x_at
219129_s_at
207895_at
215215_s_at
212629_s_at
203040_s_at
214298_x_at
202063_s_at
202522_at
206919_at
219436_s_at
209440_at
219961_s_at
213708_s_at
206972_s_at
204169_at
218691_s_at
203287_at
202733_at
204423_at
208869_s_at
208778_s_at
203812_at
218199_s_at
212796_s_at
218988_at
213095_x_at
208696_at
210926_at
211765_x_at
215606_s_at
218797_s_at
205525_at
201709_s_at
202578_s_at
218249_at
221484_at
210192_at
214725_at
208822_s_at
203853_s_at
212127_at
211701_s_at
206587_at
202206_at
213083_at
39582_at
203800_s_at
209901_x_at
208968_s_at
204334_at
213189_at
221991_at
211658_at
203662_s_at
218511_s_at
202254_at
201771_at
208206_s_at
218316_at
213394_at
209777_s_at
38487_at
217961_at
211657_at
212121_at
212715_s_at
202031_s_at
221901_at
204008_at
219545_at
202331_at
219939_s_at
212342_at
208616_s_at
210005_at
202116_at
203500_at
209970_x_at
37831_at
214791_at
204853_at
200916_at
215482_s_at
204198_s_at
204618_s_at
203320_at
211972_x_at
203894_at
222362_at
219520_s_at
220966_x_at
201146_at
217256_x_at
212157_at
206109_at
222171_s_at
201489_at
210073_at
208985_s_at
214629_x_at
221156_x_at
213203_at
203677_s_at
201361_at
205928_at
221473_x_at
211212_s_at
203661_s_at
211113_s_at
202795_x_at
211978_x_at
203037_s_at
34764_at
207571_x_at
219080_s_at
219523_s_at
201723_s_at
202998_s_at
219742_at
209332_s_at
219562_at
203797_at
207262_at
203919_at
204353_s_at
203508_at
203573_s_at
220677_s_at
212155_at
203074_at
219075_at
205231_s_at
219066_at
200673_at
213941_x_at
48031_r_at
204050_s_at
203599_s_at
209925_at
201380_at
218911_at
218032_at
202713_s_at
214177_s_at
202306_at
215990_s_at
209429_x_at
209402_s_at
200651_at
213590_at
218392_x_at
202000_at
218289_s_at
219597_s_at
204488_at
219014_at
218725_at
37022_at
214864_s_at
220108_at
213435_at
222073_at
201758_at
210401_at
218688_at
214052_x_at
216945_x_at
202613_at
201293_x_at
203249_at
221791_s_at
32094_at
208596_s_at
205398_s_at
219097_x_at
205611_at
207168_s_at
213271_s_at
208369_s_at
211031_s_at
203816_at
221928_at
218160_at
204421_s_at
212661_x_at
213556_at
200739_s_at
213217_at
203330_s_at
222221_x_at
209284_s_at
202328_s_at
40359_at
204683_at
212015_x_at
213478_at
202272_s_at
211368_s_at
200734_s_at
207071_s_at
220318_at
204912_at
215947_s_at
205823_at
200068_s_at
205479_s_at
202105_at
213113_s_at
200022_at
46665_at
208466_at
202965_s_at
218512_at
44702_at
201113_at
212409_s_at
218540_at
202449_s_at
210761_s_at
211726_s_at
218070_s_at
208786_s_at
216380_x_at
210089_s_at
208687_x_at
32259_at
219223_at
218487_at
205339_at
208112_x_at
208941_s_at
209703_x_at
218817_at
204462_s_at
203713_s_at
208964_s_at
205371_s_at
210224_at
58696_at
213326_at
219321_at
203185_at
204247_s_at
204606_at
222206_s_at
216594_x_at
205634_x_at
215059_at
202487_s_at
200788_s_at
218741_at
AFFX-
218669_at
201209_at
HSAC07/X00351_3_at
201913_s_at
218634_at
202282_at
216100_s_at
221196_x_at
214604_at
219463_at
209198_s_at
208072_s_at
218820_at
217968_at
220092_s_at
218653_at
221905_at
213699_s_at
218935_at
209391_at
202579_x_at
221807_s_at
204150_at
201239_s_at
203063_at
208759_at
209015_s_at
209421_at
215051_x_at
200657_at
212855_at
213427_at
211675_s_at
217944_at
213531_s_at
216895_at
208491_s_at
218069_at
213295_at
200809_x_at
201474_s_at
207871_s_at
209474_s_at
204378_at
200801_x_at
222234_s_at
205116_at
219255_x_at
217802_s_at
209238_at
213513_x_at
203437_at
213567_at
212861_at
219496_at
214271_x_at
202897_at
218123_at
208859_s_at
220603_s_at
204546_at
222025_s_at
201718_s_at
219203_at
212326_at
219289_at
220974_x_at
201512_s_at
212262_at
217976_s_at
207691_x_at
201672_s_at
209606_at
209262_s_at
204537_s_at
204360_s_at
213867_x_at
213912_at
213925_at
217791_s_at
203650_at
212351_at
205259_at
205441_at
208454_s_at
218101_s_at
218815_s_at
218436_at
204341_at
215023_s_at
211819_s_at
202811_at
203811_s_at
206556_at
36030_at
218636_s_at
200713_s_at
211098_x_at
212177_at
209804_at
218472_s_at
207156_at
201375_s_at
202900_s_at
214808_at
221696_s_at
212371_at
206004_at
222008_at
202322_s_at
204134_at
204295_at
215313_x_at
206492_at
211000_s_at
201629_s_at
201537_s_at
202488_s_at
215346_at
202514_at
205088_at
212433_x_at
203482_at
208659_at
219431_at
91684_g_at
200984_s_at
219676_at
201980_s_at
211036_x_at
204136_at
206831_s_at
209602_s_at
210768_x_at
205315_s_at
201077_s_at
221485_at
214442_s_at
218731_s_at
209617_s_at
204436_at
218834_s_at
221503_s_at
205761_s_at
211769_x_at
221826_at
209598_at
211558_s_at
209960_at
215300_s_at
203499_at
219786_at
219764_at
204478_s_at
210875_s_at
206533_at
218012_at
202433_at
218425_at
201614_s_at
210840_s_at
201886_at
218128_at
201385_at
216210_x_at
204034_at
212082_s_at
207833_s_at
209039_x_at
210594_x_at
218651_s_at
205617_at
206243_at
207827_x_at
202910_s_at
218209_s_at
213766_x_at
208107_s_at
200676_s_at
36475_at
201403_s_at
203252_at
209840_s_at
212740_at
217109_at
210023_s_at
210880_s_at
218252_at
202561_at
206066_s_at
202136_at
203738_at
213034_at
203569_s_at
202048_s_at
217958_at
33850_at
213188_s_at
212504_at
200740_s_at
213817_at
208821_at
43427_at
214831_at
212188_at
201613_s_at
209765_at
213610_s_at
207317_s_at
201588_at
214297_at
219307_at
60471_at
219709_x_at
217066_s_at
200691_s_at
202510_s_at
203926_x_at
200758_s_at
209317_at
202439_s_at
219428_s_at
201785_at
206722_s_at
222199_s_at
220607_x_at
212798_s_at
209433_s_at
213658_at
200875_s_at
221875_x_at
220934_s_at
205795_at
220174_at
209570_s_at
201095_at
209719_x_at
220647_s_at
200900_s_at
205512_s_at
208617_s_at
202190_at
213940_s_at
219860_at
213434_at
218180_s_at
221805_at
219575_s_at
205006_s_at
203682_s_at
212758_s_at
203458_at
221447_s_at
218509_at
220911_s_at
204088_at
209203_s_at
218133_s_at
204222_s_at
218780_at
212408_at
202852_s_at
218844_at
204675_at
203535_at
217249_x_at
207302_at
210927_x_at
204308_s_at
219771_at
209539_at
202705_at
202856_s_at
214011_s_at
219058_x_at
218198_at
220230_s_at
200088_x_at
205139_s_at
203925_at
210829_s_at
201175_at
204365_s_at
211061_s_at
220115_s_at
218481_at
202803_s_at
200925_at
213939_s_at
203154_s_at
212658_at
221206_at
211776_s_at
209323_at
210561_s_at
207563_s_at
206868_at
201478_s_at
202362_at
205140_at
205005_s_at
219324_at
205551_at
208805_at
204045_at
201682_at
218062_x_at
207831_x_at
203409_at
208405_s_at
218127_at
219188_s_at
212196_at
202604_x_at
205267_at
200750_s_at
201885_s_at
206527_at
220955_x_at
214789_x_at
210976_s_at
203621_at
202861_at
220334_at
204542_at
217835_x_at
209009_at
219874_at
243_g_at
217861_s_at
220272_at
204862_s_at
214812_s_at
222001_x_at
219451_at
203312_x_at
209435_s_at
217720_at
203909_at
221797_at
219514_at
203014_x_at
211653_x_at
206782_s_at
212792_at
218008_at
207714_s_at
204212_at
217211_at
212426_s_at
204989_s_at
204228_at
218345_at
217797_at
219670_at
221253_s_at
207069_s_at
211202_s_at
202594_at
208756_at
204215_at
204025_s_at
1294_at
202671_s_at
203567_s_at
219302_s_at
212822_at
212902_at
209083_at
217929_s_at
212169_at
218005_at
203787_at
219851_at
38671_at
207439_s_at
207838_x_at
221817_at
201021_s_at
220865_s_at
203340_s_at
201338_x_at
218332_at
202697_at
212567_s_at
204811_s_at
212294_at
210409_at
206854_s_at
209434_s_at
201828_x_at
212508_at
201506_at
201256_at
205738_s_at
204244_s_at
211203_s_at
213913_s_at
204249_s_at
221654_s_at
209297_at
218756_s_at
207705_s_at
217772_s_at
209699_x_at
212416_at
202656_s_at
203152_at
213603_s_at
210532_s_at
215222_x_at
219809_at
1405_i_at
207147_at
209702_at
212597_s_at
208096_s_at
202329_at
203726_s_at
218270_at
213395_at
212006_at
204151_x_at
202120_x_at
202617_s_at
216295_s_at
201649_at
201371_s_at
205076_s_at
214156_at
221527_s_at
212622_at
215867_x_at
218788_s_at
203503_s_at
210386_s_at
218660_at
209399_at
214937_x_at
209817_at
204834_at
220587_s_at
212565_at
218684_at
201336_at
217785_s_at
213698_at
213307_at
209563_x_at
218529_at
209194_at
201909_at
201287_s_at
202788_at
203151_at
213947_s_at
209732_at
205190_at
207397_s_at
218264_at
213261_at
219293_s_at
212441_at
200997_at
201795_at
212637_s_at
202657_s_at
221689_s_at
206382_s_at
221868_at
202378_s_at
209104_s_at
207233_s_at
204167_at
201155_s_at
214983_at
214369_s_at
206993_at
221730_at
218320_s_at
219305_x_at
212995_x_at
219025_at
213607_x_at
213151_s_at
220525_s_at
209454_s_at
220495_s_at
205082_s_at
218398_at
202158_s_at
214006_s_at
207453_s_at
210250_x_at
211997_x_at
204161_s_at
206071_s_at
221597_s_at
213386_at
220235_s_at
201022_s_at
217812_at
202784_s_at
202658_at
205079_s_at
218689_at
204682_at
203744_at
205153_s_at
220285_at
202273_at
218361_at
203883_s_at
219517_at
211473_s_at
205774_at
209834_at
203987_at
212063_at
205770_at
201108_s_at
217932_at
211458_s_at
208906_at
212660_at
218764_at
217820_s_at
210058_at
204048_s_at
217809_at
209569_x_at
218882_s_at
204482_at
212129_at
202820_at
33814_at
202478_at
204263_s_at
202756_s_at
202802_at
214656_x_at
218795_at
204438_at
200620_at
219416_at
201349_at
218631_at
203647_s_at
218084_x_at
219733_s_at
203698_s_at
213292_s_at
206600_s_at
211787_s_at
207124_s_at
220104_at
218648_at
202813_at
220326_s_at
209100_at
203794_at
35671_at
219229_at
209407_s_at
212223_at
222231_s_at
202501_at
213897_s_at
203332_s_at
218358_at
212420_at
219053_s_at
208030_s_at
200693_at
202577_s_at
202144_s_at
209365_s_at
201530_x_at
213455_at
219211_at
205559_s_at
207165_at
214577_at
218772_x_at
202957_at
221539_at
200655_s_at
202799_at
212457_at
201458_s_at
218368_s_at
201456_s_at
202552_s_at
202347_s_at
49452_at
217827_s_at
203828_s_at
214751_at
218641_at
217898_at
214624_at
202645_s_at
213138_at
204067_at
212702_s_at
212415_at
204948_s_at
201576_s_at
200791_s_at
210854_x_at
211700_s_at
201415_at
202723_s_at
214173_x_at
202508_s_at
209014_at
203756_at
201317_s_at
202003_s_at
212544_at
214211_at
221475_s_at
205100_at
221665_s_at
203104_at
201406_at
212080_at
203942_s_at
221565_s_at
204435_at
212367_at
212519_at
203281_s_at
218341_at
214460_at
204624_at
211518_s_at
208613_s_at
208763_s_at
218282_at
216944_s_at
218440_at
212259_s_at
217746_s_at
205870_at
222212_s_at
208070_s_at
202168_at
218309_at
218427_at
220975_s_at
50374_at
202371_at
203351_s_at
219561_at
206949_s_at
218831_s_at
201023_at
204670_x_at
218202_x_at
209321_s_at
220354_at
35776_at
217748_at
200920_s_at
218866_s_at
212917_x_at
205661_s_at
208671_at
217726_at
200694_s_at
219060_at
202259_s_at
218219_s_at
209582_s_at
218111_s_at
216840_s_at
218695_at
219525_at
200037_s_at
210605_s_at
201587_s_at
205648_at
213498_at
212263_at
202025_x_at
204979_s_at
202670_at
204797_s_at
221462_x_at
205207_at
200082_s_at
205529_s_at
212825_at
204011_at
219492_at
215096_s_at
201501_s_at
209081_s_at
217716_s_at
200884_at
201003_x_at
220952_s_at
212461_at
216894_x_at
207722_s_at
209437_s_at
207121_s_at
212117_at
202767_at
204854_at
202959_at
209485_s_at
202320_at
204000_at
206723_s_at
213737_x_at
205161_s_at
212851_at
201341_at
202616_s_at
218163_at
206458_s_at
217200_x_at
210762_s_at
209130_at
206375_s_at
208757_at
214823_at
202738_s_at
210201_x_at
219215_s_at
214736_s_at
209479_at
202446_s_at
204266_s_at
209075_s_at
203270_at
209506_s_at
36936_at
209307_at
209233_at
213058_at
210523_at
202575_at
218037_at
204820_s_at
219521_at
200702_s_at
201074_at
210102_at
207668_x_at
200609_s_at
208270_s_at
212494_at
204066_s_at
208679_s_at
210357_s_at
205824_at
204290_s_at
201040_at
202787_s_at
218183_at
218491_s_at
218627_at
220768_s_at
202734_at
208674_x_at
208712_at
39729_at
218284_at
209509_s_at
215000_s_at
202614_at
202047_s_at
212739_s_at
213422_s_at
200715_x_at
210973_s_at
203213_at
209069_s_at
204264_at
216033_s_at
205329_s_at
202291_s_at
216640_s_at
219165_at
218110_at
201121_s_at
205317_s_at
219489_s_at
219732_at
206813_at
203576_at
212221_x_at
209110_s_at
209546_s_at
215812_s_at
212503_s_at
201586_s_at
202117_at
209142_s_at
219370_at
204985_s_at
203501_at
221003_s_at
212111_at
212953_x_at
212518_at
201675_at
218454_at
212316_at
211944_at
209971_x_at
212158_at
217970_s_at
210968_s_at
211758_x_at
212586_at
215519_x_at
210628_x_at
205246_at
202643_s_at
206254_at
205044_at
212032_s_at
208306_x_at
200098_s_at
212119_at
218567_x_at
201730_s_at
213490_s_at
202450_s_at
209180_at
222240_s_at
217959_s_at
212179_at
202886_s_at
214660_at
210434_x_at
208335_s_at
213687_s_at
204790_at
204340_at
202464_s_at
205084_at
201311_s_at
208799_at
207118_s_at
205687_at
209967_s_at
203316_s_at
57715_at
218493_at
222024_s_at
220742_s_at
209263_x_at
215091_s_at
203749_s_at
201780_s_at
203071_at
217846_at
209596_at
204343_at
218667_at
218563_at
201721_s_at
201931_at
205805_s_at
205145_s_at
33322_i_at
214167_s_at
201605_x_at
218548_x_at
204794_at
201016_at
209343_at
208852_s_at
211796_s_at
201479_at
203518_at
203317_at
201696_at
200055_at
203597_s_at
208864_s_at
202172_at
201826_s_at
218892_at
214117_s_at
213249_at
211033_s_at
207542_s_at
202923_s_at
204260_at
208800_at
204310_s_at
208436_s_at
213170_at
209739_s_at
202765_s_at
200831_s_at
204344_s_at
203272_s_at
204491_at
217127_at
202208_s_at
200087_s_at
200611_s_at
210312_s_at
204294_at
222356_at
203156_at
65133_i_at
212120_at
212527_at
205201_at
218503_at
210632_s_at
207181_s_at
203339_at
218321_x_at
205478_at
203246_s_at
210915_x_at
202300_at
217795_s_at
200942_s_at
218723_s_at
204391_x_at
218902_at
213245_at
212878_s_at
203133_at
209312_x_at
212219_at
214085_x_at
213720_s_at
215306_at
201066_at
200905_x_at
205244_s_at
221898_at
205355_at
212197_x_at
212340_at
213519_s_at
218732_at
214894_x_at
221511_x_at
202908_at
208959_s_at
215543_s_at
212165_at
202305_s_at
218448_at
208634_s_at
218357_s_at
204803_s_at
218816_at
205857_at
202710_at
212353_at
220925_at
203889_at
201630_s_at
218152_at
202138_x_at
55081_at
213843_x_at
214771_x_at
221620_s_at
214608_s_at
211708_s_at
208760_at
216958_s_at
202931_x_at
217284_x_at
208502_s_at
219041_s_at
204730_at
211177_s_at
201743_at
217824_at
219304_s_at
203581_at
201120_s_at
201011_at
219024_at
201463_s_at
200985_s_at
201830_s_at
203028_s_at
209545_s_at
200816_s_at
219819_s_at
213316_at
218857_s_at
219985_at
219913_s_at
212549_at
205980_s_at
33323_r_at
204466_s_at
218196_at
206724_at
213348_at
207721_x_at
207966_s_at
208801_at
209645_s_at
210186_s_at
217226_s_at
218010_x_at
217997_at
201772_at
208633_s_at
218016_s_at
212561_at
221588_x_at
202878_s_at
215280_s_at
211998_at
209776_s_at
210202_s_at
39817_s_at
219534_x_at
201653_at
203233_at
202119_s_at
201648_at
213379_at
208615_s_at
212751_at
213309_at
212246_at
205782_at
200873_s_at
202821_s_at
218112_at
201752_s_at
202737_s_at
203264_s_at
214240_at
208835_s_at
203827_at
212071_s_at
202666_s_at
206710_s_at
205750_at
213182_x_at
212563_at
203639_s_at
205294_at
211990_at
218969_at
202422_s_at
201268_at
211974_x_at
202299_s_at
203068_at
212053_at
219221_at
201819_at
205898_at
208264_s_at
203964_at
214542_x_at
205577_at
219125_s_at
215706_x_at
203605_at
218376_s_at
202502_at
205348_s_at
213116_at
208146_s_at
210859_x_at
221816_s_at
203918_at
205882_x_at
221786_at
222158_s_at
202195_s_at
58916_at
205613_at
218823_s_at
217870_s_at
208848_at
204333_s_at
202156_s_at
208702_x_at
202180_s_at
219342_at
218804_at
212406_s_at
212604_at
200961_at
212923_s_at
209998_at
201859_at
201597_at
213901_x_at
205709_s_at
213075_at
214140_at
218656_s_at
213836_s_at
203017_s_at
201619_at
205961_s_at
209864_at
209374_s_at
203544_s_at
204993_at
201947_s_at
205933_at
203177_x_at
213620_s_at
203360_s_at
212510_at
201523_x_at
209379_s_at
218046_s_at
209086_x_at
213132_s_at
215146_s_at
201733_at
201869_s_at
206307_s_at
219228_at
220945_x_at
209786_at
203024_s_at
212253_x_at
208764_s_at
202432_at
219283_at
221676_s_at
208843_s_at
202341_s_at
213166_x_at
212681_at
208639_x_at
201958_s_at
200910_at
201137_s_at
218174_s_at
215333_x_at
208638_at
202242_at
201549_x_at
204655_at
209921_at
201037_at
208654_s_at
214721_x_at
201410_at
205011_at
220721_at
211991_s_at
204426_at
203695_s_at
205486_at
209298_s_at
208826_x_at
212350_at
201216_at
209787_s_at
210627_s_at
201559_s_at
213059_at
221884_at
202983_at
201995_at
214779_s_at
203685_at
209175_at
219936_s_at
213017_at
202008_s_at
212767_at
215193_x_at
203997_at
201968_s_at
218375_at
204759_at
219787_s_at
212430_at
203880_at
209846_s_at
210136_at
221870_at
211971_s_at
204640_s_at
205807_s_at
214121_x_at
213152_s_at
203178_at
203415_at
213547_at
201622_at
221666_s_at
201096_s_at
203813_s_at
203379_at
209568_s_at
214472_at
218675_at
218681_s_at
203604_at
209872_s_at
211986_at
201359_at
201566_x_at
201972_at
203619_s_at
218647_s_at
211026_s_at
218001_at
204028_s_at
204123_at
205624_at
218944_at
209691_s_at
208951_at
213135_at
212311_at
204140_at
209036_s_at
204735_at
201486_at
206453_s_at
200967_at
202132_at
209593_s_at
209612_s_at
205938_at
213015_at
214895_s_at
209197_at
212109_at
204049_s_at
215125_s_at
213306_at
208886_at
AFFX-
202207_at
221531_at
HSAC07/X00351_M_at
205622_at
213714_at
200699_at
219737_s_at
221041_s_at
208767_s_at
220584_at
37408_at
220342_x_at
202401_s_at
215923_s_at
213154_s_at
213491_x_at
201604_s_at
201659_s_at
213364_s_at
217551_at
218486_at
208074_s_at
206355_at
206103_at
212414_s_at
213119_at
201858_s_at
205875_s_at
221016_s_at
217868_s_at
203590_at
212175_s_at
201153_s_at
202233_s_at
205262_at
203148_s_at
220233_at
210087_s_at
202947_s_at
203123_s_at
202946_s_at
219036_at
212328_at
209576_at
209082_s_at
218633_x_at
204021_s_at
218073_s_at
215870_s_at
202558_s_at
200839_s_at
214096_s_at
203868_s_at
208716_s_at
203939_at
201524_x_at
222146_s_at
202712_s_at
216235_s_at
208918_s_at
203325_s_at
214214_s_at
214055_x_at
203207_s_at
205022_s_at
201091_s_at
212143_s_at
218928_s_at
221502_at
213996_at
208723_at
221827_at
202950_at
221984_s_at
204863_s_at
218272_at
202644_s_at
214855_s_at
205120_s_at
53968_at
202411_at
203582_s_at
218204_s_at
220761_s_at
205168_at
214710_s_at
213290_at
209227_at
213228_at
200804_at
212382_at
201358_s_at
201655_s_at
209007_s_at
221246_x_at
213857_s_at
207741_x_at
219061_s_at
202724_s_at
209482_at
222101_s_at
218283_at
221718_s_at
204949_at
204802_at
216338_s_at
201719_s_at
219200_at
214439_x_at
200846_s_at
212268_at
205698_s_at
218683_at
210739_x_at
209473_at
201722_s_at
209584_x_at
210296_s_at
201744_s_at
208722_s_at
205127_at
202308_at
203140_at
204039_at
210896_s_at
202425_x_at
213656_s_at
203235_at
209737_at
212688_at
203232_s_at
217927_at
211538_s_at
203721_s_at
200653_s_at
204427_s_at
219902_at
219603_s_at
204304_s_at
218039_at
209199_s_at
201115_at
203687_at
201698_s_at
205109_s_at
203139_at
212566_at
208796_s_at
200838_at
206827_s_at
201666_at
202832_at
91703_at
222155_s_at
212086_x_at
218680_x_at
212387_at
214857_at
218864_at
201736_s_at
203231_s_at
221542_s_at
205265_s_at
205293_x_at
203510_at
208787_at
204497_at
217908_s_at
222288_at
220638_s_at
213262_at
202838_at
201152_s_at
205073_at
209318_x_at
218984_at
216215_s_at
205107_s_at
201310_s_at
216064_s_at
205752_s_at
65517_at
218574_s_at
206790_s_at
221796_at
209608_s_at
215707_s_at
210946_at
212488_at
211034_s_at
201621_at
201961_s_at
205548_s_at
213129_s_at
212757_s_at
215438_x_at
212099_at
217900_at
204550_x_at
210962_s_at
205578_at
218268_at
207191_s_at
218792_s_at
201009_s_at
205019_s_at
203725_at
201520_s_at
201234_at
219762_s_at
213891_s_at
202996_at
206481_s_at
213995_at
210198_s_at
218192_at
218051_s_at
202606_s_at
33760_at
218241_at
218711_s_at
202793_at
204929_s_at
204922_at
205620_at
200889_s_at
212148_at
203484_at
202074_s_at
202603_at
220751_s_at
202346_at
212276_at
216074_x_at
201149_s_at
209300_s_at
210036_s_at
219335_at
205792_at
218972_at
204271_s_at
202543_s_at
222303_at
201264_at
213069_at
204301_at
209406_at
200968_s_at
209121_x_at
213050_at
213401_s_at
211416_x_at
209613_s_at
220189_s_at
202587_s_at
212322_at
204518_s_at
221648_s_at
203884_s_at
209064_x_at
207002_s_at
201078_at
210276_s_at
204392_at
213381_at
218291_at
209242_at
212305_s_at
211002_s_at
211936_at
221671_x_at
217964_at
201482_at
202064_s_at
209270_at
204927_at
209959_at
203201_at
212489_at
202918_s_at
201868_s_at
205876_at
210751_s_at
209218_at
45297_at
200820_at
202898_at
210816_s_at
204517_at
211404_s_at
201508_at
209150_s_at
210105_s_at
218500_at
201425_at
209662_at
202762_at
201098_at
204058_at
218439_s_at
216331_at
221941_at
203002_at
203971_at
213982_s_at
212496_s_at
219506_at
212536_at
209447_at
202418_at
202609_at
213234_at
212690_at
208653_s_at
218236_s_at
201892_s_at
201368_at
205593_s_at
203753_at
218275_at
212817_at
220094_s_at
205251_at
218981_at
214767_s_at
204175_at
201865_x_at
214005_at
213134_x_at
220741_s_at
204149_s_at
203102_s_at
202796_at
203225_s_at
203256_at
208802_at
212386_at
219848_s_at
205381_at
210886_x_at
216887_s_at
203008_x_at
215382_x_at
218206_x_at
203411_s_at
217790_s_at
205743_at
218888_s_at
201151_s_at
202096_s_at
201286_at
213301_x_at
209090_s_at
201568_at
221773_at
210024_s_at
209305_s_at
201005_at
208963_x_at
200806_s_at
212793_at
205812_s_at
206117_at
214522_x_at
210145_at
209873_s_at
216264_s_at
200929_at
216565_x_at
209265_s_at
201312_s_at
213308_at
221651_x_at
213410_at
203607_at
201953_at
204205_at
221882_s_at
215127_s_at
200803_s_at
203886_s_at
219048_at
221900_at
202655_at
37005_at
218826_at
201599_at
218326_s_at
205383_s_at
201790_s_at
201536_at
205164_at
201148_s_at
218704_at
207761_s_at
206557_at
201387_s_at
218701_at
1598_g_at
205594_at
206104_at
219217_at
212239_at
208840_s_at
204422_s_at
216305_s_at
221045_s_at
202194_at
210613_s_at
204386_s_at
209264_s_at
214307_at
201012_at
203775_at
212646_at
214281_s_at
212463_at
202395_at
212669_at
204608_at
219829_at
200048_s_at
218678_at
208910_s_at
205364_at
203165_s_at
218934_s_at
200599_s_at
221766_s_at
218532_s_at
202917_s_at
204127_at
203585_at
220942_x_at
215388_s_at
202211_at
202720_at
210243_s_at
202228_s_at
210241_s_at
203066_at
210907_s_at
202465_at
202660_at
208430_s_at
219065_s_at
204115_at
212623_at
204059_s_at
221586_s_at
214464_at
212410_at
AFFX-
212805_at
205077_s_at
HSAC07/X00351_5_at
211747_s_at
218421_at
205538_at
215464_s_at
211754_s_at
202157_s_at
201219_at
208965_s_at
201339_s_at
202388_at
218883_s_at
201185_at
214875_x_at
201008_s_at
205160_at
212195_at
218213_s_at
210471_s_at
206299_at
201272_at
213365_at
213993_at
201401_s_at
213158_at
204967_at
209135_at
218328_at
218502_s_at
202406_s_at
210072_at
217871_s_at
209287_s_at
221688_s_at
201867_s_at
204332_s_at
210517_s_at
201943_s_at
204037_at
213600_at
206359_at
211497_x_at
58780_s_at
204331_s_at
221276_s_at
212741_at
212240_s_at
218003_s_at
206022_at
209250_at
212358_at
203431_s_at
219647_at
213399_x_at
212845_at
217986_s_at
201289_at
218989_x_at
211962_s_at
209759_s_at
212535_at
202296_s_at
203810_at
204160_s_at
204114_at
212307_s_at
204455_at
202960_s_at
211984_at
212116_at
219427_at
204142_at
204755_x_at
200636_s_at
212203_x_at
213518_at
219505_at
201284_s_at
201329_s_at
206429_at
209604_s_at
219920_s_at
209200_at
212685_s_at
209883_at
64486_at
212354_at
218676_s_at
213004_at
208872_s_at
202766_s_at
208612_at
204621_s_at
215227_x_at
212077_at
211574_s_at
209505_at
214358_at
201389_at
218608_at
203636_at
201135_at
203688_at
212064_x_at
213110_s_at
219076_s_at
218435_at
201955_at
221583_s_at
220625_s_at
214724_at
204233_s_at
217023_x_at
221920_s_at
206932_at
206351_s_at
201602_s_at
208689_s_at
214077_x_at
200052_s_at
202086_at
200863_s_at
201315_x_at
212749_s_at
204688_at
202857_at
57588_at
209326_at
212151_at
217645_at
213274_s_at
202279_at
212554_at
205937_at
200808_s_at
218145_at
202759_s_at
212279_at
201109_s_at
200895_s_at
202794_at
221637_s_at
207547_s_at
201004_at
211564_s_at
209796_s_at
202728_s_at
218049_s_at
203570_at
201962_s_at
213016_at
201941_at
201850_at
202785_at
204072_s_at
211899_s_at
203088_at
201976_s_at
217890_s_at
218027_at
209047_at
218962_s_at
212526_at
221739_at
212274_at
217755_at
206211_at
217483_at
203254_s_at
203524_s_at
200904_at
220753_s_at
205303_at
218961_s_at
209293_x_at
208950_s_at
206874_s_at
50400_at
212501_at
207655_s_at
212587_s_at
219362_at
205304_s_at
200807_s_at
212190_at
213988_s_at
216733_s_at
212922_s_at
204777_s_at
217962_at
209897_s_at
221823_at
212242_at
218194_at
203620_s_at
213713_s_at
206701_x_at
200652_at
203637_s_at
212314_at
213974_at
218557_at
209470_s_at
208309_s_at
202686_s_at
201791_s_at
204990_s_at
219133_at
218298_s_at
210018_x_at
219179_at
213501_at
217996_at
217800_s_at
213438_at
209149_s_at
212344_at
204905_s_at
218499_at
204238_s_at
210084_x_at
220642_x_at
213275_x_at
213280_at
211323_s_at
214315_x_at
201060_x_at
215471_s_at
221755_at
204168_at
201565_s_at
203116_s_at
204749_at
217956_s_at
203295_s_at
209357_at
202071_at
213441_x_at
201069_at
218592_s_at
205051_s_at
222262_s_at
203921_at
215696_s_at
204418_x_at
220892_s_at
208816_x_at
204404_at
204099_at
201890_at
202554_s_at
218261_at
209663_s_at
218996_at
211981_at
208583_x_at
218854_at
202836_s_at
221814_at
212186_at
208944_at
209224_s_at
201601_x_at
203641_s_at
211671_s_at
218923_at
214022_s_at
210541_s_at
201136_at
91816_f_at
209285_s_at
206352_s_at
214071_at
200825_s_at
202760_s_at
202721_s_at
205683_x_at
200093_s_at
209101_at
218546_at
210095_s_at
219166_at
212886_at
222216_s_at
205433_at
218789_s_at
219440_at
218652_s_at
212624_s_at
217825_s_at
203640_at
219301_s_at
204687_at
205757_at
209656_s_at
209164_s_at
213411_at
203517_at
206377_at
209694_at
218223_s_at
207809_s_at
203632_s_at
221345_at
212677_s_at
212570_at
209154_at
202778_s_at
208636_at
203224_at
201560_at
217803_at
204352_at
202961_s_at
201426_s_at
201912_s_at
201328_at
219115_s_at
213675_at
211075_s_at
213010_at
200044_at
211577_s_at
202540_s_at
207134_x_at
220080_at
217764_s_at
217851_s_at
218330_s_at
222118_at
202664_at
214274_s_at
211160_x_at
203629_s_at
210764_s_at
208398_s_at
213005_s_at
201940_at
202551_s_at
214097_at
65718_at
207414_s_at
213001_at
219038_at
204223_at
205768_s_at
218901_at
218605_at
212419_at
221590_s_at
212104_s_at
209502_s_at
202732_at
203931_s_at
208228_s_at
219276_x_at
219922_s_at
216251_s_at
209583_s_at
214157_at
201603_at
218387_s_at
209469_at
222125_s_at
201243_s_at
220980_s_at
217762_s_at
202889_x_at
211535_s_at
203557_s_at
202729_s_at
218865_at
205802_at
208841_s_at
218285_s_at
217758_s_at
216474_x_at
219551_at
212764_at
210371_s_at
201170_s_at
209147_s_at
221760_at
203228_at
212675_s_at
218458_at
219064_at
201543_s_at
214696_at
212202_s_at
216321_s_at
211498_s_at
204430_s_at
207949_s_at
204754_at
211778_s_at
209205_s_at
201579_at
221584_s_at
203594_at
222108_at
200894_s_at
209466_x_at
212474_at
37996_s_at
202939_at
204424_s_at
214437_s_at
208370_s_at
206656_s_at
204748_at
203663_s_at
214266_s_at
200852_x_at
212647_at
212652_s_at
221127_s_at
200947_s_at
202719_s_at
218434_s_at
209016_s_at
209665_at
211985_s_at
211715_s_at
201841_s_at
202941_at
212423_at
203115_at
208949_s_at
209605_at
209436_at
201647_s_at
201369_s_at
211733_x_at
204268_at
202718_at
209655_s_at
212347_x_at
208690_s_at
212204_at
203603_s_at
213244_at
217763_s_at
211417_x_at
205803_s_at
221428_s_at
204971_at
217168_s_at
206433_s_at
209108_at
219410_at
212989_at
212914_at
201825_s_at
212993_at
209228_x_at
203748_x_at
203545_at
206580_s_at
221245_s_at
218824_at
203616_at
204472_at
203124_s_at
205608_s_at
201116_s_at
201430_s_at
210996_s_at
201313_at
220226_at
211562_s_at
201760_s_at
202075_s_at
200654_at
204163_at
209919_x_at
204396_s_at
205925_s_at
202133_at
213812_s_at
209465_x_at
218720_x_at
201215_at
205155_s_at
213924_at
217894_at
218094_s_at
205420_at
207935_s_at
217942_at
204753_s_at
207131_x_at
218162_at
212160_at
204442_x_at
202843_at
213194_at
218654_s_at
203680_at
210547_x_at
205952_at
211297_s_at
213400_s_at
211576_s_at
206391_at
202599_s_at
202403_s_at
217919_s_at
218518_at
217761_at
217437_s_at
201761_at
211965_at
218966_at
209868_s_at
220547_s_at
214104_at
202178_at
210096_at
221923_s_at
205200_at
214109_at
213524_s_at
212694_s_at
209621_s_at
218140_x_at
202949_s_at
201661_s_at
208962_s_at
203630_s_at
205934_at
208523_x_at
209821_at
200698_at
212509_s_at
209905_at
212713_at
201127_s_at
201030_x_at
218388_at
212736_at
212916_at
200696_s_at
203009_at
202822_at
205074_at
202177_at
209109_s_at
212848_s_at
207606_s_at
209542_x_at
203765_at
207266_x_at
214919_s_at
208029_s_at
209917_s_at
201300_s_at
202183_s_at
212288_at
209916_at
204855_at
217043_s_at
204940_at
208783_s_at
212135_s_at
211048_s_at
210427_x_at
207260_at
212667_at
207981_s_at
201893_x_at
207980_s_at
205573_s_at
218582_at
205083_at
212680_x_at
209337_at
214243_s_at
206392_s_at
220030_at
200911_s_at
205003_at
204793_at
219649_at
206631_at
213900_at
213800_at
204170_s_at
213572_s_at
203215_s_at
207016_s_at
217826_s_at
201792_at
218423_x_at
210986_s_at
209302_at
212551_at
217749_at
208637_x_at
203387_s_at
219654_at
214308_s_at
211864_s_at
209836_x_at
200878_at
212816_s_at
200795_at
202016_at
211980_at
215794_x_at
202393_s_at
221610_s_at
205229_s_at
221782_at
211737_x_at
202539_s_at
219935_at
218931_at
204938_s_at
203966_s_at
823_at
201197_at
219090_at
211935_at
202073_at
201691_s_at
201617_x_at
202109_at
204602_at
201900_s_at
214039_s_at
209600_s_at
213258_at
203011_at
220532_s_at
201013_s_at
220765_s_at
220816_at
203370_s_at
220187_at
209550_at
222140_s_at
209863_s_at
213143_at
214761_at
200946_x_at
215813_s_at
218218_at
212361_s_at
204026_s_at
201798_s_at
204567_s_at
212091_s_at
218465_at
200824_at
205309_at
201462_at
208284_x_at
211966_at
201735_s_at
210987_x_at
203138_at
204359_at
206170_at
211813_x_at
221754_s_at
211964_at
201704_at
205128_x_at
200903_s_at
200600_at
220606_s_at
207836_s_at
204143_s_at
213338_at
221788_at
203705_s_at
211494_s_at
201616_s_at
205833_s_at
204030_s_at
218924_s_at
200982_s_at
202061_s_at
214265_at
207431_s_at
201061_s_at
204957_at
213503_x_at
202871_at
206434_at
209113_s_at
209356_x_at
206385_s_at
207826_s_at
205042_at
201590_x_at
203130_s_at
204345_at
203593_at
203638_s_at
221027_s_at
202920_at
216483_s_at
213156_at
201734_at
213293_s_at
212692_s_at
204412_s_at
219395_at
206332_s_at
214446_at
202504_at
205078_at
203710_at
204121_at
212887_at
213423_x_at
218974_at
206069_s_at
216598_s_at
219152_at
200974_at
212573_at
211343_s_at
213943_at
205384_at
212899_at
203892_at
219121_s_at
203571_s_at
202363_at
219747_at
207362_at
210078_s_at
207824_s_at
209118_s_at
209772_s_at
202350_s_at
219933_at
218694_at
207549_x_at
206070_s_at
218556_at
211340_s_at
201660_at
208789_at
202929_s_at
209087_x_at
205316_at
218963_s_at
219555_s_at
204963_at
212282_at
207961_x_at
221927_s_at
209191_at
218531_at
207957_s_at
213148_at
209129_at
200681_at
200930_s_at
202503_s_at
204964_s_at
205566_at
204041_at
209625_at
217767_at
203164_at
221935_s_at
210108_at
213564_x_at
202023_at
202994_s_at
209504_s_at
221872_at
207275_s_at
209488_s_at
222315_at
203562_at
201130_s_at
218224_at
218979_at
209685_s_at
217823_s_at
204731_at
201577_at
219250_s_at
221781_s_at
203498_at
215407_s_at
204036_at
37117_at
203881_s_at
205133_s_at
211126_s_at
205942_s_at
201147_s_at
209367_at
201438_at
215380_s_at
213994_s_at
200970_s_at
214212_x_at
219518_s_at
206938_at
202605_at
213568_at
200971_s_at
205609_at
63825_at
201631_s_at
221874_at
201645_at
205505_at
202440_s_at
212978_at
209496_at
218025_s_at
212977_at
210720_s_at
212067_s_at
206110_at
221541_at
218188_s_at
204364_s_at
204942_s_at
200923_at
201724_s_at
212236_x_at
217111_at
220595_at
208737_at
212813_at
203219_s_at
204284_at
218909_at
218380_at
204019_s_at
208747_s_at
209531_at
212230_at
212295_s_at
203131_at
201417_at
218418_s_at
209855_s_at
201242_s_at
202893_at
205132_at
221024_s_at
204463_s_at
218086_at
200931_s_at
221865_at
204464_s_at
51158_at
209427_at
203386_at
201843_s_at
219411_at
204288_s_at
210719_s_at
202748_at
218258_at
218730_s_at
221880_s_at
202018_s_at
201583_s_at
218980_at
220432_s_at
208966_x_at
209825_s_at
213371_at
202546_at
209209_s_at
222121_at
203706_s_at
211423_s_at
200897_s_at
204388_s_at
205856_at
217736_s_at
209487_at
219850_s_at
221748_s_at
207098_s_at
210869_s_at
204389_at
200907_s_at
200606_at
211896_s_at
215108_x_at
222162_s_at
219388_at
219295_s_at
201196_s_at
209286_at
213085_s_at
209335_at
209478_at
204955_at
200078_s_at
211663_x_at
214733_s_at
212843_at
206860_s_at
202566_s_at
205769_at
205157_s_at
202668_at
204570_at
209030_s_at
204069_at
218248_at
209074_s_at
201014_s_at
200953_s_at
219584_at
201348_at
202005_at
203851_at
211559_s_at
201957_at
206068_s_at
205725_at
206303_s_at
202202_s_at
203029_s_at
212226_s_at
205248_at
213428_s_at
203430_at
208131_s_at
217776_at
201497_x_at
219015_s_at
200621_at
201963_at
213992_at
200700_s_at
211748_x_at
202769_at
218611_at
212181_s_at
207977_s_at
213325_at
212254_s_at
205102_at
207876_s_at
209585_s_at
209948_at
204319_s_at
206116_s_at
208580_x_at
217757_at
200670_at
204273_at
202790_at
204457_s_at
266_s_at
201787_at
204141_at
221505_at
210787_s_at
209651_at
218696_at
201540_at
206770_s_at
204931_at
209514_s_at
200986_at
214106_s_at
202283_at
210480_s_at
200906_s_at
203042_at
209687_at
212744_at
203729_at
210715_s_at
201842_s_at
209934_s_at
218718_at
212448_at
201431_s_at
215432_at
214091_s_at
212115_at
209156_s_at
202428_x_at
202196_s_at
87100_at
202269_x_at
217014_s_at
204400_at
200656_s_at
202007_at
209693_at
201105_at
213892_s_at
219167_at
211596_s_at
209288_s_at
208658_at
201150_s_at
222258_s_at
214505_s_at
203030_s_at
202565_s_at
204394_at
200762_at
220014_at
209616_s_at
208788_at
212136_at
217912_at
214247_s_at
213288_at
203423_at
210293_s_at
209283_at
209031_at
201641_at
211724_x_at
212187_x_at
221589_s_at
213093_at
202148_s_at
217728_at
213712_at
202995_s_at
221019_s_at
201539_s_at
201951_at
204939_s_at
212183_at
210298_x_at
203180_at
204894_s_at
201193_at
205547_s_at
208190_s_at
215016_x_at
201582_at
207030_s_at
203642_s_at
210139_s_at
208527_x_at
209167_at
218211_s_at
219685_at
202770_s_at
209291_at
202826_at
201495_x_at
210951_x_at
213068_at
208180_s_at
203065_s_at
212745_s_at
209351_at
219017_at
205549_at
207843_x_at
209170_s_at
219405_at
203324_s_at
217775_s_at
202222_s_at
205645_at
219478_at
40093_at
202992_at
203717_at
209210_s_at
212252_at
213746_s_at
201079_at
203323_at
204776_at
208791_at
209389_x_at
212768_s_at
210738_s_at
208792_s_at
210041_s_at
204135_at
222067_x_at
205564_at
202688_at
213071_at
201848_s_at
204734_at
210652_s_at
202274_at
205221_at
201058_s_at
203946_s_at
209540_at
209366_x_at
205382_s_at
202088_at
209355_s_at
219266_at
205242_at
202457_s_at
33767_at
210337_s_at
201496_x_at
200832_s_at
201615_x_at
201131_s_at
202722_s_at
209541_at
202786_at
209706_at
212724_at
208546_x_at
204583_x_at
213139_at
202740_at
220933_s_at
212233_at
220926_s_at
214404_x_at
203903_s_at
211070_x_at
213246_at
207480_s_at
213920_at
222209_s_at
208790_s_at
209094_at
200969_at
210299_s_at
220380_at
213285_at
221747_at
215779_s_at
202429_s_at
205935_at
202708_s_at
210387_at
201820_at
213106_at
203911_at
209292_at
200790_at
217875_s_at
212992_at
209911_x_at
221802_s_at
202409_at
208490_x_at
201128_s_at
203766_s_at
204751_x_at
219118_at
203186_s_at
212310_at
219667_s_at
212730_at
203041_s_at
210130_s_at
212097_at
216623_x_at
203739_at
217897_at
214329_x_at
204231_s_at
203951_at
212281_s_at
215726_s_at
200859_x_at
210317_s_at
205052_at
222043_at
217850_at
214765_s_at
221667_s_at
218922_s_at
201849_at
211276_at
213555_at
209460_at
201667_at
201413_at
222277_at
214752_x_at
217752_s_at
213587_s_at
212865_s_at
210222_s_at
210377_at
218087_s_at
204582_s_at
213622_at
203296_s_at
221561_at
222075_s_at
208937_s_at
202286_s_at
202525_at
214027_x_at
74694_s_at
204485_s_at
202555_s_at
209806_at
212543_at
207390_s_at
209163_at
220116_at
209763_at
212255_s_at
214774_x_at
204083_s_at
205924_at
203304_at
208650_s_at
218035_s_at
203644_s_at
201596_x_at
217901_at
205597_at
214463_x_at
209844_at
219127_at
217973_at
201562_s_at
209459_s_at
219117_s_at
202427_s_at
218254_s_at
214290_s_at
221582_at
214469_at
209696_at
219312_s_at
216905_s_at
209623_at
200935_at
219736_at
203485_at
211137_s_at
202687_s_at
46323_at
212640_at
219856_at
202089_s_at
218186_at
218189_s_at
206302_s_at
214651_s_at
212686_at
201952_at
203007_x_at
215017_s_at
202454_s_at
208837_at
206558_at
203857_s_at
202043_s_at
212812_at
214087_s_at
209935_at
205830_at
201662_s_at
209173_at
204973_at
205780_at
200644_at
218280_x_at
204305_at
204875_s_at
220161_s_at
209369_at
201923_at
202890_at
221732_at
205776_at
208579_x_at
212789_at
219806_s_at
221669_s_at
202489_s_at
218638_s_at
201563_at
217979_at
217080_s_at
36830_at
214455_at
218835_at
210328_at
203954_x_at
211478_s_at
210339_s_at
209340_at
203397_s_at
210788_s_at
220192_x_at
203716_s_at
209114_at
206214_at
209398_at
219476_at
212449_s_at
204667_at
211689_s_at
215071_s_at
203216_s_at
209854_s_at
206858_s_at
203917_at
212445_s_at
205862_at
201690_s_at
200862_at
212412_at
203474_at
203243_s_at
209624_s_at
211303_x_at
212218_s_at
204623_at
201688_s_at
215363_x_at
205542_at
205347_s_at
201839_s_at
219360_s_at
202345_s_at
203196_at
213506_at
203953_s_at
218313_s_at
205860_x_at
214598_at
216920_s_at
221424_s_at
215806_x_at
217487_x_at
221577_x_at
216804_s_at
211144_x_at
201689_s_at
209813_x_at
204934_s_at
209425_at
217771_at
209426_s_at
203908_at
209424_s_at
203242_s_at
201739_at
209854_s_at
33322_i_at
209706_at
205780_at
205780_at
201577_at
202404_s_at
209706_at
200795_at
211194_s_at
200931_s_at
214800_x_at
201615_x_at
202088_at
207169_x_at
205541_s_at
202436_s_at
209854_s_at
203084_at
209283_at
207956_x_at
202088_at
201995_at
202088_at
205645_at
209771_x_at
201577_at
202089_s_at
201394_s_at
209771_x_at
202525_at
201839_s_at
214460_at
205834_s_at
209935_at
211834_s_at
221788_at
210930_s_at
212230_at
202089_s_at
201409_s_at
201555_at
33322_i_at
217487_x_at
201744_s_at
201215_at
211748_x_at
221788_at
215564_at
201555_at
33322_i_at
211964_at
221233_s_at
KIAA1411
216839_at
LAMA2
215231_at
ABP1
216814_at
217321_x_at
ATXN3
216819_at
202865_at
DJB12
206490_at
DLGAP1
207479_at
219688_at
BBS7
220791_x_at
SCN11A
207465_at
AFFX-
PheX-5_at
204884_s_at
HUS1
217392_at
CAPZA1
214702_at
FN1
214636_at
CALCB
208181_at
HIST1H4H
215228_at
NHLH2
220507_s_at
UPB1
205539_at
AVIL
220869_at
UBE1L2
204945_at
PTPRN
217048_at
215053_at
SRCAP
221617_at
TAF9B
214222_at
DH7
210520_at
FETUB
220832_at
TLR8
211310_at
EZH1
221414_s_at
DEFB126
206731_at
CNKSR2
215615_x_at
RERE
222048_at
ADRBK2
212743_at
RCHY1
213631_x_at
HP
222176_at
PTEN
213909_at
LRRC15
215611_at
TCF12
221409_at
OR2S2
220793_at
SAGE1
206730_at
GRIA3
217112_at
PDGFB
215560_x_at
MTRF1L
216422_at
PA2G4
220776_at
KCNJ14
206249_at
MAP3K13
220764_at
PPP4R2
215768_at
SOX5
216536_at
OR7E19P
207615_s_at
C16orf3
203866_at
NLE1
205336_at
PVALB
207254_at
SLC15A1
203998_s_at
SYT1
207236_at
ZNF345
215652_at
214675_at
NUP188
210712_at
LDHAL6B
214655_at
GPR6
221049_s_at
POLL
219997_s_at
COPS7B
219928_s_at
CABYR
204191_at
IFR1
219711_at
ZNF586
215249_at
RPL35A
215868_x_at
SOX5
211402_x_at
NR6A1
214245_at
RPS14
207409_at
LECT2
217612_at
TIMM50
207902_at
IL5RA
210695_s_at
WWOX
216340_s_at
CYP2A7P1
217171_at
SMPD1
214842_s_at
ALB
221905_at
CYLD
205610_at
MYOM1
210197_at
ITPK1
207045_at
FLJ20097
210701_at
CFDP1
212308_at
CLASP2
201763_s_at
DAXX
216661_x_at
CYP2C9
220122_at
MCTP1
211318_s_at
RAE1
205915_x_at
GRIN1
208281_x_at
DAZ1 /// DAZ3 /// DAZ2
/// DAZ4
218564_at
RFWD3
213971_s_at
SUZ12 /// SUZ12P
213957_s_at
CEP350
203839_s_at
TNK2
214283_at
TMEM97
217830_s_at
NSFL1C
207331_at
CENPF
218621_at
HEMK1
207455_at
P2RY1
220444_at
ZNF557
201208_s_at
TNFAIP1
204283_at
FARS2
202885_s_at
PPP2R1B
203383_s_at
GOLGA1
209072_at
MBP
203171_s_at
KIAA0409
202550_s_at
VAPB
205851_at
NME6
217721_at
210005_at
GART
207735_at
RNF125
212087_s_at
ERAL1
222184_at
205238_at
CXorf34
214526_x_at
PMS2L1
219543_at
MAWBP
204883_s_at
HUS1
217094_s_at
ITCH
214756_x_at
PMS2L1
207511_s_at
C2orf24
219854_at
ZNF14
213893_x_at
PMS2L5 /// LOC441259 ///
LOC641799 ///
LOC641800 ///
LOC645243 ///
LOC645248
207505_at
PRKG2
203436_at
RPP30
205829_at
HSD17B1
201905_s_at
CTDSPL
214507_s_at
EXOSC2
209677_at
PRKCI
208676_s_at
PA2G4
207347_at
ERCC6
201961_s_at
RNF41
209029_at
COPS7A
219797_at
MGAT4A
219596_at
THAP10
221984_s_at
C2orf17
222006_at
LETM1
222192_s_at
FLJ21820
217136_at
PPIAL4 ///
202004_x_at
SDHC /// LOC642502
LOC653505 ///
LOC653598
220037_s_at
XLKD1
217586_x_at
206962_x_at
218540_at
THTPA
204111_at
HNMT
215198_s_at
CALD1
214681_at
GK
217931_at
TNRC5
213888_s_at
TRAF3IP3
202801_at
PRKACA
212284_x_at
TPT1
202821_s_at
LPP
203015_s_at
SSX2IP
208157_at
SIM2
204551_s_at
AHSG
218636_s_at
MAN1B1
214327_x_at
TPT1
202924_s_at
PLAGL2
220491_at
HAMP
219222_at
RBKS
210931_at
RNF6
213328_at
NEK1
219901_at
FGD6
214473_x_at
PMS2L3
207503_at
TCP10
210187_at
FKBP1A
219634_at
CHST11
200786_at
PSMB7
212869_x_at
TPT1
209222_s_at
OSBPL2
201319_at
MRCL3
205355_at
ACADSB
219616_at
FLJ21963
214481_at
HIST1H2AM
208018_s_at
HCK
214315_x_at
CALR
213273_at
ODZ4
221838_at
KLHL22
214543_x_at
QKI
216315_x_at
UBE2V1 /// Kua-UEV
213443_at
TRADD
205047_s_at
ASNS
208929_x_at
RPL13
218026_at
CCDC56
221356_x_at
P2RX2
204173_at
MYL6B
209929_s_at
IKBKG
211127_x_at
EDA
220673_s_at
KIAA1622
207831_x_at
DHPS
214649_s_at
MTMR2
218711_s_at
SDPR
206715_at
TFEC
203190_at
NDUFS8
201025_at
EIF5B
202406_s_at
TIAL1
217687_at
ADCY2
52651_at
COL8A2
221447_s_at
GLT8D2
212684_at
ZNF3
209826_at
EGFL8 ///
201791_s_at
DHCR7
LOC653870
212961_x_at
CXorf40B
206667_s_at
SCAMP1
206801_at
NPPB
214117_s_at
BTD
218182_s_at
CLDN1
203368_at
CRELD1
219594_at
NINJ2
218658_s_at
ACTR8
203652_at
MAP3K11
219278_at
MAP3K6
221907_at
C14orf172
207156_at
HIST1H2AG
213688_at
CALM1
214460_at
LSAMP
204989_s_at
ITGB4
65884_at
MAN1B1
202055_at
KP1
221058_s_at
CKLF
217362_x_at
HLA-DRB6
202903_at
LSM5
219055_at
SRBD1
201685_s_at
C14orf92
206987_x_at
FGF18
209231_s_at
DCTN5
201309_x_at
C5orf13
212862_at
CDS2
203017_s_at
SSX2IP
219736_at
TRIM36
203227_s_at
TSPAN31
212283_at
AGRN
207616_s_at
TANK
202186_x_at
PPP2R5A
221901_at
KIAA1644
209527_at
EXOSC2
202302_s_at
FLJ11021
200868_s_at
ZNF313
210933_s_at
FSCN1
209247_s_at
ABCF2
222148_s_at
RHOT1
204089_x_at
MAP3K4
213095_x_at
AIF1
214695_at
UBAP2L
212613_at
BTN3A2
215203_at
GOLGA4
218013_x_at
DCTN4
203189_s_at
NDUFS8
210831_s_at
PTGER3
218830_at
RPL26L1
211776_s_at
EPB41L3
221860_at
HNRPL
212535_at
MEF2A
208523_x_at
HIST1H2BI
201594_s_at
PPP4R1
218996_at
TFPT
58780_s_at
FLJ10357
203593_at
CD2AP
209658_at
CDC16
219125_s_at
RAG1AP1
202000_at
NDUFA6
218403_at
TRIAP1
205479_s_at
PLAU
208490_x_at
HIST1H2BF
211323_s_at
ITPR1
221261_x_at
MAGED4 /// LOC653210
210473_s_at
GPR125
208527_x_at
HIST1H2BE
215051_x_at
AIF1
205501_at
219078_at
GPATC2
209078_s_at
TXN2
212371_at
C1orf121
206110_at
HIST1H3H
200978_at
MDH1
202098_s_at
PRMT2
202286_s_at
TACSTD2
208546_x_at
HIST1H2BH
203705_s_at
FZD7
208579_x_at
H2BFS
216583_x_at
219538_at
WDR5B
210102_at
LOH11CR2A
212744_at
BBS4
203177_x_at
TFAM
214472_at
HIST1H3D
218534_s_at
AGGF1
215779_s_at
HIST1H2BG
204215_at
C7orf23
208180_s_at
HIST1H4H
218454_at
FLJ22662
214469_at
HIST1H2AE
202794_at
INPP1
211474_s_at
SERPINB6
204037_at
EDG2 ///
208583_x_at
HIST1H2AJ
LOC644923
213233_s_at
KLHL9
215978_x_at
LOC152719
212222_at
PSME4
217775_s_at
RDH11
204222_s_at
GLIPR1
213789_at
204456_s_at
GAS1
214455_at
HIST1H2BC
211945_s_at
ITGB1
209210_s_at
PLEKHC1
217798_at
CNOT2
203567_s_at
TRIM38
203854_at
CFI
200982_s_at
ANXA6
216231_s_at
B2M
209901_x_at
AIF1
209083_at
CORO1A
215116_s_at
DNM1
215411_s_at
TRAF3IP2
212314_at
KIAA0746
218047_at
OSBPL9
210273_at
PCDH7
217732_s_at
ITM2B
208070_s_at
REV3L
204150_at
STAB1
208985_s_at
EIF3S1
201278_at
DAB2
209550_at
NDN
213741_s_at
KP1
210285_x_at
WTAP
201887_at
IL13RA1
206117_at
TPM1
213716_s_at
SECTM1
202693_s_at
STK17A
212500_at
C10orf22
219179_at
DACT1
219140_s_at
RBP4
203868_s_at
VCAM1
212294_at
GNG12
204298_s_at
LOX
215313_x_at
HLA-A
205698_s_at
MAP2K6
220955_x_at
RAB23
203300_x_at
AP1S2
209191_at
TUBB6
210915_x_at
TRBV19 ///
TRBC1
200033_at
DDX5
202810_at
DRG1
218396_at
VPS13C
204114_at
NID2
204364_s_at
REEP1
219687_at
HHAT
201590_x_at
ANXA2
209168_at
GPM6B
201060_x_at
STOM
212203_x_at
IFITM3
213258_at
TFPI
202450_s_at
CTSK
204244_s_at
DBF4
210416_s_at
CHEK2
209932_s_at
DUT
208146_s_at
CPVL
203153_at
IFIT1
214252_s_at
CLN5
203961_at
NEBL
204168_at
MGST2
40489_at
ATN1
209034_at
PNRC1
201280_s_at
DAB2
213572_s_at
SERPINB1
212586_at
CAST
203323_at
CAV2
221816_s_at
PHF11
219370_at
RPRM
201506_at
TGFBI
201540_at
FHL1
211429_s_at
SERPI1
218656_s_at
LHFP
210275_s_at
ZA20D2
201842_s_at
EFEMP1
201061_s_at
STOM
209648_x_at
SOCS5
222088_s_at
SLC2A3
203706_s_at
FZD7
201132_at
HNRPH2
210139_s_at
PMP22
212149_at
KIAA0143
214257_s_at
SEC22B
214022_s_at
IFITM1
218741_at
C22orf18
221523_s_at
RRAGD
220595_at
PDZRN4
201601_x_at
IFITM1
202446_s_at
PLSCR1
206662_at
GLRX
201560_at
CLIC4
206332_s_at
IFI16
217741_s_at
ZA20D2
202609_at
EPS8
202936_s_at
SOX9
209154_at
TAX1BP3
203305_at
F13A1
212824_at
FUBP3
208296_x_at
TNFAIP8
209498_at
CEACAM1
217832_at
SYNCRIP
212533_at
WEE1
213193_x_at
TRBV19 ///
TRBC1
204472_at
GEM
205898_at
CX3CR1
200887_s_at
STAT1
209170_s_at
GPM6B
209488_s_at
RBPMS
210986_s_at
TPM1
204036_at
EDG2
208966_x_at
IFI16
202283_at
SERPINF1
203640_at
MBNL2
203810_at
DJB4
210072_at
CCL19
213791_at
PENK
212230_at
PPAP2B
210987_x_at
TPM1
205110_s_at
FGF13
212097_at
CAV1
215716_s_at
ATP2B1
200935_at
CALR
218162_at
OLFML3
201645_at
TNC
203710_at
ITPR1
211864_s_at
FER1L3
204939_s_at
PLN
202430_s_at
PLSCR1
209487_at
RBPMS
202037_s_at
SFRP1
204135_at
DOC1
206991_s_at
CCR5 ///
LOC653725
200836_s_at
MAP4
209167_at
GPM6B
212417_at
SCAMP1
210299_s_at
FHL1
209288_s_at
CDC42EP3
212671_s_at
HLA-DQA1 ///
HLA-DQA2 ///
LOC650946
209684_at
RIN2
201310_s_at
C5orf13
201196_s_at
AMD1
202269_x_at
GBP1
201798_s_at
FER1L3
204955_at
SRPX
201787_at
FBLN1
209687_at
CXCL12
202291_s_at
MGP
219117_s_at
FKBP11
207826_s_at
ID3
218730_s_at
OGN
209291_at
ID4
209541_at
IGF1
204464_s_at
EDNRA
201030_x_at
LDHB
204172_at
CPOX
217546_at
MT1M
203453_at
SCNN1A
203932_at
HLA-DMB
205498_at
GHR
213293_s_at
TRIM22
218087_s_at
SORBS1
205158_at
RSE4
216598_s_at
CCL2
213975_s_at
LYZ /// LILRB1
221510_s_at
GLS
202258_s_at
PFAAP5
205097_at
SLC26A2
202333_s_at
UBE2B
218589_at
P2RY5
202935_s_at
SOX9
213564_x_at
LDHB
214836_x_at
IGKC /// IGKV1-5
204070_at
RARRES3
206392_s_at
RARRES1
218331_s_at
C10orf18
204259_at
MMP7
217028_at
CXCR4
221872_at
RARRES1
201650_at
KRT19
This example relates to the use of linear models to predict the tissue component of prostate samples based on microarray data. This strategy can be used to estimate the proportion of tissue components in each case and thereby reduce the impact of tissue proportions as a major source of variability among samples. The prediction model was tested by 10-fold cross validation within each data set, and also by mutual prediction across independent data sets.
Prostate Cancer Microarray Data Sets:
Four publicly available prostate cancer data sets (datasets 1 through 4) with pathologist-estimated tissue component information were included in this study (Table 13). For all data sets, four major tissue components (tumor cells, stroma cells, epithelial cells of BPH, and epithelial cells of dilated cystic glands) were determined from sections prepared immediately before and after the sections pooled for RNA preparation by pathologists. The tissue component distributions for the four data sets are shown in Table 13.
Four publicly available microarray data sets (datasets 5 through 8) also were collected. These included a total of 238 arrays that were generated from 219 tumor enriched and 19 non-tumor parts of prostate tissue, as shown in Table 14. Dataset 5 consists of two groups (37 recurrence and 42 non-recurrence) for a total of 79 cases. The samples used in these four datasets do not have associated details of tissue component information.
Selection of Genes for Model-Training:
Subsets of genes were selected to train the prediction model using two strategies. In the first strategy, each gene was ranked by the correlation coefficient between its intensity values and the percentage of a given tissue component across all samples. In the second strategy, the genes were ranked by their F-statistic, a measure of their fit in the multiple linear regression model as described below. The two strategies produced very similar results.
Multiple Linear Regression Model:
A multi-variate linear regression model was used for prediction of tissue components. This is based on the assumption that the observed gene expression intensity of a gene is the summation of the contributions from different types of cells:
where g is the expression value for a gene, pj is the percentage of a given tissue component determined by the pathologists, and βj is the expression coefficient associated with a given cell type. In this model, C is the number of tissue types under consideration. In the current study, only β's of two major tissue types, tumor and stroma, were estimated to minimize the noise caused by other minority cell types. The contribution of other cell types to the total intensity g is subsumed into β0 and e. Note that βj is suggestive of the relative expression level in cell type j compared to the overall mean expression level β0. The regression model was used to predict the percentage of tissue components after the parameters were determined on a training data set.
Cross-Validation within Data Sets:
Ten-fold cross-validation was used to estimate the prediction error rates for each data set. Briefly, one tenth of the samples were randomly selected as the test set using a boot strapping strategy and the remaining nine tenths of the samples were used as training set. Prediction models are constructed using the training sets with a pre-defined number of genes selected with the strategy mentioned above. The prediction is then tested on the test set. The sample selection and prediction step are repeated 10 times using different test samples each time until all the samples are used as test samples only once. This whole procedure is repeated five times using different sets of 10% of the data in each iteration to generate reliable results.
Validation Between Data Sets:
Mutual predictions were performed among datasets 1, 2, 3 and 4 to assess the applicability of prediction models across different data sets. Because the microarray platforms differ among the four data sets, quantile normalization are applied to preprocess the microarray data (Bolstad et al. (2003) Bioinformatics 19:185-193) with one modification. Quantile normalization method was applied on the test data set with the entire training set as the reference. This change means that the training set that is used to build prediction models will not be re-calculated and the prediction models will likely stay the same.
The mapping of probe sets from different Affymetrix platforms is based on the array comparison files downloaded from the Affymetrix website (World Wide Web at affymetrix.com). Probe sets of Probes in Affymetrix U133A array are a sublist of those in Affymetrix U133Plus2.0 array, and the DNA sequences of the common probes of two platforms are identical, suggesting these two platforms are very similar. The Illumina DASL platform used in data set 4 only provided gene symbols as the probe annotation, which was used to map to Affymetrix platforms. The numbers of genes mapped among different platforms are shown in Table 15.
Prediction on Data Sets that do not have Pathologist's Estimates of Tissue Proportions:
Datasets 5, 6, 7, and 8 do not have previous estimates of tissue composition (Table 14). Datasets 1, 5, and 6 were generated from Affymetrix U133A arrays. Thus, the prediction models constructed with data set 1 were used to predict tissue components of samples used in datasets 5 and 6. Likewise, datasets 2, 7, and 8 were generated with Affymetrix U133Plus2.0 arrays, so prediction models constructed with dataset 2 were used to predict tissue components of samples used in datasets 7 and 8. The modified quantile normalization method described above was used for preprocessing the test data sets.
Comparison of in Silico Predictions and Pathologist's Estimates within the Same Data Set:
Four sets of microarray expression data for which tissue percentages had been determined by pathologists (Table 13), were used to develop in silico models that could predict tissue percentages in other samples that had array data but did not have pathologist data on tissue percentages. The discrepancies between in silico predictions and pathologist's estimates were measured by the mean absolute difference between values predicted in silico and the observation values estimated by pathologists. Ten-fold cross-validation was used to estimate the prediction discrepancies for datasets 1, 2, 3 and 4. To determine the best number of genes for constructing prediction model, the most significant 5, 10, 20, 50, 100 or 250 genes were compared. The prediction results are shown in
Among the four datasets, dataset 1 has the most similar in silico prediction to the pathologist's estimation, with 8% average discrepancy rate for tumor and 16% average discrepancy rate for stroma using the 250-gene model. This may because: 1) this dataset has four pathologists' estimation of tissue components, which will certainly be more accurate than that by one pathologist; 2) fresh frozen tissues were used which generate intact RNA for profiling; and/or 3) relatively larger sample size. Dataset 4 has the least accurate prediction, which may be because: 1) the dataset was generated from degraded total RNA samples from the FFPE blocks; and/or 2) the total number of genes on the Illumina DASL array platform are much less than that of other array platforms (511 probes versus 12626 or more probe sets for the other data sets).
The predictions of tumor components are slightly better than that of stroma, which may be explained in part by the fact that prostate stroma is a mixture of fibroblast cells, smooth muscle cells, blood vessels et al.
As shown in
Dataset 2 contains twelve laser capture micro-dissected tumor samples, the average in silico predicted tumor components for these samples are 91% in average. Assuming these samples really are all nearly pure tumor then the error rate is 9% or less for these samples, which is close to the average error rates of all samples in dataset 2.
The possibility of predicting of two other prostate cell types—the epithelial cells of BPH and dilated cystic glands by extending the current multi-variate model—also were explored. It was found that in silico prediction on these two tissue components are much less accurate than tumor and stroma component, largely because their percentage values are usually small and the pathologists differed in their estimates of these tissues. The extended prediction model including these tissues also slightly lowers the prediction accuracy of tumor and stroma components.
In the original study for dataset 3, agreement analysis on the tissue components that were estimated by four pathologists were assessed as inter-observer Pearson correlation coefficients. The average coefficients for tumor and stroma were 0.92 and 0.77. This is better than the correlation coefficients between in silico prediction and pathologist's estimation for the same dataset, which is 0.72 for the tumor component and 0.57 for stroma component. However, pathologists reviewed the same sections and the tissue components of the adjacent but non-identical samples processed for array assay may differ.
One indication that the prediction model may be optimized to the limits of the data available is the fact that the discrepancy between in silico predicted tissue components and pathologist's estimate for the predictions made on the test sets is often barely 1% different from that of the predictions made on the training set. See the example of 250-gene model as below. Data on other models were very similar.
Data set 1 (training/test): tumor 7.6%/8.1%; stroma 11.7%/12.8%.
Data set 2 (training/test): tumor 8.4%/9.5%; stroma 11.5%/12.5%.
Data set 3 (training/test): tumor 10.3%/11.4%; stroma 15.2%/17.3%.
Data set 4 (training/test): tumor 11.9%/12.5%; stroma 14.7%/15.4%.
To construct the best prediction models from each data set, a 10-fold permutation strategy was adopted to select the most suitable genes to be used in the final prediction model. To construct a n (i.e., 5, 10, 20, 50, 100, 250) gene model for each data set, only nine tenths of randomly chosen samples were used in the multi-variate linear regression analysis for selecting the n most significant genes. This step was repeated nine more times until all the samples were used nine times, which also means that all samples were skipped once. All selected genes (n×10) were pooled and ranked by their incidence. The n genes with the most hits, which are listed in Table 18, were used to construct prediction models that are integrated into CellPred program, as described below.
Comparison Between in Silico Predictions Across Data Sets and Pathologist's Estimates:
Discrepancies for predictions made across different data sets are shown in Table 19. The 250-gene model is used for the mutual prediction. The prediction models constructed on fewer genes also were performed, and the prediction was less accurate than the 250-gene model. In general, the in silico predictions across different datasets are less similar to the pathologist's estimates than the in silico prediction made within the same dataset. However, the discrepancy in predictions across datasets is similar to the discrepancy within datasets when the array platforms are very similar (Affymetrix U133A and U133Plus2.0) and sample types are the same (i.e., fresh frozen sample). For the example of datasets 1 and 2, the prediction discrepancy is 11.0% for tumor and 16.7% for stroma when data set 1 was used as a training set, whereas vice versa, the numbers are 11.6% for tumor and 11.8% for stroma. In the case that microarray platforms and sample types vary (between fresh frozen and FFPE, for example), the cross data set prediction error rates increase and vary largely from 12.1% 28.6% for tumor and 14.7% to 38.2% for stroma depending on the comparison. The mutual prediction results strongly suggest that the feasibility of tissue components prediction across data sets when array platform and sample type are the same. For other cases, prediction of tissue percentages is also possible, but has a large error.
In Silico Prediction of Tissue Components of Samples in Publicly Available Prostate Data Sets:
The in silico predicted tumor and stroma components of 238 samples used in datasets 5, 6, 7, and 8 are documented in Table 17. When 219 of 238 samples were prepared as tumor-enriched prostate tissue, the in silico predicted tumor proportions for these 219 samples showed a wide range from 0 to 87% tumor cells. There are 44 (20.1%) samples predicted with less than 30% tumor cells, as shown in
Dataset 5 includes information regarding recurrence of cancer after prostatectomy for patients, which was used to divide the samples into two groups for comparison (Stephenson, supra). The average tumor tissue component predicted for the recurrence group (58.5%) was noted to be about 10% higher than that of non-recurrence group (48.0%), as shown in
To further illustrate this effect, the percentage of tumor predicted on dataset 5 using the dataset 1 in silico model was plotted as the x axis in a heat map with the non-recurrence and recurrence groups plotted separately. The Y axis consists of the expression levels in data set 5 of the top 100 (50 up- and 50 down-regulated) significant differential expressed genes between tumor and normal tissue identified in dataset 6. The gradient effects from left to right on two groups (non-recurrence and recurrence group) of samples from dataset 5 shows that expression levels of tissue specific genes selected from dataset 6 greatly correlate with the in silico predicted tumor contents with the prediction models developed from dataset 1. Moreover, samples in the recurrence group show slightly higher expression levels in up-regulated genes and lower expression level in down-regulated genes (also shown in
Software for Prostate Cancer Tissue Prediction:
CellPred, a web service freely available on the World Wide Web at webarraydb.org, was designed for prediction of the tissue components of prostate samples used in high-throughput expression studies, such as microarrays. CellPred was developed on a LAMP system (a GNU Linux server with Apache, MySQL and Python). The modules were written in python (World Wide Web at python.org) while analysis functions were written in R language (World Wide Web at r-project.org). The R script for modeling/training/prediction is downloadable from the World Wide Web at webarraydb.org/softwares/CellPred/. Users have the option to choose the number of genes for constructing the model. Genes used for generating the model are provided as an output file. Other details about the program can be found in the online help document.
Users can upload their own data sets for construction of prediction models. However, as an example, data has already been uploaded to allow prediction models constructed on datasets 1, 2 and 3 to be used for making predictions for a user-supplied data set. The user needs to upload the Affymetrix Cel file or any other type of microarray intensity file processed appropriately to make it compatible for making predictions. The most accurate prediction is made for Affymetrix U133A, U133Plus2.0 and U95Av2 array data using the prediction models developed on dataset 1, 2, or 3 respectively. For all other types of microarray platforms, prediction is likely quite noisy. In such cases, probes/probe sets on the platform of the test sets will be mapped to the probes on the training set of choice based on the gene symbols, gene IDs (i.e. GenBank IDs, refSeq IDs) or a mapping file (Xia et al. (2009) Bioinformatics 25:2425-2429). Modified quantile normalization is integrated for preprocessing the intensity values of the test arrays. Then the prediction is made on the test sets using the prediction models constructed with the training set. High-throughput expression sequence tags are accepted by the program if the data are condensed into a file equivalent to an intensity file, along with gene names or IDs that can be mapped to the training data sets.
Homo sapiens, clone IMAGE: 4413783, mRNA
Homo sapiens, clone IMAGE: 4512785, mRNA
Homo sapiens, clone IMAGE: 5309572, mRNA
Genes specifically expressed in different cell types (tumor, stroma, BPH and atrophic gland) of prostate tissue were identified.
Using linear models based on a small list of tissue specific genes, the tissue components of samples hybridized to the array is predictable. These genes are listed in Table 20.
Some tissue specific genes showed significant expression level changes between relapse and non-relapse samples. The gene list is shown in Table 8 above.
cerevisiae)
cerevisiae)
Cancer gene expression profiling studies often measure bulk tumor samples that contain a wide range of mixtures of multiple cell types. The differences in tissue components add noise to any measurement of expression in tumor cells. Such noise would be reduced by taking tissue percentages into account. However, such information does not exist for most available datasets.
Linear models for predicting tissue components (tumor, stroma, and benign prostatic hyperplasia) using two large public prostate cancer expression microarray datasets whose tissue components were estimated by pathologists (datasets 1 and 2) were developed. Mutual in silico predictions of tissue percentages between datasets 1 and 2 correlated with pathologists' estimates for tumor, stroma and BPH (pairwise comparisons for each tissue p<0.0001). The model from dataset 2 was used to predict tissue percentages of a third large public dataset, for which tissue percentages were unknown. Then datasets 1 and 3 were used to identify candidate recurrence-related genes. The number of concordant recurrence-related markers significantly increased when the predicted tissue components were used. The most significant candidates are listed herein. This is the first known endeavor that finds genes predicative of outcome in two or more independent prostate cancer datasets. Given that tumors are highly heterogeneous and include many irrelevant changes, some markers in adjacent stroma or epithelial tissues could be reliable alternative sensors for recurrent versus non-recurrent cancers. The candidate biomarkers associated with recurrence after prostatectomy are included here.
Previously, a modification of the linear combination model of Stuart et al. 2004 was demonstrated and validated. This method is then employed to correct the independent data to that expected based on cell composition. The corrected data is used to validate genes discovered by analysis of the data to exhibit significant differential expression between non-recurrent and recurrent (aggressive) prostate cancer. The biomarkers of this and previous approaches are compared.
Herein, the result of further manipulation of the data is presented in Table form. A list of genes is provided that cross validate across the U01/SPECS dataset (dataset 1, which has tissue percentage estimated) and the dataset of Stephenson et al. (supra), dataset 3 where tissue percentages are estimated by applying a model based on tissue percentages in Bibilova et al. (supra).
Previous reports summarized efforts toward the development of enhanced methods and specification of genes for the prediction of the outcome of prostate cancer. The current report summarizes continued development of predictive biomarkers of Prostate Cancer.
The goals of this study are to continue development of predicative biomarkers of prostate cancer. In particular the goal of the work summarized here is to use independent datasets to validate genes deduced as predictive based on studies of dataset 1 (infra vide). Here “dataset” refers to the array-based RNA expression data of all cases of a given set together with the clinical data defining whether a given case recurred or remained disease free, a censored quantity. Only the categorical value, recurrent or non recurrent, is used in the analyses described here.
For the purposes of the present work, recurrent prostate cancer is taken as a surrogate of aggressive disease while a non-recurrent patient is taken as indolent disease with a variable degree of indolence that is directly proportional to the disease-free survival time. The dataset 1 contains 26 non-recurrent patients, 29 recurrent patients, the dataset 2 contains 63 non-recurrent patients, 18 recurrent patients, and the dataset 3 contains 29 non-recurrent patients and 42 recurrent patients. The data used for this analysis are subsets of previous datasets. Only samples containing more than 0% tumor and follow-up times longer than 2 years for non-recurrent and 4 years for recurrent cases were included for this particular analysis. The first two datasets' samples have various amount of different tissue and cell types, including tumor cells, stroma cells (a collective term for fibroblasts, myofibroblasts, smooth muscle, and small amounts of nerve and vascular elements), BPH (epithelial cells of benign prostate hypertrophy) and dilated cystic glands (AKA “atrophic” cystic glands), as estimated by four pathologists (Stuart et al., supra) for dataset 1 and one pathologist for dataset 2. Dataset 3 samples were tumor-enriched samples, as claimed by the authors (a coauthor of that study, Steven Goodison, is also a coauthor of Stuart et al. PNAS 2004). In this study, published datasets 2 and 3 were used for the purpose of validation only. A major goal of this study is to use “external” published datasets to validate the properties deduced for genes based on analysis of the dataset 1.
Linear regression analysis was performed on the SPECS (dataset 1) and Goodison (dataset 3) arrays, separately. Estimates of significance of association with recurrence were determined as described in previous updates. The accompanying table filters this data as follows. First, genes associated with recurrence with p<0.1 in any tissue in either dataset were retained. Those genes that showed expression changes that were concordant between datasets were retained. However, the confidence in tissue assignment is not great because stroma and tumor tissue percentages are naturally anti-correlated. Thus, the data was also filtered for genes with p<0.1 which appeared to move in opposite directions in these two tissues across datasets as these are about as likely to be real changes and concordant changes in one tissue across datasets. In addition, genes that had a p<0.01 in one tissue in one dataset were also retained even if the other dataset did not show a significant change, if the fold change in either stroma or tumor was consistent across datasets and there was at least a two-fold change in both datasets. Following these procedures and criteria we observed the results listed in Table 21.
This is the first known endeavor that finds genes predicative of outcome in two or more independent prostate cancer datasets. In addition, some of the identified prognosticators are likely to occur in stroma or in BPH rather than in tumor. Such markers in stroma or BPH may be more easily observed as these tissues are more prevalent and more genetically homogeneous than tumor cells.
Datasets Used in this Study
The two datasets used for this study include 1) 148 Affymetrix U133A arrays from 91 patients we acquired (publicly available in the GEO database as accession no. GSE8218, not otherwise published, also referred to as “our data”) which is the principal data set utilized in previous studies; 2) Illumina (of Illumina Inc., San Diego) beads arrays data from 103 patients as analyzed on 115 arrays, a published data set (Bibikova et al., supra);
The two datasets samples have various amount of different tissue and cell types, including tumor cells, stroma cells (a collective term for fibroblasts, myofibroblasts, smooth muscle, and small amounts of nerve and vascular elements), BPH (epithelial cells of benign prostate hypertrophy) and dilated cystic glands (AKA “atrophic” cystic glands), as estimated by four pathologists (Stuart et al., supra) for dataset 1 and one pathologist for dataset 2.
Linear models (Model 1˜3, below) were applied to microarray data from prostate tissues with various amounts of different cell types as estimated by a team of four pathologists. We identified genes specifically expressed in different cell types (tumor, stroma, BPH and dilated cystic glands) of prostate tissue following our published methods (Stuart et al. 2003).
Cell composition can also be considered as two different cell types; one specific cell type versus all the other cell types, grouped together.
G
i=(βtumor·Ptumor+βnon-tumor·Pnon-tumor)i
G
i=(βstroma·Pstroma+βnon-stroma·Pnon-stroma)i
G
i=(βBPH·PBPH+βnon-BPH·Pnon-BPH)i
The correlation (between probe hybridization intensity and tissue percentages) parameters, such as intercept, slope, probability, standard error, was developed for all the genes on the array from model 1, 2 and 3 using dataset 1 and dataset 2.
Using linear models 1-3, the approximate percents of cell types in samples hybridized to the array may be estimated using only the microarray data based on a sub-list of genes on the array. For example, each gene employed in Model 1 provides an estimate of percent tumor cell composition. We used the median of the predictions based on multiple genes for each tissue type. In our case, only a very limited number of the best tissue-specific genes (5-41 genes) were used for the prediction. Even fewer genes might be sufficient.
In order to validate the method of tumor or stroma percent composition determination, we utilized the known percent composition figures of data set 1 to predict the tumor cell and stroma cell compositions for data set 2 with known cell composition. For example, the number of genes used for cell type (tumor epithelial cells, stroma cells or BPH epithelial cells) prediction between dataset 1 and dataset 2 ranges from 5 to 41 non-redundant genes, which are listed in Table 20 herein. The Pearson correlation coefficient between predicted cell type percentage (tumor epithelial cells, stroma cells or BPH epithelial cells) and pathologist estimated percentage ranges from 0.45˜0.87.
Since dataset 1 and dataset 2 data were based on different array platforms, the cross-platform normalization were applied using median rank scores (MRS) method (Warnat et al., supra).
The method of deducing cell type percentage from array data of whole prostate tissue as illustrated here is claimed as novel.
Dietary factors have long been considered major factors influencing the development and progression of prostate cancer and Dr. Gordon Saxe of UCSD has published small scale clinical trials showing that diet and life style alterations have a significant impact on the progression of relapsed prostate cancer (Nguyen, Major et al. 2006); (Saxe, Major et al. 2006)). The UCI SPECS study has accepted a “piggy back” project funded by a subcontract from UCSD (G. Saxe, P I) for carrying out a computerized survey of dietary habits of all patients recruited into the SPECS trial at UCI and UCSD. The questionnaire is self administered by providing a laptop computer to postoperative patients and is directly transmitted to Viocare (world wide web at viocare.com), the developers for the questionnaire, where the results are evaluated and provided with comparative statistics for study use. Blood samples are obtained and assessed for carotenoid carotenoids, vitamin D, and other dietary markers (as a validation of reported habits), as well as sex steroid hormones, IG-1, IGFBP-3, and cytokines. Body mass and BMI is measured by standard anthropometry and dexascanning will be introduced shortly to enable more precise evaluation of body composition. The information will be used to independently model diet/nutrition—disease outcome associations and also correlated with our gene expression results to examine diet-gene interactions.
Bioinformatics Identification and Technical Validation of Expression Biomarkers Using Independent Test Sets of Prostate Cancer Cases.
This is focused on the technical and experimental validation of candidate genes that have been identified as differentially expressed in relapsed (aggressive) and non-relapsed (indolent, good prognosis) prostate cancer. Efforts utilized standard approaches such as recursive partitioning (Koziol 2008)PAM, and VSM to identify potential biomarkers. These efforts showed that genes could be defined that preferentially identified cases that relapse early, within two years of prostatectomy, but were not general. This may be due to the heterogeneity of expression in prostate cancer and the need to identify different signatures for different subclasses of prostate cancer, i.e. the development of a true classifier drawn from the appropriate signatures. Efforts have led to significant progress toward this goal. Two factors are particularly significant. First we have made extensive use of multiple linear regression (MLR) analysis first developed by us for analysis of expression of prostate cancer during the predecessor “Director's Challenge” project (Stuart 2004). Second, we have utilized our data set of 147 U133 arrays together with five additional independent data sets of expression data (Table 22). The data sets of Table 22 are a unique resource for validation. The extended MLR approach provides for determining cell-type specific gene expression for four cell types in non-relapsed prostate cancer cases and for the determination of significant changes in expression for the four cell types for relapsed cases, i.e. significantly differentially expressed genes by cell-type in high risk cases. This model is summarized in equation 1:
G
i=β′tumor,iPtumor+β′stroma,iPstroma+β′BPH,iPBPH+β′dilcys gland,iPdilcys gland+rs(γtumor,iPtumor+γstroma,iPstroma+γBPH,iPBPH+γdilcys gland,iPodilcys gland) (eqn. 1)
where Gi is the observed Affymetrix total Gene expression, the β are the cell-type specific expression coefficients, the P's are the percent of each cell type of the samples applied to the arrays, and the γ's are the differentially expressed component of gene expression for the relapsed cases. When rs=0, no relapse cases are included and the equation is that for gene expression by nonrelapse cases only. The percentages, P, may be determined by examination of H and E slides of the tissue used for RNA preparation by a team of four experienced pathologists. Only two of the six data sets (our cases and those of the Illumina data set, Table 22) have had P's determined by pathologists. Therefore it was first necessary to estimate the percent cell type distribution in all cases of the other four data sets. This was done by using profiles of 40-80 genes for each cell type identified as described (Stuart 2004) that do not vary whether a case is relapse or nonrelapse and are independent of Gleason etc. This method was validated by predicting the percent tumor and stroma cell content of the cases of the Illumina data set which confirmed that the method was accurate (Wang 2007; Wang 2008).
We then applied equation one to our data to identify genes with significant (p<0.01) differential expression in relapsed cases. To validate these genes the process was repeated with each of the five data sets. For each data set we considered a gene as validated if (1) the γ again exhibited p<0.01, (2) were represented by identical Affymetrix probe sets or mapped probe set, and (3) exhibited the same direction change in differential expression. For the tumor cells and stroma cell probe sets, the magnitude of differential expression (the γ) of the two data sets are highly correlated (rpearson>0.7). Approximately 1000 probe sets were identified that were validated in our data set and one other data set. The number of genes validated in this way is highly significantly greater than the number that may be expected to meet the validation criteria for two data sets by chance. These probe sets represent approximately 693 unique genes owing to a number of genes that were validated in two or more pairs of data sets. Numerous genes correspond to those previously reported by others as related to outcome in prostate cancer and these and many others are functionally related to processes thought important in the progression of prostate cancer. For example several members of the Wnt signal transduction pathway are apparent and are being examined using the TMA.
Discussion. The statistical and biochemical properties of many of these genes support the conclusion that an important signature of outcome for prostate cancer has been obtained. We believe that this is the first use of multiple independent data sets for the validation of signatures of outcome for prostate cancer. Not all validated genes exhibit significant differential expression on all data sets. This provides a picture of the diversity of expression of genes as they appear in independent data sets. Thus, it is possible to construct a true classifier that represents the diversity of all six data sets and this effort is underway. The recognition of diversity among published data sets by a consistent set of criteria provides an explanation for the difficulty of finding a signature based on analyses of one or two data sets.
Experimental Validation.
As originally proposed, archived prostate cancer cases of the predecessor “Director's Challenge” program that have not been examined by expression analysis are being measured using the U133 plus 2 platform. These cases were recruited in the period 2000-2004. Approximately 25% of these cases have exhibited evidence of relapse. Thus, these cases provide additional valuable material for validating the predictive properties of the recently developed classifiers. The candidate biomarker genes and their ability to function in classifiers identified above will be tested by comparison of the categorization of these new cases with observed survival results. Approximately 300 fresh frozen prostate cancer cases with clinical follow-up have been characterized with respect to tumor content and approximately 80 have sufficient tumor content for analysis. The percent cell-type distribution has been determined by one pathologist and will be refined by use of the four pathologist analysis. Nearly all cases analyzed have yielded excellent RNA and to date 63 cases have been applied to U133 plus 2 arrays and 27 of these cases also have been applied to EXON arrays. Purified RNA and DNA have been banked from all of these cases and may be used, for example, for PCR validation. The analyzed cases were chosen to (2) maximize tumor content and (2) to be approximately equally divided among relapse and nonrelapse cases in order to maximize statistical power for the testing of differential expression. Owing to these criteria, only 15-20 additional cases from the set of 300 will be useful.
The goal of this set of studies is to identify SNP variations and to determine whether particular SNPs correlate with gene expression changes. The potential significance of this study is that SNP sequence maybe determined for any patient from somatic cells such a blood cells or buccal smears. Thus SNP changes that are found to correlate with predictive expression changes may provide to a much more versatile predictive assay. Moreover this information may provide an understanding of the basis of the of the differential expression changes in terms of the properties of location of the correlated SNP.
The platform that is being utilized by D. Duggan is the Illumina one million SNP array and technology. This is the largest coverage array available and provides for sampling of >1 million SNP sequences. The arrays focus on SNP sites near known genes. Over half of all sampled SNPs are within 10 Kb of a gene.
Twenty one nontumor samples from tumor-bearing prostates have been provided and have now been examined on the Illumina platform. These samples are taken from the same 300-case validation set being analyzed by U133 plus 2 and Exon arrays. Approximately equal numbers of know relapse and nonrelapse cases have been provided. All cases have been used to prepare both RNA and DNA. The RNA is archived while the DNA has been applied to the Illumina platform. All cases analyzed have yielded over 90% present calls indicating excellent DNA qc. The data from these first 42 samples will be used for an interim analysis. Owing to the open ended nature of correlating all differentially expressed genes with multiple SNPs, power of the analysis increases with sample numbers and the current plan is to utilize all samples provided to U133 plus 2 arrays to the SNP analysis included relapse and nonrelapse cases.
Tissue Microarray Development.
The goal is to fabricate prostate cancer TMAs to (1) validate newly identified biomarkers, (2) to validate cell-type specific express on the protein level, and (3) to identify antibody reagents for prognostic assay development. To date 494 prostate cancer cases have been provided and 254 have been used for TMA fabrication (Table 23). The major criterion for the selection of cases is that >5 years of survival data be available (except for normal prostate controls) and most of the cases from UCI and LBVA (Long Beach Veterans Administration Medical Center, an associated hospital of the UCI SOM) have 10-19 years of survival data. The original clinical slides of all cases are examined by two pathologists (P. Carpenter and J. Wang-Rodriquez) who regrade Gleason scores and color-encircle zones for core punching. Cores are taken to represent tumor, BPH, tumor-adjacent stroma, far stroma, dilated cystic glands and, where applicable, PIN. TMA fabrication is carried out at the Burnham Institute for Medical Research (S. Krajewski and J. Reed), All chosen fields are represented by two cores. Thus typically each case is represented by 5×2=10 cores. To date 254 cases array contains ˜1000 cores. The four cell types are placed on separate slide arrays so that specialized studies of one cell type do not needlessly consume material. The 494 cases that have been collected for the TMA are entirely independent of all other cases of this study. For approximately two dozen “Director's Challenge” cases that have been used for U133 plus 2 expression analysis there is FFPE tissue which will be applied to the TMA as a means of directly comparing RNA expression and IHC results.
In addition to multiple cell types, several unique features are being developed. Normal prostate control tissue is being incorporated to represent the same cell types as for the cancer cases. These are provided by Sun Health Research Institute (T. Beach and J. Rodgers) based on their rapid autopsy program. These cases are carefully vetted by two pathologists (P. Carpenter and J. Wang-Rodriquez). In addition the time from death to freezing for all cases is recorded and averages 4.25 h for all 65 cases acquired so far but 3.9 h for the cases of the last year. As a further assessment of quality, RNA has been assessed using the Agilent Bioanalyzer for 38 cases (Y. Wang and H. Yao) which indicates intact RNA in 80% of cases and degraded RNA in 10% of cases. Thus, these normal prostates promise to provide an extensive and approximately age-appropriate control panel. A small number of cases contain prostate cancer and may provide an opportunity to determine protein expression differences between clinical and occult disease.
Another unique feature of the TMAs is the collaborative development of quantization being carried out between the BIMR and Aperio Biotechnologies of San Marcos, Calif. This system provides very high resolution line scanning which is stored on a devoted server at BIMR. Specialized software allows retrieval of high power images of any field for remote viewing by participating pathologists via a secure web-based portal (Scancope). Thus finished TMAs are being examined by two pathologists to determine that selected cores indeed represent the Gleason pattern and cell type intended. Moreover, the software provides a database for the survival data associated with each case. Algorithms have been developed by Allen Olson and colleagues of Aperio for the separation of two colors of TMAs labeled with two antibodies developed with different chromagens. In this method a standard antibody that identifies tumor such a AMACR is used for IHC in parallel with a test antibody (second color). Only pixels of the test antibody labeling that colocalizes with AMACR are then selected for correlation with survival data. An example of two color separation using our TMA was published recently (Krajewska, Olson et al. 2007). Quantification is in advanced stages of development.
Numerous antibodies have been screened for use on FFPE sections and 36 have been optimized, applied to one or more of the TMA slides, and digitized as summarized in Table 24. Several antibodies with known behavior in prostate cancer (anti-PSMA, AMACR, E-Cadherin, beta-Catenin, etc.) have been chosen to characterize the arrays while others (anti-Frzd7. SFRP1, PAP, ANX2, etc.) correspond to predicative biomarkers of this study. A number of apoptosis related biomarkers have be identified and the use of BCL-B as a biomarker in prostate and other epithelial tumors has been published recently (Krajewska 2008; Krajewska 2008b).
It is planned to (1) emphasize visual and electronic scoring of the IHC-labeled TMA, (2) validate electronic scoring and (3) evaluate the relationship of antibody labeling and outcome parameters using the Cox-proportional hazard analysis of Kaplan-Meier plots. A second priority will be to continue to expand the TMA to the full 594 case array.
Prognostic Test of Predicative Gene Profiles.
The goal is to recruit new prostate cancer cases and utilize fresh surgical specimens and biopsies to assess outcome using the current predictive gene profile and to prospectively compare the predicted outcome to observed outcome during year five and as a follow-on long term project. Cases for this study are being recruited in four centers: NWU, UCI, UCSD (SDVA and Thornton Hospitals), and SKCC (Kaiser Permanent Hospital, San Diego). In addition, plans are underway to add the UCI-associated hospital in Long Beach, LBVA. The total number of cases recruited over the past year and from the inception of the study is summarized in Table 25 and associated Demographic, Grading, and Staging data is summarized in Tables 26 and 27. Nearly 1500 cases have been recruited by informed consent to date, over 1300 frozen tissues obtained of which approximately 520 contain tumor. The original goal is to validate selected biomarkers by PCR. Should array costs continue to decrease it may be possible to carryout complete pangenomic expression analysis. By present RNA requirements, conservatively 260 samples would support this effort. Many of these cases have provided blood and post-DRE urine specimens (Table 25) as a further basis for the determination of biomarker expression in more accessible fluids. Shadow charts with baseline data and follow-up data are being developed for all cases.
Diet SPECS Study.
Patients being recruited for the prostate cancer prospective are being consented to participate in the “piggy back” SPECS diet survey study. To date 27 cases have been consented of which 21 have had blood drawn and provided to the NIH-sponsored General Clinical Research Centers of USCD and UCI (Table 28). In addition 8 patients have completed the computerized questionnaire (Table 28). It is the planned to extend the UCI study to include a second clinic of Dr. D. Ornstein at UCI in addition to the present clinic of A. Ahlering and to continue to enroll all future patients that will be recruited for the prospective study at UCI and UCSD over the coming year. A longer range goal of this study is to utilize the present observational study as a proof of principle that sample acquisition and data base resources are available for the development of a potential phase II trial in which relapsed patients may be offered participation in a randomized intervention trial to test the efficacy of diet and life style change to modify the subsequent course of disease. This initiative will require the development of a new proposal for follow-on funding to the SPECS study.
The goal of these studies remains the development of a multigene profile that identifies at the time of diagnosis, prostate cancer patients with poor prognosis and good prognosis. Biomarkers have been identified that are validated in at least one independent data set of six data sets available. Moreover the biomarkers represent the diversity of expression among independent data sets. Thus, a true classifier may be formed for the prognosis of prostate cancer.
Current biomarker information is be utilized to develop a test based on the use of FFPE patient tissue, a widely available resource, that may provide improved guidance for prostate cancer patients.
A 254-case TMA is being used to validate selected biomarkers at the protein expression level. The TMA is composed of cases that are independent of the cases utilized to define the biomarkers. Antibodies that perform well may be useful reagents for the development of an IHC-based assay for determining outcome using FFPE prostatectomy tissue or using preoperative biopsy tissue.
Pangenomic expression data has been collected on 60 cases archived from the “Director's Challenge” program and 25 of these cases have also been profiled on the Illumina million SNP chip. This analysis will continue and when suitable numbers are available, SNP alterations that correlate with expression changes will be determined in order that blood cells may provide a means to determine susceptibility to expression of genes associated with behavior to define SNPs with predictive properties. SNPs can be assessed from any tissue, buccal smears or prostate cancer. Patients that are reliably recognized as belonging to either of these groups will be provided with increased knowledge of the likely outcome of their disease and, therefore, may opt for a wider and more appropriate spectrum of treatment.
Patients are being recruited for prospective testing. In addition, certain dietary features are being determined by questionnaire and blood analysis. Patient of this cohort that relapse but do not seek immediate hormonal or radiation therapy may be offered a diet-life style intervention trial. In particular, the over use of radical prostatectomy may be reduced at considerably decreased morbidity, anguish, and expense.
A variety of efforts have been initiated to translate the results into practical tests. High throughput gene expression analysis will allow us to use all 1000 probe sets that we have determined have predictive value to assess risk and compare the assessment to the clinical indicators of risk such as preop PSA, Gleason, and stage and well as outcome over the next few years. Strong indications of predictive value will indicate that biopsy samples should routinely be made available in the fresh state for RNA analysis and provide preoperative information about patients at high risk of disease that may not be cured by surgery and may provide guidance of who would profit from adjuvant therapy. Finally, patients that relapse following surgery commonly have slowly rising PSA values (low PSA doubling time) and many specialists do not immediately recommend hormone or radiation treatment. Such cases may be offered a diet regimen. Our current “piggy back” observational diet study may set the frame work for evaluating the role of diet. In addition the gene signature of such patients will be known and correlations may be carried out to assess whether there is a signature predictive of response. Similarly, by correlating the response to treatment with the known gene expression results, other signatures predictive of response-to-therapy may be determined. These possibilities require that our prospective cohort be examined by expression analysis which requires a large number of arrays not provided for in the original proposal. Thus, work with the prospective cohort will require additional funding for continuation of the translation of the SPECS studies and planning needs to focus on this issue.
1a,b
2a
3c
aContains data on tissue percentages.
bThese data sets contain information on follow-up time. Relapse was defined as PSA reaches detectable level after prostatectomy within the first four years. All non-relapse cases were cases followed-up over two years and showed no sign of relapse.
cThese data sets contain information on follow-up time. Relapse was defined as three consecutive PSA increases >0.1 ng/ml within the first four years. All non-relapse cases were cases followed-up over two years and showed no sign of relapse.
dNumber of target transcripts represented on the array.
indicates data missing or illegible when filed
The Challenge of Developing Predictive Signatures for the Outcome of Newly Diagnosed Prostate Cancer Based on Expression Analysis and Genetic Changes of Tumor and Non-Tumor Cells
Linear regression analysis was used to determine the average gene expression profile of four cell types, including tumor and stroma cells, in a set of 88 prostatectomy samples (1). By combining these cases with 55 additional cases with Affymetrix U133A gene expression data, we were able to select 63 cases in which disease relapsed over a period of three or more years following prostatectomy. Linear regression analysis of the non-relapse and relapse sets revealed changes in hundreds of gene expression values, including genes primarily expressed in stroma cells that were associated with the relapse status. These genes were used to generate classifiers using two other independent Affymetrix expression datasets generated from enriched prostate tumors. One dataset of 79 samples (37 relapse, Affymetrix U133A array; training-set) was used as the training set (2), and one dataset of 48 samples (23 relapse, Affymetrix U95Av2/U95B/U95C array was used as the test-set (3). Probe sets across platforms were mapped using the Affymetrix array comparison spreadsheet and normalized using quantile discretization (4). Classifier genes were determined by use of recursive partitioning (RP) in which a handful of genes are used sequentially for classification (5), as well as Prediction Analysis of Microarrays (PAM)(6), in which case outcomes were predicted via a nearest shrunken centroid method from gene expression data (1). RP classification trees using up to five genes, and sometimes including pre-operative PSA, routinely classified each independent dataset into three survival groups, non-relapse, early relapse, and late relapse with p<0.005. Classifiers generated by PAM using tumor specific genes predicted by linear regression as input was as good (accuracy, sensitivity, specificity) as the best classifiers using all of the expression data, indicating an enrichment for relevant genes by the linear regression method (SVM was dropped from here since it did not perform better than PAM). However classifier performance decreased with increased disease-free survival of the cases. A 59-gene classifier determined by PAM using all cases of the training set with times-to-relapse of <2 years yielded a specificity of 75.9% and a sensitivity of 88.0% with an overall accuracy of 73.4% when tested with the second independent data set for cases of the same time period. All three performance values decreased continuously upon inclusion of longer time periods to <4 y. No reliable PAM classifiers could be generated for late relapse cases. RP consistently yielded a major group of nonrelapse cases and two classes of relapse cases, one of which consists of very early relapse cases with disease-free survival of <2 years. The distinction of late relapse cases from nonrelapse cases using PAM remains a challenge and may reflect the similarity of gene expression profiles of nonrelapse cases from those destined to relapse relatively late after diagnosis. Prediction of early relapse at the time of diagnosis may be a realistic goal. 1. Stuart, R., et al. PNAS 2004; 201:615-20; 2. Stephenson et al. Cancer. 2005; 104:290-8. 3. Yu Y., et al. J. Clin. Oncol. 2004; 22:1790.4. Warnat, P., et al. BMC Bioinformatics. 2005; 6:265. 5. Koziol, J., et al. Cancer Res. 2003; 9:5120-6. 6. Tibshirani, R. et al. PNAS 2002; 99:6567-72.
Prostate cancer is the most common malignancy of males. However, the majority of cases are “indolent” and may not threaten lives. In order to improve disease management, reliable molecular indicators are needed to distinguish the indolent cancer from the cancer that will progress. Statistical methods, such as hierarchical clustering, PAM and SVM, have been widely used for classifier development for various cancers. However, those methods can not be immediately applied to prostate cancer research because the tissue samples collected from patients are very heterogeneous in cell composition. The observed expression level of any gene for a given sample is not solely for tumor cells; rather, it is the sum of contributions from all types of cells within that sample. In current study, we propose a novel method where the expression level of any gene is illustrated with a linear model considering the contributions from different types of cells and their interactions with aggression phases (relapse or non-relapse). ANOVA is used to identify cell specific relapse associated genes that possess discriminative power. The expression patterns of those selected genes may be described using two Gaussian models on the basis of disease phases; thus they can be used for predicting outcomes of newly diagnosed. The new method is compared to other conventional methods based on simulated data. A predictive classifier is created by training a real dataset generated for prostate cancer research. The performance of the new classifier is compared to the nomogram and other clinical parameters with predictive value.
Differences in RNA levels that correlated with relapse versus non-relapse were calculated for two public expression microarray data sets using two models. One model did not take into account tumor and stroma tissue percentages in each sample, and the other used these percentages in a linear model. The latter model led to a highly significant increase in the number of candidate relapse-associated biomarkers cross-validated between both data sets. Many of these relapse-associated changes in transcript levels occurred in adjacent stroma. Estimates of tissue percentages based on expression data applied between data sets correlated almost as well as multiple pathologists correlated with each other within a data set. This in silico model to predict tissue percentage was applied to a third public data set, for which no tissue percentages exist. Cross-validation of relapse-associated genes between data sets was again highly significantly improved using the linear model, and included changes in stroma. The third data set was heavily skewed towards a previously unrecognized higher tumor percentage in relapse versus non-relapse cases, a bias that is taken into account by the linear model. In summary, the use of tissue percentages determined by a pathologist or inferred from in silico data increased the power to detect concordant changes associated with a clinical parameter in separate data sets, and assigned these changes to different tissue compartments. The strategy should be applicable for biomarkers other than RNA and for samples from any type of disease that contains measurable mixed tissues.
Although many studies of detecting RNA-based prognosticators for prostate cancer have been performed, they have limited agreement with each other. One contributing factor may be the variations in the proportion of tissue components in prostate tissue samples, which leads to considerable noise and even misleading results in mining microarrays data.
We assembled six microarray data sets for RNA expression in prostate cancer samples with associated relapse information, including two large data sets of our own. Our two datasets, and one other, included estimates of tissue percentages made by pathologists. These data sets were used to identify genes that were then used to build a simple linear model for tissue percentage prediction. Estimates of tissue percentages based on expression data applied between data sets correlated almost as well as multiple pathologists correlated with each other within a data set.
Using a multiple linear regression (MLR) model which integrates tissue component percentages, we identified a list of tumor- and reactive stroma-associated prognostic RNA biomarkers in all six data sets. The level of each RNA is expressed as a linear model of contributions from the different cell types and their interactions with relapse status
where g is expression intensity, C is the number of cell types, RS is relapse status indicator, e is random error, and b's and γ's are regression coefficients. ANOVA is used to identify cell specific genes that are differentially expressed between relapsed and non-relapsed cases, i.e., the genes with significant γ's. Markers were then cross-validated between the six different microarray data sets. There were 185 genes that occurred in more than one data set, and 152 of 185 (82.2%) showed the same direction of change in differential expression between relapse and non-relapse patient samples (p<10−18). Most of these prognostic markers were not previously identified by other studies and some were potentially differentially expressed in stroma.
In summary, the use of tissue percentages determined by a pathologist or inferred from in silico data increased the power to detect differential expressed genes associated with a clinical parameter and assigned these changes to different tissue compartments. The strategy should be applicable for biomarkers other than RNA and for samples from any type of disease that contains measurable mixed tissues. A Bi-Model Classifier that Allows RNA Expression in Mixed Tissues to Be Used in Prostate Cancer Prognosis
Introduction:
Reliable molecular indicators are needed to distinguish indolent prostate cancer from cancer that will progress. Statistical methods, such as hierarchical clustering, PAM and SVM, have been widely used to develop classifiers of prognostic molecular markers that estimate risk. However, one barrier to the efficient use of classifiers in prostate cancer is the variable mixture of different cell types in most clinical samples. The observed level of any marker for a given sample is due to the sum of contributions from all types of cells within the tumor. Elsewhere [1], we propose a novel classification method in which the expression level of any gene is expressed as a linear model of contributions from the different cell types and their interactions with relapse status. While this method provides biomarkers with greater confidence by deconvoluting the effect of tissue percentages in each sample, the problem of how to construct a classifier for mixed populations remains.
We propose that the expression patterns of prognostic RNAs may be described using either of two Gaussian models, one for relapsed cases and the other one for non-relapsed cases, both of which include calculation with cell constitute information. A likelihood-ratio statistic (LR) can be developed by contrasting the probability of being risk free to the probability of undergoing relapse based on fitting expression values of selected biomarkers and the cell composition data of each sample to these two differential models. A patient is diagnosed as having high risk of relapse if LR≧k1, or is diagnosed as being of low risk if LR≦k2, where k1 and k2 are pre-selected cutoffs with k1>1>k2.
In a simulation study, the new method outperformed the conventional classification methods PAM and SVM. A prognostic classifier was then created by training an expression dataset generated from Affymetrix U133P2 arrays from prostatectomies with known tissue compostion, which yielded a 50 gene classifier with an accuracy of 94% following cross validation. When the predictive classifier was applied to an independent “test” data set based on 35 Affymetrix U133A arrays, an accuracy of 80% was achieved
This novel classifier may be useful for assessing risk of relapse at the time of diagnosis in clinical samples with variable amounts of cancer tissue.
The prostate tumor microenvironment exhibits numerous differentially expressed genes useful for diagnosis
There are over one million prostate biopsies performed in the U.S. annually. Pathology examination misses the tumor entirely in a few percent of cases. In an additional 10-20% of cases the biopsies are not definitive due to atypical foci, PIN, or other caveats, often leading to a “repeat biopsy” in 6-12 months. We observed that the microenvironment of prostate tumor cells exhibits numerous differential gene expression changes compared to remote stroma tissue of the same cases. Such changes could be useful to form a classifier for the diagnosis of prostate cancer when tumor is present in very low amounts or is barely missed by a biopsy.
A training set of 105 prostate cancer cases was created with known cell type composition for the three major cell types of tumor tissue (tumor epithelial cells, epithelial cells of BPH and stroma cells) as assessed by four pathologists. RNA expression was measured on U133plus2 GeneChips. A linear model defined the total signal as the sum of expression values of the three cell types each weighted by its percent composition figure for a given case:
Gi=βtumor Ptumor+Pstroma Pstroma+βBPHPBPH
where Gi is the fluorescence intensity for a gene of a case, Pi are the percents of the indicated cell type and βi are cell-specific expression coefficients (signal/percent cell type). The model was applied separately to tumor-bearing tissues and tumor-free remote stroma tissues. Differential gene expression was derived by subtraction of the values for the two series.
The ˜200 most significant differences were used as input to PAM. Tenfold cross-validation dichotomized the training set into tumor-bearing and remote stroma tissues, yielding a classifier of 36 genes that had a 94% accuracy. This classifier was then tested using an independent set of 82 cases, as well as 13 control normal prostate stroma tissues. The classifier had an accuracy of 83% on the test set. Correct classification was also achieved for five of six biopsies from normal males and all seven cases from the rapid autopsy. Several genes such as myosin VI, collagen IX, and destrin, known to be highly expressed in mesenchymal derivatives, are preferentially expressed in tumor-adjacent stroma.
The differential gene expression changes observed here most likely represent differences in expression between tumor-adjacent stroma and remote stroma. These differences may be due to paracrine or “field effect” mechanisms involving interaction with the tumor adjacent to the affected stroma. The reaction of stroma to nearby prostate cancer is well-known but, as observed here, involves many more gene changes than previously recognized. These changes can be exploited to develop a classifier that accurately categorizes tumor-bearing tissues, remote tissues of the same cases and normal tissues. Such a classifier could enhance diagnosis from false negative and equivocal biopsy results.
Projects that use antibodies for clinical diagnosis or prognosis must take into account the huge biological differences that occur between patients and between clinical samples. One way to minimize the clinical variation is to use a panel of diagnostic or prognostic antibodies, each of which are known to capture relevant information in a subset of patients or a subset of clinical samples. However, there are also technical challenges that cause difference in staining within and between samples. One way to minimize the impact of technical variation would be to multiplex diagnostic and prognostic markers together with “reference” antibodies that that identify within tissues particular cell type rather than outcomes. These reference antibodies, under the same technical influences and in the same tissue section, can then be used to identify the signals observed for the diagnostic and prognostic antibodies of the relevant cell types which can then be quantified far more accurately than would be possible using separate hybridizations. In the case of prostate cancer, where diagnostic and prognostic antibodies are likely to be relevant in a highly variable and often rare fraction of the cancer cells or adjacent stroma cells in a patient or clinical sample, and where changes from normal tissue may often be subtle rather than “all-or-nothing”, it is likely that only the inclusion of reference antibodies in the same visualization will make it possible to identify the distinct clinically relevant regions with any confidence.
Fortunately, the technology that would be able to perform multiplex antibody staining of individual samples exists with the use of fluorescent dyes. The overall goal over this two phase project is to develop an automated quantitative image-based assay of the expression level of a panel of 5-10 diagnostic and 5-10 prognostic antibody biomarkers in Prostate cancer. Quantification of each antibody biomarker will be carried for specific cell types by utilizing colocalization of each test antibody biomarker of the panel with a reference antibody that is known to specifically identify total epithelium or tumor epithelial cells or tumor-adjacent stroma cells.
In Phase 1 of this project we will focus on the identification and characterization of the reference antibodies that reliably identify total epithelium or tumor epithelium or tumor adjacent stroma in both formalin-fixed and paraffin-embedded (FFPE) and frozen tissue sections. It is likely that a set of reference markers that distinguish different types of epithelial/tumor and fibroblast/smooth muscle stroma, could be useful for automated screening of samples for diagnosis. Phase II will then build on this reference set with additional markers of diagnostic and prognostic use.
In phase I, whole frozen and FFPE sections as well as prostate cancer tissue microarrays (TMAs) will be used to survey candidate reference antibodies and the reproducibility, variability, and accuracy of labeling will be determined for all cases of the TMA as well as by comparison to standard cell lines and normal prostate tissue specimens. This aim is non-trivial as antibodies can have optima for immunohistochemistry that differ markedly from each other. Optimizing a multiplex application may require examining may different types of antibody for each marker as well as a variety of conditions in order to uncover a standard conditions and a standard set of antibodies. Reproducibility, variability, and accuracy of the intensity data will be carefully assessed using positive and negative controls, TMA statistics, and repeated hybridizations on different days for adjacent slices of tissue, including the TMAs. Data storage consistent with the DICOM standard will take place by porting our data to a freeware database and visualization system (ConQuest).
The quantitative properties of the multiplex antibody system will be generated automatically using the proprietary scanning microcytometer developed by Vala Sciences Inc. using multiple fluorphores and validated by comparison to direct visual assessment of the binding location and intensity of representative candidate antibody biomarkers. Each section used for quantitative immunofluorescence (IF) will then be used to prepare DAB (bisdiazobenzidene) chromagen labeled version with hematoxyl counter stain and provided to a panel of four pathologists for estimation of labeling intensity and percent positively labeled epithelial cells or tumor epithelial cells or tumor-adjacent stroma cells. Visual scores for DAB and for fluorescence labeled sections will by quantitative compared to the automated output of the Vala system, using a linear model of the relationship between automated intensity and visual intensity. There is no strict necessity for an antibody to map exactly to a tissue type as assessed by a pathologist, but the scorings should be consistently different for any particular sample, in order to be confident that the antibody is measuring something slightly different, consistently. Zones of authentic tumor and stroma will be defined and the coincidence with colocalized pixels or cells will be quantitatively evaluated.
Workflow will be streamlined and then an SOP created to allow automatic image analysis to be completed with 4-5 days.
Despite advances in our understanding of cancer and the development of new therapeutics, cancer remains the number two killer in the US with mortality rates of many cancers remaining relatively unchanged for decades. Prostate cancer is the most common cancer and second leading cause of cancer-related death among males of Western countries [1-3]. While PSA screening has been a valuable marker increasing early detection of prostate cancer, PSA testing currently suffers from several limitations including lack of specificity and inability to accurately predict disease progression [1, 2, 4-8]. There is a critical unmet need to identify reliable novel biomarkers to assist in early detection of prostate cancer, and, most critically, to determine risk of prostate cancer rercurrence following initial therapy such as prostatectomy. Currently the major treatment modality for newly diagnosed prostate cancer remains radical prostatectomy. Radical prostatectomy provides an excellent outcome for organ-confined disease. However, 15%-20% or more of all surgical patients ultimately experience rercurrence indicating the presence of residual disease, local invasion and/or metastatic deposits at the time of surgery [7-11]. Traditional clinical parameters including tumor staging, Gleason score, and PSA levels, stage or their combinations based on preoperative values have not adequately predicted the patient risk of rercurrence [11, 12]. It is now recognized that prostate cancer exhibits hundreds of altered gene expression changes many of which may represent genes that directly influence outcome [13-19]. However a recent consensus statement by a panel of prostate SPORE leaders (the Inter-SPORE Prostate Biomarkers Study and NBN Pilot group) has tersely summarized that few or none have proven reliable enough to advance to clinical use (http://prostatenbnpilot.nci.nih.gov/aboutpilot_ipbs.asp).
We are developing a new test using novel methods that identify cell-specific biomarkers that can be applied at the time of diagnosis to determine whether the tumor has the potential to recur after surgery. The development of a clinical test capable of distinguishing indolent and aggressive forms of the disease at the time of diagnosis will provide crucial guidance. First, this information will provide guidance as to who needs treatment thereby providing the option of avoiding surgery and the associated morbidity for those patients with a high risk of recurrence. Second, this information will also provide guidance as to who may profit from postsurgery or immediate adjuvant therapy thereby utilizing a period of many months or years during which recurrence otherwise could develop unopposed. Moreover, integration of gene expression signatures with clinical data has recently been shown to improve the accuracy of predicting progression, and metastasis [13, 14, 20]. One purpose of this proposal is the translation of a prostate cancer gene expression classifier into an antibody panel capable of rapid and reliable prediction of disease recurrence using (a) generally available clinical material such as biopsy specimens or, (b) as a guide to adjuvant therapy and patient counseling using post prostatectomy surgical pathology blocks. A crucial advantage of protein markers over RNA markers is that the protein markers provide spatial resolution of cell types and can detect cell-type-localized co-expression of markers, information that is lost in bulk RNA samples.
Moreover there remain critical challenges to diagnosis by biopsy. Over one million prostate biopsies are carried out per year in the U.S. Most are negative. Approximately 20% of these negative biopsies are judged insufficient for a definitive diagnosis owing to small foci or read as “atypical glands” only seen or other ambiguities, i.e. ˜100,000 such cases per year. The microenvironment of these sites contains potential information for diagnosis. We have observed that the tumor adjacent stroma of prostate cancer exhibits hundreds of altered mRNA expression changes and have derived a gene list that accurately identifies tumor adjacent stroma tissue. Thus, antibodies of selected gene products may be potentially useful to assist in diagnosis of traditionally nondiagnositic biopsies.
To date, only a limited number of diagnostic biomarkers that are differentially regulated in prostate carcinoma have been identified such as prostate-specific antigen [2, 5, 6, 23-25], prostate specific membrane antigen [26, 27], and human glandular kallikrein 2 [10, 28-32], and PCA3. While these antigens have been useful in the development of early diagnostics and for the directed delivery of therapeutics to prostate cancer in preclinical models [33, 34] these markers do not address the need to identify biomarkers that characterize early or advanced stages of prostate carcinogenesis and metastasis. Recent studies have identified circulating urokinase-like plasminogen activator receptor forms that may be used alone or in combination with other prostate cancer biomarkers (hK2,PSA) to predict the presence of prostate cancer [35]. Other potential prognostic markers include early prostate cancer antigen (EPCA), AMACR, human kallikrein 11, macrophage inhibitory cytokine 1 (MIC-1), PCA3, and prostate cancer specific autoantibodies [5, 36-42].
The search for novel prostate cancer biomarkers has turned to the use of global genomic and proteomic profiling to facilitate the discovery of multiple markers with both diagnostic and prognostic significance [5, 18, 36-42]. Gene-expression profiling comparing gene expression from normal prostate tissue, BPH tissue, and prostate cancer tissue has identified many potential genes that are differentially regulated in prostate cancer [14, 15]. These include hepsin, a serine protease, alpha-methylacyl-CoA racemase (AMACR), macrophage inhibitory cytokine (MIC-1), and insulin-like growth factor binding protein 3 (IGFBP3) [40], TGF131, IL-6, and many others. Validation of these markers at the protein level from patient tissue or serum samples and clinical validation of these markers as true diagnostic and prognostic tools are necessary. While some of these candidates have appeared in meta analyses (e.g., Rhodes, 2002), as noted, the recent consensus statement of the InterSPORE study has noted that none have proven sufficiently reliable for clinical use and none have been used to form a panel that predicts outcome of multiple independent case sets.
Current clinical parameters including Gleason score, PSA, and tumor staging have been inadequate in predicting patient outcome. Combinations of clinical criteria have been assembled into predictive nomograms in attempts to improve diagnosis of indolent vs. advanced disease [11, 12]. While these studies suggest improved diagnostic and prognostic capabilities, those based solely on preoperative clinical values perform less well and they await widespread clinical validation. One major challenge has been that the majority of prostate cancers share similar histological features (Gleason score) or clinical markers (PSA) but exhibit widely different clinical outcomes. Recently multigene profiles of biomarkers that are predictive of the outcome of prostate cancer at the time of diagnosis have been developed [14, 20, 44-46]. Singh identified a 5-gene classifier capable of predicting prostate cancer recurrence better than clinical parameters of preop PSA or tumor stage [46]. Stephenson identified a set of 10 genes highly correlative with prostate cancer recurrence. An analysis combining clinical variables with the 10-gene classifier greatly improved prediction of clinical outcome [20]. Henshall identified >200 genes that correlate with prostate cancer recurrence better than preoperative PSA [14]. From these studies it is clear that molecular correlates have the potential to provide a considerable increase in information related to outcome than current clinical parameters. In addition to prediction of outcome, it is likely that several of these unique biomarkers are functional and therefore provide intervention opportunities. The proper identification of the molecular determinants predictive of prostate cancer rercurrence, their validation at the protein level, and the translation of the data into a robust clinical test is the challenge addressed in our current proposal. We have developed improvements in both the identification and validation of candidate genes that will enable a rapid and robust transition to a clinical test.
We have developed new methods that have helped in the development of gene signatures for the diagnosis and for prognosis based on expression values of tissue obtained at about the time of the original diagnosis. First, as described herein, we have used a linear combination model together with knowledge of cell composition as determined by a panel of four pathologist to determine gene expression by cell type [18]. These studies revealed cohorts of genes that are differentially expressed by tumor epithelium compared to epithelium of PBH or dilated cystic glands or stroma [18]. This observation has important practical considerations. While most global genome studies have looked at differences between normal and cancerous prostate epithelial cells, considering the contribution of stromal cells as “contamination”, we have found that stroma exhibit dozens of significantly differential gene expression changes between tumor-adjacent stroma and stroma remote from tumor sites [18] and dozens of differential expression changes between tumor-adjacent stroma of recurrent PCa cases compared to nonrecurrent cases [43]; [44]. We have identified two separate subsets of genes. The first consists of tumor epithelium specific and stroma cells specific genes that are differentially expressed between recurrent PCa (“aggressive” cancer, relapsed PCa) and nonrecurrent PCa (“indolent” cancer, nonrelapsed PCa). Since nearly all PCa tissue specimens contain stroma or reactive stroma in the immediate microenvironment of tumor, the proper inclusion of antibodies sensitive to stromal change provides an important ingredient of a “classifier” for prognostic use. These expression changes may be used to predict outcome ([43] [44]).
Second, we have identified a separate subset of tumor-adjacent stroma specific genes. These genes are differentially expressed between tumor-adjacent stroma and remote stroma. These expression changes may be used to detect tumor-adjacent stroma at foci of “nondiagnostic” or “atypical” tumor in biopsies of equivocal cases thereby potentially converting “nondiagnostic” cases to a definitive determination. We propose to use these gene lists as the starting point for the development of panels of 5-10 antibodies for application to biopsy or postoperative FFPE tissue specimens that are routinely available for all patients with a confirmed or suspected diagnosis of prostate cancer. While RNA may be retrieved from these samples, the preservation of a particular set of transcripts with the crucial information in all cases and in proportion to the amounts in fresh tissue is problematic. In contrast, antibody based diagnosis from FFPE is well established. In Phase II we plan to utilize a high throughput scanning microscope to identify the best antibodies for inclusion in the panels. TMAs consisting of 254 prostate cancer cases, normal prostate tissue and defined cell lines will be used for the survey. The TMAs to be used here have been constructed to contain cores especially rich in tumor-adjacent stroma and remote stroma. These cores will allow us to evaluate whether the differential expression observed between relapsed and nonrelpased cases may be observed in adjacent nontumor tissue or even in remote nontumor tissue and to confirm that diagnosis based on tumor-adjacent stroma is reliable. Additional potential applications include the detection of tumor-adjacent stroma in “negative” biopsies that may have narrowly “missed” frank tumor. This possibility is of considerable significance given that most of the million biopsies performed each year are “negative”.
The heterogeneous nature of DNA changes in prostate cancer makes it unlikely that a single biomarker will be adequate for proper determination of prostate cancer severity and risk of rercurrence. What is needed is the identification of a panel of biomarkers that can be shown to correlate with different aspects of disease progression and risk of rercurrence in the population of cancer patients. The screening of tissue by use of microarrays (TMAs) is ideal for identification of markers that statistically correlate with disease progression and outcome [45-48]. Screening of TMAs is a powerful tool for validation of the microarray results, for extension of the RNA expression results to protein expression and for the identification of antibodies of biomarkers that are widely expressed and readily available from samples routinely taken at time of diagnosis. TMAs are constructed using hundreds of different patient samples that span the entire range of clinical pathology and outcome. Furthermore, it requires only small amounts of tissue that can be collected at the time of diagnosis such as biopsy samples and is amendable to high throughput analysis using multiple antibody probes. TMAs may be made from selected archived cases with clinical annotation spanning many years detailing survival and other parameters, such as treatment history.
Numerous studies have used TMAs to identify or validate prostate cancer biomarkers associated with disease progression, response to therapy, rercurrence, and metastasis [45-48, 49, 50]. TMA analysis was used to validate a seven antibody panel derived from a 48 gene expression signature enabling more accurate classification between Gleason grade 3 and 4 tumors [47]. Multiple TMA studies have identified several markers indicative of prostate cancer progression including Amacr (alpha-methyl acyl racemase) AMACR, AR, Bcl-2, CD10, ECAD, Ki67, and p53 [45]. TMA analysis has identified 13 genes associated with prostate cancer rercurrence. These include AKT, □-catenin, NFκB, Stat-3, hMSH2, Hepsin, PIM1, syndecan-1, Bcl-2, Ki67, and ECAD [45]. Few have been formed into a coherent predictive panel and evaluated as a panel. Therefore, the performance of a panel compared to individual antibodies and the potential of combinations to overcome the diversity of prostate cancer is unknown. Nearly all studies ignore the stroma although smooth muscle alpha actin has been examined by Rowley and coworkers [51]. Others suffer the caveats noted by interSPORE group. Several, such as AMACR are utilized as an aid to diagnosis in surgical pathology but are not used routinely in risk assessment. We propose the systematic evaluation of over 50 predicted prognostic biomarkers (Phase I and Phase II) taken from a predictive panel of known performance at the RNA level.
The current study will address several obstacles that have precluded the development of a rapid and reliable biomarker panel ready for clinical testing. While TMAs contain a wealth of potential data, the ability to properly identify and quantify the cell-specific staining patterns of antibodies currently relies on manual identification or pattern recognition programs that are both time consuming and subject to bias and error. Therefore we will utilize an automated digitizing scanning system developed by Vala Sciences Inc. (http://www.valasciences.com/). This system can rapidly record histological sections labeled with up to 10 distinct fluorophores with pixel level subcellular resolution including for TMAs and display each color separately. The system has been acquired by Beckman Coulter Instruments Inc. (Fullerton, Calif.) (http://www.beckmancoulter.com/hr/pressroom/oc_pressReleases_detail.asp?Key=4764&Date1=Dec. 11, 2003) and developed as the Beckman-Coulter IC 100 system. Our application requires only two colors. The reference antibody will be applied to locate all epithelial cells or the subset of epithelial tumor cells or stroma cells and a test antibody will be applied in with a second fluorophore and the pixels of colocalization of test antibody with bona fide epithelia or tumor or stroma will be determined as well as the pixels of not colocalized with target cells. The intensity of antibody labeling at target sites will then be integrated, normalized and compared to nonlocalized binding or to the known clinical outcome. Thus specificity, sensitivity, and accuracy may be determined by existing technology and software. As a gold standard, Phase I will establish the utility of the reference antibodies in comparison to the visual results of a panel of pathologists.
While the importance of the tumor microenvironment on tumor progression and metastasis has been well documented [19, 40, 49, 51-54], very few studies such as Tuxhorn et al. (2002) [51] and [55] have identified genetic markers of reactive stroma. We have utilized linear regression to define expression profiles of the four major cell types contained within prostate tissue samples including tumor cells, stromal cells, and two additional normal epithelial components [18]. In the linear model, the observed expression of any gene (the expression array result for that gene) in a complex piece of dissected prostate tissue used for RNA preparation and Affymetrix analysis is considered to be due to the sum of contributions from the principal cell types in the sample. Each contribution is in turn due to the proportion or percent of each cell type in the sample and the characteristic expression coefficient for the particular gene in a particular cell type:
G
i=β′tumor,iPtumor+β′stroma,iPstroma+β′BPH,iPBPH+β′dilcys gland,iPdilcys gland. (egn. 1)
where Gi is the observed Affymetrix total Gene expression, β′ are the cell-type specific expression coefficients, and the P's are the percent of each cell type of the sample used for the array. The percentages, P, may be determined by examination of H and E slides of the tissue used for RNA preparation by a team of four experienced pathologists. The expression coefficients are determined by multiple linear regression (MLR) analysis. For grossly microdissected tissue enriched in tumor, there are four major cell types as expressed in eqn. 1. We showed that there is very high and statistically significant agreement both between and amongst the four pathologists for the determination of cell-type percentages [18]. In this initial study we sought to determine genes that were consistently expressed predominately by one cell type or another without regard to outcome, i.e. genes that were characteristic of cell type in prostate cancer specimens. We observed 3384 genes were statistically significantly expressed predominately by one cell type. For example, 1096 were consistently expressed by tumor epithelial cells while 496 genes were significantly associated with BPH epithelial cells. Cell type specific expression has been validated by comparison to the literature, by quantitative PCR of LCM samples, and by immunohistochemistry [18].
C.1.A. Diagnostic multigene signature. These initial studies indicate that numerous, perhaps hundreds, of genes may be differentially expressed in the microenvironment of tumor cells which may be useful in diagnosis in supplement to or even in the absence of data from the tumor cell component [18]. Three methods have employed to identify such genes. We adopted the model that it is mainly tumor-adjacent stroma that exhibits the most and largest differential expression changes between the microenvironment around tumor cells and normal or remote stroma. We also assumed that stroma remote from tumor sites of PCa-bearing prostate glands could be used to approximate the expression of normal stroma. We utilized publicly available expression data from 91 cases applied to 148 U133A Affymetrix GeneChips (GEO accession number GSE8218). These cases were the same as those previously studied on the U95av platform [18] plus additional cases. The percent cell composition determined exactly as described [18]. The goal is to find the genes that have altered expression levels between normal stroma cells and the stroma cells close to the tumor cells. We divided U133A samples into two subgroups: 91 tumor-bearing cases and 57 non-tumor-bearing portions of tissue from the same cases. These portions are largely remote stroma. We then applied eqn. 1 to each set thereby determining two β values for stroma: tumor-adjacent stroma and tumor-remote stroma. Note that neither recurrence status or any other clinical parameter such as the Gleason score indicating differences among the tumor bearing portions was considered. Thus only β characteristic of stroma were determined together with a least-squares estimate of error for each β value. Note also that β which are large relative to error must be uniformly or characteristic of tumor-adjacent stroma or remote stroma, i.e. independent of clinical values such as Gleason scores that might indicate differences in aggressiveness. Such β favor high T values in significance tests. The significant differences between the β values for tumor-adjacent stroma and remote stroma were determined. This method produced 208 genes. These significant genes are candidate genes as specifically differentially expressed in the tumor-adjacent microenvironment.
In a second method eqn 1 was extended to include a cross-product:
The cross-product term is used for modeling the interaction between tumor and stroma cells. The significant interaction can be treated as the altered expression trait of stroma caused by the adjacent tumor cells. Egn 2 was applied to the U133A plus data set thereby 1820 significant cross-product terms (˜8% of the probe sets). Finally a third gene list was determined by application of Egn. 2 to and independent set of 91 cases measured on the pangenomic Affymetrix U133A plus2 GeneChips (unpublished data, D. Mercola). This third data set could be used as a test set for the genes determined using the U133A arrays however the differences in platform means that testing can not be applied without cross platform normalization, a process that introduces additional error. Therefore we applied eqn. 2 to the third data set ab initio and sought genes that met the same significance criterion yielding 4533 significant cross-product terms (also ˜8% of probe sets).
Finally we asked which of these genes were common with to all three determinations (the maximum intersect is 208 genes). This three-way intersect yielded 90 genes, i.e. 90 genes which appeared on all three calculations using the two different case sets. These genes may be used to diagnosis the presence of tumor-adjacent gene changes entirely from stroma tissue in the absence of tumor cells.
To test the consistency of these genes PAM (Prediction Analysis for Microarrays) was employed using all 90 genes as a classifier to distinguish tumor and nontumor tissues of the U133A and the U133 plus2 data sets. This method does not utilize information of percent cell type composition.
First, we extracted relevant expression values for these 90 genes from U133plus2 data as a training set. Then we used PAM to analyze these extracted expression data, with tumor/non-tumor as relevant classification variable. Via cross validation, PAM identified 21 genes out of 90 as the best predictor for classification variable. The classifier was tested on the U133A data which yielded a specificity of 100% and a sensitivity of 94.4% (accuracy >94.4%).
Conclusions.
The observations indicate that it is possible to diagnosis the presence of prostate cancer in a large proportion of cases solely from an analysis of the expression of tumor-adjacent tissue, i.e. in the absence of tumor cells. This has a very important potential application to the understanding of patient biopsy material. Moreover, by repeating the above analysis by applying egns. 1 and 2 only to U133A, (two list input in forming the intersect) the final analysis would be free of any input from the test set and stringently objective. We plan to the 21 gene set in this way and to use the resulting list as the starting point for the identification of antibodies suitable for formation of a diagnosis panel for Phase II.
C.1.B. Prognostic Multigene Signature.
MLR may be extended to identify genes differentially expressed by a given cell type between indolent and aggressive tumor cases where “aggression” is defined by chemical recurrence. In the simplest application of this method, eqn. 1 is applied separately to each class of cases—indolent or aggressive cases—and significant differences in β for these two classes of cases for each cell type are determined. Using these methods for a series of 91 patients examined on 131 U133A GeneChips, we observed 1212 genes were significantly and differentially expressed by tumor cells (p<0.05).
In order to validate these differential expression changes, the process was then repeated using the independent 86 cases assessed on the U133A plus2 platform. Again, no cross platform normalization is required. 1373 significantly differentially expressed (p<0.05) genes were identified. “Validated” genes were then defined by four criteria: (i) two or more probe sets of each platform mapped to the same gene; (ii) where multiple probe sets for the same gene were present, all probe sets for the same gene met criteria (iii) and (iv); (iii) differential expression changes for each case set were significant with p<0.05, (iv) the differential expression of identified genes are in the same direction for each case set. We observed that 18 tumor cell specific genes and 19 stroma cell specific gene met these criteria. The chances that that 37 genes could appear to meet the significance criteria for both case sets and be of the same sign by chance is a vanishingly small p<zx indicating supporting that the validated gene list is specific. Moreover, the magnitude of differential express of these genes for the two cases sets is positively and significantly correlated (
Conclusions.
These preliminary calculations indicate that it is readily possible to identify multigene signatures that exhibit reproducible differential expression changes that discriminate indolent for aggressive disease. These calculations account for the cell type heterogeneity that is an essential part of the structure of prostate cancer and leads to the heterogeneity of sample collections assessed by others. Therefore our approach may overcome a major problem plaguing the development of a reliable prognostic classifier. In addition we employed two independent data sets. As a result of accounting for percent cell type composition, we have observed separate gene signatures for tumor epithelial cells and for tumor-adjacent stroma cells. Thus, it may be possible to utilize tissue with sparse tumor content to enhance the prognostic value of the specimens. We plan to use the 38 identified genes as a starting point for the identification and screen of antibodies for our antibody panel in Phase II. This study with TMAs will further validate the prognostic properties of our signature. Numerous additional studies are in progress. We need to test our classifier on published independent data sets by calculation of operating characteristics. We plan to use PAM to further refine our gene list and assess the accuracy by as for the diagnostic profile. These and other refinements are in progress.
C.2. Fully Automated Fluorescence and Absorption Microscopy Analyses.
The scanning microscopy and separate image representation from multiple color labeled slides to be used here has been developed by Vala Sciences Inc. of San Diego by J. Price, President and CEO, and coworkers and has been utilized for a variety of publications (61-84). This system, known as the Q3DM Eidaq™ 100 robotic microscopy instrument runs on the Beckman Coulter's CytoShop™ version 2.0. This instrument includes a Nikon (Melville, N.Y.) Eclipse microscope with an automated stage interfaced to a fluorescence light source and filter wheel of up to 10 narrow band base optical filters in the range 413 nm-663 nm. Numerous supporting software packages has been developed. The system is supported by a variety of antibody-based kits prepared by Vala. Each product contains staining reagents that are targeted towards particular proteins of interest along with a software program (Thora™) that can be used on virtually any computer system. The original instrumentation was developed by a predecessor company, Q3DM Inc. by J. Price focused on the development of high throughput microscopy instrumentation oriented primarily toward automated fluorescence image cytometry (61-84). This instrumentation was designed with accurate image segmentation (81, 83, 84), fluorescent excitation arc lamp stabilization (68, 82), and autofocus for producing fluorescence imaging (69). This system was sold to Beckman Coulter and developed as the Beckman-Coulter IC 100. The current instrumentation is a further generation scanning microcytomer and includes a slide holder hotel for automated scanning of 100 prepared slides.
Two modes, immunofluorescence (IF) with fluorophore-labeled antibodies and immunohistochemistry using absorption chromophores will be employed in the present study. For both methods spectral separation of multiple labeled sections is achieved by capturing multiple images using multiple fixed band pass filters. Up to ten fixed band pass filters are automatically rotated into the optical path of the light either in front of the light source or in front of the camera. Therefore up to 10 images per section are recorded on a monochrome CCD camera creating a “spectral stack”. Spectral unmixing from the data of the spectral stack is sensitive to errors in registration of images of the spectral stack to chromatic aberration. Multiple precautions have been included in the software correct for effects.
For IF the narrow emission of fluorophores of different colors are resolved directly by the appropriate filter of the spectral stack and the corresponding image may be used for pixel-level analysis (for examples see Progozhina et al 2007).
For IHC the broad absorption bands of typical chromophores such as DAB (bisdiazobenzidene), hematoxyln, and others require analysis of multiple images of the spectral stack as previously developed (3). Briefly, spectral unmixing of the observed intensity is based on a model expressed in matrix notation as a linear combination of chromophores where each chromophore contribution is the product of amount of binding and fluorescence intensity or absorption in a given wavelength range. Emission and absorption spectra for all chromophores to be used here are known and the desired unknown are relative amounts of each chromaphore contributing to a given pixel intensity. These are determined by the method of Non-negative Matrix Factorization (NMF) (Rabinovitch et al. unpublished). Effective multicolor separation of tissue images usually requires knowledge of the individual chromophores interacting with the tissue. Based on NMF, the Vala system is the first system capable of performing this color decomposition in a fully automated manner without reference to individual chromaphore-tissue absorption or fluorescence spectra. Instrumentation and software implementing these methods have been developed, characterized and validated on TMAs using objective standards and expert visual scoring and the results are described in reference (Rabinovitch et al. unpublished, Rabinovich et al. 2006).
Supportive additional features of imaging technology and software include: (i) the ability to regroup broken core images which are common in TMA fabrication. None of the currently available software other than that of Vala has addressed this to our knowledge. This problem solved this problem by using the K-means clustering algorithm (53, 54), which provides an automatic method for grouping objects (e.g., pixels) based on distance. Details can be found in the Vala TMA software “framework” article (Rabinovich et al. 2006). (ii) Online viewing, computerized entry of TMA Scoring and Storage is implemented. The tissue microarray core images are organized by software for viewing, interactive entry of expression scores and storing of the data in an organized format. The user can click on any of these thumbnails to view an enlarged image of the entire core and/or a full magnification subfield of the image of the core. Data can then be entered by selecting the data entry pop-up window. The storage format for the images is standard TIF or BMP. Further details can be found in reference (Rabinovich et al. 2006). (iii) Fully Automated Densitometry IF- or IHC-labeled TMAs using Unsupervised Multispectral Unmixing has been developed and implemented (Rabinovich et al. 2006).
We propose to utilize reference antibodies in one color to identify particular cell types and double label the same section in a second color to localize a candidate or test antibody binding. The amount of test antibody binding to target cells such as tumor cells will be determined by colocalization: determination of the pixels of test antibody binding at the site (pixels) of reference antibody labeling. The integrated pixel values of non-colocalized test antibody also will be determined as a measure of lack of specificity.
Two separate uses of colocalization are planned. For routine high throughput screening of candidate antibodies (Phase II), IF will be used as IF has is more sensitive, enjoys greater dynamic range and more amenable to the application of multiple proven antibodies to patient material. For characterization of reference antibodies (Phase I) by comparison to the gold standard of visual score by an expert panel of pathologist, IHC will used in order to provide slides that can be directly assessed by pathologists and compared to the results of colocalization by spectral deconvolution.
C.3. Accuracy of Spectral Unmixing of IHC Labeled TMAs: Comparison to Single Labeling and to Visual Scoring.
Cell type specific labeling of candidate biomarkers in an automated fashion proposed here relies on colocalization of candidate antibodies with the cell of interest as identified by a reference antibody using a second color. The resolution of separate fluorophore labeling patterns from multiple labeled tissue section may be obtained directly from images of multiple narrow band base filters. However absorption/transmission based images of IHC are more challenging and require spectral separation using nonmatrix factorization (NMF). We have evaluated this approach by using double labeled TMAs by the following procedure. Using a set of 97 cores, we first applied the DAB stain and captured 437 multispectral image stacks 9), an average of 4.5 fields of view per core. We then added the hematoxylin stain and acquired a second image stack. The second stack served as the input to our algorithm and the resulting decomposition, which estimated the DAB staining, was compared with the first stack, which serves as the ground truth. We then experimentally evaluated the use of NMF for the color decomposition problem. While reconstruction error represents a quantitative measure, it does not provide a standard for judging how accurately the estimated components represent the dye concentrations. We quantified the performance by comparing the ground truth single-stained image to the corresponding automatically extracted component of the doubly-stained tissue sample as proposed by Rabinovich et al. (Rabinovitch et al. unpublished).
Using this procedure the average decomposition error over all samples was 6.73% with standard deviation of 1.81%. This therefore provides one objective assessment of the accuracy of spectral devolution in comparison to the single chromophore labeled section.
With the accuracy of densitometry via multispectral unmixing established, we asked how this quantitative measurement compares with the subjective scoring of a human expert. A panel of four trained pathologists (M. Krajewska, S. Krajewski, D. Mercola, A. Shabaik) evaluated the 97 tissue biopsies for the expression of antibody protein (DAB). The scoring was performed according to pathology conventions and each tissue section was graded on a scale from 0.0 to 3.0 in increments of 0.5. For correlation of the visual and analytical results, we analyzed the performance of a linear model y=mx+c, where x is the score reported by NMF decomposition, y is the pathologist's score, m is the slope and c is the y-intercept. Linear regression was used to fit the model. The fitting error for regression may be an indication of the prediction error of the model. However, depending on the complexity of the model and the amount of data available, the regression error can be significantly different from the true prediction error of the model. Thus, an effort was made to estimate the prediction error and report it instead of the fitting error. The simplest and most widely used method for reporting prediction error when the data is scarce is cross-validation (86). Ten-fold cross validation resulted in a mean squared error of 0.02 with a standard deviation of 0.01. This is equivalent to a root mean squared (RMS) error of 0.163, which also translates to an average of 5.4% error on the pathologist's scale. A major result of the validation study is that the 5.4% error is considerably larger than the corresponding signal: noise ratio of the camera detector. Thus the validation makes available a greatly increased dynamic range of electronic signal detection of the camera-based microscope over the visual system with a “noise” value of ˜3×5.4%=16.2% vs. <1% for the camera. The increased dynamic range for quantified antibody binding overcome a major limitation of antibody labeling using visual or IHC methods and greatly increases the ability to identify antibodies that correlate with survival data and other important clinical co variants. This advantage is extended many times for fluorescence-based antibody labeling.
Another decomposition of the form A=BC that is widely used is Independent Component Analysis (ICA) (Hyvarinen, J., Karhunen, and E. Oja, Independent Component. Analysis, John Wiley & Sons, 2001). ICA is based on the assumption that the matrix A is the result of the superposition of a number of stochastically independent processes. This is a more reasonable description of the staining process where each stain can be assumed to be independent of the other stain. Classically, however, ICA algorithms do not enforce non-negativity and that makes them unsuited for stain recovery as well. We experimentally evaluated the use of NMF and ICA for the color decomposition problem. While reconstruction error represents a simple quantitative measure, it does not provide a standard for judging how accurately the estimated components represent the dye concentrations. We quantify the performance by comparing the ground truth single-stained DAB image to the corresponding automatically extracted component of the doubly-stained DAB/hematoxyln tissue sample. Quantitatively, the overall for four images sets was 50% larger for ICA compared to NMF (the images are available at hppt://vision.ucsd.edu/). Both NMF and ICA provide good results however there is an observable increase in fidelity to ground truth for the NMF analysis. We propose to utilize NMF for the studies proposed here.
Conclusions. 1. These Studies Provide Support for the Ability to Successfully Decompose Multicolor Labeled TMAs to Component Images.
The application proposed here is simpler as separate 2D images are unnecessary. We plan to extract a subset of pixel intensities, those of chromaphore A that are co-localized with the pixels of chromaphore B where chromaphore A predominately binds to cells of interest such as tumor or epithelial cells or stroma cells. We have not completed this task however only minor modifications to existing software, pixel integration, is required and is proposed as a milestone of Phase I. The data of co-localized chromaphore B, the test chromaphore, would then be analyzed by Cox-regression and ANOVA analysis with covariates of disease progression currently available for the cases of the PCa TMA. 2, The automated ability to scan TMAs and extract quantified data will greatly facilitate antibody screening.
C. 4. Multicolor IF Separation at the Subcellualar Level.
The design goal of the Vala scanning robotic microscope is subcellular segmentation using pixel level resolution. It is important to note, therefore, that this capability exceeds the needs of cellular resolution required here which is well within current level of the instrumentation development. This was insured by the successful development of an automated membrane algorithm of the Thora package (Prigozhina 2007). For example mouse skin tumors were labeled with three fluorophores, two to identify proteins of interest, the membrane binding E-cadherin and the epithelial localizing antibody anti-K-14, and a cell localizing label for nuclei, DAPI. In this context, K14 is a putative marker for tumorigenic epidermal cells that invade the deeper skin layers. Cells exhibiting K14 signal (high red channel fluorescence) were clustered within the tumor loci. Areas of the section that stained brightly for K14 stained relatively dimly for cadherins, whereas surrounding tissue stained poorly for K14 and brightly for cadherins. To quantify K14 and cadherins, Thora separated the three primary cellular compartments (membrane, nucleus, and cytosol) from the dualcolor image of pan-cadherin and nuclear fluorescence. Thora estimated the cell boundaries in both the normal cells bordering the tumor where the cadherin signal was strong and in the tumor where it was relatively weak. To measure cadherin reduction in K14-positive cells, TMIs (total membrane intensity by pixel integration by boundary recognition) in the cadherin channel were collated for K14 cells with ACT (average cytoplasmic intensity) of 30 (the ACT range was 0 ACT 255 for the 8-bit images). By visual inspection and comparison of the intensity measurements of different cellular regions, ACT values below 30 arose from background staining that was not cell-specific. The mean pan-cadherin TMI for K14-positive cells was just 34% of that for K14-negative cells, and this difference was highly significant (P<0.01). Thus, the K14-positive cells representing invading tumor exhibited quantifiably reduced cadherin expression relative to the surrounding cells. Other examples and details of the development have been described in detail (Prizozina 2007).
For the applications proposed in this SBIR project membrane boundary recognition is less crucial as it is only necessary to identify zones of tumor epithelial cells and zones of nonepithelial stroma and those subareas of test antibody labeling that colocalize with either tumor or, for nonspecific labeling nontumor labeling. It is of course important to recognize that colocalized tumor labeling may only be increased on average compared to non tumor labeling and, like cadherin, this may be readily quantified.
The Prostate cancer TMAs to be used here have been fabricated as part of the NIH-supported UCI SPECS (Strategic Partners for the Evaluation of Cancer Signatures) consortium at the Burnham Institute of Medical Research, a consortium member of the UCI SPECS program and are available here as an NIH resource of NIH-sponsored projects. The TMAs have been specifically fabricated to validate the cell-specificity of candidate biomarkers of prostate cancer. 272 cases with known clinical outcome have been included to date. FFPE blocks and clinical follow-up were retrieved from two participating institutes of the SPECS consortium according to an IRB-approved and HIPPA-compliant protocol and consist of cases provided by SKCC (60 cancer cases, 12 normal cases) with the rest of the cases drawn from UCI that have 10-19 years of clinical follow-up with clinical characteristics as previously described in T. Ahlering and coworkers [75]. All cases have been re-examined by two clinical pathologists who confirmed the Gleason score and defined areas of tumor, BPH, stroma adjacent to tumor, stroma away from tumor, and epithelium of dilated cystic glands and PIN cores. In order to validate cell-specific binding properties of candidate biomarker antibodies, each case on the TMAs is represented by 4-5 cores from 4-5 zones of pure cell types as defined by two pathologists. Duplicate cores from the chosen zones were used for array fabrication so that all zones are represented in duplicate. Thus these TMAs are unusual in that they have 4−5×2 cores per case on the array. The TMAs are under continuous construction with the next phase to include 100 additional UCI cases so that the arrays available for the proposed study will exceed the present 272 case set. The prototype array at the 66 case stage have been utilized for the evaluation of several potential antibody by markers including Claudin I and Bcl-B (Krajewska et al. 2007; Krajewska et al. 2008).
The studies of Krajewska et al. (Krajewska 2007;Krajewska 2008) utilized double antibody labeling of the same TMA section using anti-Claudin I and anti-cytokeratin in the double chromagen mode. For colocalization the two color were separated using a segmentation program developed by Aperio Technologies and represented individually and provide clear indication of the epithelial binding pattern of anti-Claudin-I. Pixel count and quantification of colocalization as well as nonlocalized binding is readily possible although non specific binding for anti-Claudin-I is negligible in this example. The method is less easily generalized to three or more colors or to IF as yet and therefore is less versatile than the Thora system of Vala preferred for this application however it provides further illustration of our early experience in the methods proposed here.
Conclusions.
Candidate gene expression levels for diagnosis and prognosis have been derived. Methods for the high throughput and quantitative assessment of labeling by corresponding antibodies are available. The wedding of this methods promises to provide the means of developing reference and assessment antibodies for new ICON-compliant clinical assays which solve significant unmet needs.
Phase I.
Here we focus on attaining milestones that support the goal of demonstrating that reference antibodies and methods are available for the reliable and quantitative identification of cells of interest for use in Phase II, the systematic assessment of candidate biomarker antibodies for the development of panels for the multiplex determination of diagnosis and prognosis
Milestone 1.
Develop an automated optimized imaging assay and SOP for prostate stroma and epithelial/tumor cells using three or more antibodies for immunohistochemistry and immunofluorescence.
Unstained sections of formalin-fixed paraffin-embedded prostate tumors, unstained sections of our prostate cancer TMAs and frozen sections of frozen prostate carcinoma-bearing tissues will be utilized. FFPE blocks will be taken from the extensive collection used for construction of the TMAs. Frozen tissues are available from the UCI SPECS program. Antibodies for the labeling of all epithelial structures, just tumor epithelium, and the fibroblast/myofibroblasts component of stroma will be optimized separately for all three tissue preparations. Screening studies will be carried out using chromagen labeling by indirect IHC using DAB for ease of visual monitoring and optimization will be extended to indirect IF.
Panepithelial labeling.
Panepithelial labeling will be used as a reference to define candidate antibody biomarker labeling that colocalizes with bona fide epithelium in prostate cancer sections and therefore to derive a ratio of epithelial:nonepithelial labeling as a measure of specificity. Panepithelial labeling will be optimized for two antibodies and the best one of these used for all subsequent studies. Anti-high molecular cytokeratin (anti-HMW keratin; Dako clone 34βE12 mouse monoclonal anticytokeratin) will be used at the starting conditions that we have previously employed for the prostate cancer TMAs (Krajewski 2007). The antibody labels squamous, ductal and complex epithelia containing cytokeratins 1, 5, 10, and 14 (68, 58, 56.5′ and 50 kDa proteins).
A second anti-panepithelial antibody is AE3/AE4 (Dako AE3/AE4 MNF116 mouse monoclonal antihuman) which is in standard clinical use in the Pathology Department at UCI for the identification of epithelial components especially in the investigation of metastatic spread of carcinomas in distant tissues. The antibody labels multiple cytokeratins (65-67, 64, 59, 58, 56.5, 56, 54, 52, 50, 48 and 40 kDa cytokeratins) in either FFPE or frozen tissue.
Tumor Epithelial Cell Labeling.
Tumor epithelial cell labeling will be used as a reference to define the colocalization of labeling by candidate antibody biomarkers with bona fide tumor cells and therefore to derive the ratio tumor cell labling:non tumor cell labeling as a measure of specificity. Prostate cancer tumor epithelial cell labeling provides a more specific reference site for co-localization studies to be carried out in Phase II but is a challenging reference target owing to the limited number of antigens accepted as expressed in prostate cancer epithelial cells independent of the degree of differentiation or other histological properties such as Gleason score. We previously examined the expression pattern at the RNA level for a series of 55 tumors where expression could be resolved to the principal cells types (tumor epithelial cells, BPH epithelial cells, dilated cystic gland lining epithelium and stroma) which revealed that several classically expressed antigens such as PSMA (prostate specific membrane antigen), PAP (prostate acid phosphatase), and AMACR (α-methyl acyl CoA racemase) where significantly expressed at the RNA in nearly all tumor cells independent of grade and stage (Stuart et al. 2004). In this study we validated the protein expression was specific in seven representative cases (Stuart et al. 2004) using IHC.
Anti-AMACR is now in widespread clinical use for the identification of metastatic prostate cancer and has been reviewed extensively (e.g. Rubin 2004). In an analysis of anti-AMACR labeling of a prostate cancer TMA of 70 cases including “foamy” cell carcinoma with low expression of AMACR, labeling was detected in 91% percent of cases (Rubin 2004). Specificity and sensitivity were examined by quantitative receiver operator characteristic which yields an AUC was 0.9 (p<0.00001). These values are highly encouraging for the approach proposed here. It is not necessary to identify all prostate cancer cells but rather label a statistically valid sampling in order to assess, on this sample, the colocalization properties of candidate antibody biomarkers. Thus, a 91% labeling efficiency is very acceptable. We will employ the same commercial antibody and procedures as for Rubin et al. (Rubin 2004): mouse monoclonal anti-AMACR p504s (Zeta Corp., Sierra Madre, Calif.) at a starting dilution for optimization (see below) of 1:25. The optimization protocol to be used here encompasses the conditions of Rubin et al. (Rubin 2004). A major potential advantage of anti-AMACR is that the weak or absent labeling of normal epithelial components will facilitate quantification of nonspecific labeling (“noncolocalized labeling”) by candidate biomarker antibodies to be developed in Phase II.
Other potential tumor epithelial cell antibodies include anti-PSMA, anti-PSA, and anti-PAP. Antibodies to these products react with epithelium of normal and malignant cells. Anti-PSMA is extensively studied, is FDA approved (clone 7E11) for radiological detection of PCa metastases, labels nearly 100% of tumors in histological sections, and consistently label tumors at greater intensity that benign prostate epithelium (Chang 2004). We will optimize the labeling of FFPE, TMAs, and frozen sections test with our quantitative IF methods can exploit this property to distinguish tumor from benign labeling in comparison to anti-AMACR and visual scoring. We will utilize a mouse monoclonal anti-human PSMA (Dako clone 3E6).
Stroma Cell Labeling.
“Stroma” as used here is a collective term consistent largely of fibroblasts, myofibroblasts and less proportion of vascular, neural, and other elements. Fibroblast and myofibroblasts labeling will be used as a reference to identify colocalization of stroma-binding candidate biomarker antibodies and to derive the ration of stroma:nonstroma labeling by the candidate antibodies. Widely accepted markers that may make suitable reference antibodies consist of anti-desmin, anti-vimentin, and smooth type α-actin and others (Castellucci 1996; Tuxhorn 2002; Ayala 2003; Tomas 2004: Ao 2006; Jiang 2007). We have previously utilized anti-desmin for the IHC analysis of prostate cancer (Stuart 2004). Considerable literature has accumulated indicating that Vimentin and smooth muscle type α-alpha vary in expression in PCa depending on the extent of epithelial-mesenchymal transformation and reactive stroma formation, two processes that correlate with aggression (Tuxhorn 2002; Ayala 2003; Hyanagisawa 2007; Yang 2008)). These phenomena appear to be proximal to the site of PCa. These markers therefore have the potential to delimit the “field” effects that are associated with differential gene expression of tumor-adjacent stroma. These observation correlate well with our observations that tumor-adjacent stroma contain numerous differentially expressed genes useful for diagnosis and for prognosis. Indeed, as noted, the mRNA levels of desmin and vimentin are significantly increased in stroma of our PCa samples compared to the epithelial components (Stuart et al. 2004). We plane, therefore, to optimize all three antibodies and determine their suitability as reference antibodies for stroma in general and tumor-adjacent stroma in particular. Previously characterized stroma reference antibodies include: anti-desmin mouse monoclonal antibody Dako clone D33 (Stuart 2004); anti-vimentin goat polyclonal sera cat. No. AB1620 from Chemicon (Temecula, Calif.) (Tuxhorn 2002); and anti-smooth muscle α-actin Dako clone IA4 (Tuxhorn 2002). For the development of stable renewable reagent sources it is highly desirable to work with monoclonal antibodies where source licensing can be organized. Therefore for anti-vimentin we will also examin mouse monoclona antibody from Dako, clone V9.
Optimization and SOP Development.
The primary antibodies will be applied using an automated immunostainer (DAKO Universal Staining System) and employing the Envision-Plus-horseradish peroxidase system (DakoCytomation, Inc.) secondary labeling system for DAB. FFPE sections will be deparaffinized by xylene overnight followed by microwave treatment and 0.4 power for 30 min. in a 6.0-pH citrate buffer. No enzymes or other “antigen retrieval” processes will be applied here or any of the labeling conditions considered here in order to minimize the variables required in developing panels of multiple antibodies with compatible protocols (Phase II). Sections will be pre-treated with normal mouse serum for 40 min. and washed in PBS with automated stirring three times. For optimization, primary antibodies will be applied at room temperature for 40 min in two-fold serial dilution from 1:30 through 1:960 or higher dilutions if practical. The optimal titre (as well as the preceding and following titre value) as judged by visual appearance (D. Mercola, F.C.A.P.) of specific labeling intensity to background labeling intensity will be re-tested on sections with increased deparaffinization steps (see IF procedure) including an over night baking step and reduced as well as extended microwaving to check for an improvement in signal to background labeling intensity. Finally, the time and temperature of application of the primary antibody will be optimized by comparing exposure to primary antibodies for 2 h and 24 h at room temperature and 24 at 4 deg. C.
These steps will be applied to both FFPE and frozen sections of fresh tissue. In the case of fresh tissue, we will utilize samples that have been cryopreserved in liquid nitrogen from the time of initial freezing. All samples for the UCI SPECS project are obtained directly from the O.R. and processed by an expedited surgical pathology grossing procedure. Sample for research are taken from tissue adjacent to the grossly identified tumor site or, for “remote” tissue control samples, taken from the contralateral prostate. Tracking sheets are maintained on all samples giving the elapsed time from the O.R. to freezing. Representative samples are used for RNA q.c. as an indication of preservation by analysis of total RNA using an Agilent Bioanalyzer which indicates high levels of preservation in over 95% of samples. Frozen sections will be prepared from these tissues directly from the frozen state without thawing. The sections will be fixed for 60 sec. in 95% methanol or 100% acetone or 70% EtOH all at −22 deg. C., air-dried, and used directly for antibody optimization.
TMA Confirmation.
Optimized labeling protocols developed on FFPE sections will be tested by application to our TMA with 272 cases including cores of tumor-adjacent and remote stroma. Labeling of the TMAs will provide information of the generality of labeling across cases and the reproducibility of specific labeling for tumor and stroma. To insure that optimization has been achieved for the TMAs, the last steps of the optimization procedure will be repeated using the TMA sections, i.e. the application of primary antibody using the three best titre values and the following steps. Progress will be monitored by visual inspection of the DAB labeled slides (D. Mercola, F.C.A.P). Optimal conditions will be judged by the most cases of the TMA that reflect the desired criteria of the greatest differential expression between target cell type with “background” intensity. All informative slides will be stored in a temperature controlled laboratory for scanning and quantitative assessment of variability, accuracy, and reproducibility assessment of Milestones 3 and 4.
Immunofluorescence.
Immunofluorescence is the intended method of choice owing to the much higher dynamic range and sensitivity of antigen detection. Indeed, we anticipate that primary antibodies can be extended to high titres by factors of 10× or more. The major challenge is selection of conditions that minimize “background” or “autofluorescence”. Background fluorescence can be minimize by using fluorophores with long wavelength emission (>500 nm), use of sections with rigorous deparaffinization procedures (i.e. the overnight deparaffinzation xylene treatment and used of prolong baking of unstained FFPE sections, above), use of pretested acid washed slides and coverslipping reagents, and use of a configuration of the robotic microscope with optical filter wheel located before the monochrome CCD camera. These methods have been optimized previously (Rabinovich 2006). The characterized fluorophore-conjugated secondary antibodies to be used previously that will be applied here are: Texas Red-labeled goat anti-mouse (catalog number 115-075-146, Jackson Laboratories, Bar Harbor, Me.) and Alexa Fluor 488-labeled goat anti-mouse (catalog number A21121, Molecular Probes, Eugene, Oreg.). These reagents can be used at dilutions in the range 1:1,000 to 1:10,000. The optimum concentration will be determined for sections of our TMAs.
Visual assessment of optimum conditions require counter staining. Sections will be stained with DAPI (Molecular Probes, Eugene, Oreg.) at 75 ng/ml (in 10 mM TRIS, 10 mM EDTA, 100 mM NaCl) for 45 min prior to sealing with coverslips. Visual assessment will be carried out by J. Price and D. Mercola.
Milestone 2.
Storage and visualization will utilize exiting technology of the Vala Sciences Inc. system. All data will also be placed in a free database that is DICOM compliant.
In this project the bulk of data collection, storage, and analysis will be by the Vala Science robotic scanning microscope and associated software and storage capacity. As reviewed here (Preliminary Studies), Throra and associated software for data acquisition, analysis and storage are advanced. These are most completely described in the specialty publications of Rabinovich et al. (Rabinovich 2006) and Prignoshima et al. (Prigoshina 2007). Moreover Proveri Inc. and Vala Sciences Inc. are committed to the development of completely DICOM complaint storage and data sharing (http://www.sph.sc.edu/comd/rorden/dicom.html). The primary data of the assay proposed here, a multiplexed antibody assay utilizing indirect IF, will consist of a spectral stack of multiple color images of histological section of biopsies or postprostatectomy tissue sections together with standard hematoxylin and eosin stained sections of the same section used for IF labeling. Such images represent a novel data set for diagnosis and prognosis without direct precedent in the DICOM standard. Since Phase II is focused on product development for diagnosis and prognosis in the CLIA reference lab setting, Vala Science Inc. is very interested in developing a DICOM-compatible format for the storage and transmission of primary tissue images. It is planned to develop a demonstration format using DICOM heading and other features in analogy of other imaging systems.
SOPs Will be Developed for Specimen Collection, Processing, and Stability of the Cell Types in the Imaging Assay.
SOPs for the acquisition of tissues and blocks have been developed by the UCI SPECS program and are maintained as date pdf files and in an SOP workbook. These SOPs describe procedure for informed-consent based patient recruitment at all participating sides and methods of tissue collection at O.R rooms, expedited processing and storage together with diagrammatic illustrations of dissection procedures and additional tracking forms for each specimen. All procedures are UC11RB-approved and HIPPA-compliant. In addition the UCI SPECS program maintains “shadow charts” for all recruited patients including the signed witness informed consent, tracking sheets, and CRFs of baseline clinical data together with source documentation of all values recorded in the SPECS data base. The data base is maintained on a devoted server hosted by a participating institute, the Sidney Kimmel Cancer Center of San Diego, in a locked server room under the control of the SKCC IT department. The server is accessed remotely via a password protected web-based portal by approved clinical coordinators and the data base manager. All personnel are UCI employees. The SOPs will be incorporated into the SOPs generated for phase I of this project.
SOPs describing the optimized procedures and reagents of Milestone 1 will be developed as final conditions are determined. The methods for the fabrication of the TMAs will be included. These will include methods for periodic testing to insure stability of the labeling results. The current TMAs contain cores of fixed cultured prostate cells including standard tumor cells (LnCAP, PC3, DU145, M12) and normal immortalized cells (RWPE1, p69) will will be used to record quantified labeling intensity. Upon the completion of Milestone 1, multiple section of the TMA block containing cell cores will be prepared as a master lot for periodic qc and for standardizing new lots of renewable reagents. These procedures will be included in the SOPs.
It is a major goal of phase II to initiate a prospective validation program using newly recruited clinical patients and UCI and applying the multiplex panel to research biopsies and post surgery tissue specimens in the CLIA lab of the molecular pathology core of the UCI Department of Pathology and Laboratory Medicine. In anticipation of this study, All SOPs, master lot preparations, and DICOM-compatible image storage will be coordinated with CLIA requirements of this laboratory.
Antibodies against known prostate cancer antigens and against putative prostate cancer biomarkers identified by gene expression analysis will be obtained from commercial sources and characterized using Western blotting and immunohistochemistry. Candidate antibodies that demonstrate the ability to detect discrete proteins on Western Blots prepared from fresh prostate tissue samples (stroma or tumor) and the ability to differentially label cell types in paraffin-embedded prostate cancer tissue sections will identified. Their ability to predict clinical outcome will be tested in specific aim 2.
D.1.a. Description of Antibodies
Commercial antibodies will be purchased, if available. Other antibodies will be generated (Lampire Biologicals, San Diego, Calif.). Numerous antibodies used in our separate projects have been developed in cooperation with Lampire Biologicals [50, 68-74].
1. Antibodies that label prostate tumor cells, normal epithelium, or stromal cells to be used as internal standards will be used to identify specific cell-types within prostate tissue samples. Those on hand of particular importance for the identification of epithelial components include anti-high molecular weight cytokeratin (HMW cytokeratin), anti-PSA, anti-PAP, anti-PSMA, and anti-Amacr. Those intended for the identification of stroma include anti-Desmin and anti-smooth muscle alpha actin (Anti-ACTA). We have optimized all of these for use with FFPE tissue sections and described results in previous studies [18, 67].
2. Antibodies against potential prognostic markers identified by gene expression analysis. Twelve commercially available antibodies against predicted antigens have been obtained and screened using standard sections of FFPE prostate cancer tissue blocks. Five of these antibodies are very promising for detailed characterization as proposed here. Antibodies that are not available or exhibit poor labeling or background properties in screening will be commissioned de novo as described below.
3. The selection and screening of additional antibodies will be prioritized by starting with antibodies to gene products that exhibit the largest differential labeling (largest difference in immunoscore or normalized pixel intensity) between nonrecurrent and recurrent prostate cancer cases. As noted above, approximately half of the antibodies screened so far do exhibit excellent signal to background properties on test sections of FFPE prostate cancer.
D.1.b. Criteria for Inclusion of Antibodies for TMA Analysis Will Include: Path to Monoclonal Antibody Production.
1. Antibodies are suggested by the results of MLR (Preliminary Data, Section C1). Candidate antibodies first will be vetted by Western analysis to test for the detection of antigen of correct molecular weight in prostate tumor tissue extracts or alternative molecular weights previously reported as prostate cancer-variants. Previous experience [18] has revealed that an important factor in meeting these criteria is knowledge of the origin of the antigen. The linear regression results identify probe sets of Affymetrix GeneChips which correspond to precise genes and introns of genes. Commercial antibodies against recombinant proteins or large fragments of proteins likely correspond to the identified gene product and so are useful for testing whether genes of probe sets are expressed at the protein level. Similarly, commercial antibodies against highly pure native proteins of a carefully characterized molecular weight that agrees with that expected value on the basis of the Affymetrix-predicted gene product also may be expected to be confirmed by Western analysis. However, antibodies produced against proteins purified from natural sources may contain alternative spliced products and/or other gene family member proteins as well as closely related proteins or fragments that are difficult to separate during purification may lead to antibodies reactive to a range of molecular weights with an unclear relationship to the gene product corresponding to the Affymetrix probe set. Monoclonal antibodies against recombinant or synthetic peptides more often meet the need for single gene product specificity and will be preferred. In addition monoclonal (mouse, rat) define a potentially renewable resource that may be contracted as a stable supplier of test kit reagents. Therefore, all polyclonal antibodies characterized here for inclusion on the final antibody classifier will replicated by the commissioned preparation of the corresponding monoclonal antibody as part of phase II.
2. Consistent and robust IHC signal of antigens from formalin-fixed and paraffin-embedded (FFPE) tissue. TMAs provide a major advantage in that the fraction of cases exhibiting increased or decreased IHC signal may be quantified readily. In order to develop an assay with maximum reproducibility, methods that minimize reliance on “antigen retrieval” strategies will be adopted. This will select for robust antibodies capable of recognizing antigens on archived samples.
3. Consistent and robust IHC signal of antigens from archived (>10 years) FFPE tissue. IHC labeling intensity for each antibody will be correlated with the age of the sample on the TMA. An advantage of our TMAs is the presence of cases from 2 to 19 years old.
4. Cell-specific labeling. Cell identity (normal epithelium, stroma, BPH) will be determined by manual inspection or staining with cell-specific antibodies. IHC intensity for each antibody will be immunoscored for staining intensity and cell specificity as described below (Sections D.2.c. or D.3.b.)
D.1.b. Tissue Source for Western Blotting.
Tissues will be obtained from the UCI SPECS prostate project tissue bank. This is a resource of the NIH-supported UCI SPECS prostate project. Prostate samples were obtained from patients (UCI) that were preoperatively staged as having organ-confined prostate cancer. Institutional Review Board-approved informed consent for participation in this project was obtained from all patients. Tissue samples were collected in the operating room, and specimens were immediately transported to institutional pathologists who provided fresh portions of grossly identifiable or suspected tumor tissue and separate portions of uninvolved tissues that were excess to patient care needs (surgical pathology staging and confirmatory diagnosis). All excess tissue was snap frozen upon receipt and maintained in liquid nitrogen until used for frozen section preparation at −22° C. Fifty five percent of all cases collected in this series contained histologically confirmed tumor tissue. Portions of frozen samples enriched for tumor, stroma, BPH, and dilated cystic glands are identified by examination of frozen sections. When suitable tissues are identified, thick frozen sections of 20 microns are collected in separate Eppendorf tubes for lysis and Western analysis.
Additionally, the ability of antibodies to visualize antigens of correct MW on Western blots from tissue extracts established from a panel of human prostate cell lines will be determined. This panel will include androgen resistant prostate cancer cells (PC3, DU145), androgen sensitive prostate cancer cells (LnCAP), primary immortalized RWPE-1 epithelial cells. Cancer cells of alternative derivation (lung, breast, colon), and several normal cell lines (fibroblasts, myoblasts) (ATCC) (these cells have also been applied to the TMAs as sections of formalin-fixed cell pellets).
D.1.c. Western blotting
Tissues or cultured cells will be lysed in either 1× Laemmli solution lacking bromophenol blue or in RIPA buffer (0.15 mM NaCl/0.05 mM Tris·HCl, pH 7.2/1% Triton X-100/1% sodium deoxycholate/0.1% sodium dodecyl sulfate) containing protease inhibitors including the caspase inhibitors 100 μM Z-Asp-2,6-dichlorobenzoyloxymethyl-ketone (Bachem) and Z-Val-Ala-Asp-fmk (Calbiochem). Total protein content will be quantified by either the Bradford or bicinchoninic acid methods (Pierce). SDS/PAGE and immunoblotting with enhanced chemiluminescence-based detection (Amersham Pharmacia) will be performed [50, 69-71].
Antibody reactivity will be semiquantified by comparison of reaction intensity of tissue and cellular extracts with extracts of prostate cancer cells (PC3, LNCaP) and negative control cells (bacterial cultures and female normal breast epithelial cells, MCF10A) of known total protein mass.
Our methods for optimization and detection of antibody labeling have been described extensively [50, 68-74]. Briefly, the cell specificity of the identified antibody for normal and malignant prostate tissue will be tested by comparing the binding patterns on a series of normal and malignant prostate tissue specimens. FFPE tissue sections (5 μm) will be deparaffinized, microwave-heated, and immunolabeled by indirect staining using either a conjugated secondary antibody for avidin-biotin complex formation with horseradish peroxidase (HRP) using the Vecta labeling reagents (Vector Laboratories) followed by addition of diaminobenzidine (DAB) for colorimetric detection or the Envision-Plus-HRP system (Dako) with a Dako Universal Staining System. A range of antibody concentrations will be tested to optimize signal detection and specificity. For all tissues examined, the immunostaining procedure will be performed in parallel by using either preimmune serum (polyclonals) to verify specificity, or the antiserum reabsorbed with 5-10 μg/ml of synthetic peptide or recombinant protein immunogen where available. Positive controls for cell-type specificity will be determined by staining sections with a “cocktail” of antibodies directed against pan-cytokeratin (Sigma) to identify epithelial cells and antibodies against Desmin, alpha-smooth muscle actin, or prolyl-4-hydroxylase to identify stromal cells
Our TMAs have been constructed from archived prostate tissue samples with known clinical outcomes from SKCC and UCI. IHC staining will be performed using antibodies developed in Specific Aim 1. IHC staining levels will be immunoscored (below) and compared to clinical outcomes by Kaplan-Meier analysis. Significance of discrimination of survival groups will be determined by the Cox Proportional Hazards model.
Visual determination is carried out by three pathologists (SK, MK, and DAM) and averaged. Candidate antibodies demonstrating the greatest sensitivity, specificity, and accuracy for the prediction of clinical outcome by the Kaplan-Meier criterion will be selected for the antibody panel for prognostic validation of clinical samples in Phase II.
D2.b. Immunohistochemistry on TMAs.
Immunohistochemistry on TMAs will be performed as described previously [50, 69-71] and above (Section D.1.d.)
D.2.c. Immunoscoring of TMA Readouts
Immunoscores are determined visually and are formed as a product of the percent of a given cell type that is positive 1-100 percent) times the intensity on a three point scale yielding a range of values from 1-300 [68-70, 72, 73]. For the three-point scale intensity is j judged as 0, negative; 1+, weak; 2+, moderate; and 3+, strong [70]. Samples will be additionally scored for percentage of immunopositive malignant cells, estimating the percentage in increments of 10% (0%, 10%, 20%, 30%, and so on) from a minimum of five representative medium-power fields. The scoring will then be based on the percentage of immunopositive cells (0 to 100) multiplied by staining intensity score (0/1/2/3), yielding scores of 0 to 300. Scoring is conducted in a joint session of the three pathologists utilizing the original glass slides and a multihead microscrope in order to insure identical viewing times and field exposures. The reproducibility and agreement among pathologists following this format has been assessed [18] and immunoscoring using the above scales has been used in several studies [50, 69-71].
D.1.d. Statistical Analysis
Data will be analyzed using the JMP Statistics software package (SAS Institute, Cary, N.C.), and STATISTICA Software (StatSoft, Tulsa, Okla.). Comparisons of antibody immunostaining data with patient survival will be made using the Cox proportional hazards model and the comparison of Kaplan-Meier survival curves. An unpaired t test method was used for correlation of immunoscores with the available patient data. All statistical methods will be supervised by our biostatistician, Zhenyu Jia, consultant for Phases I and II of this project (see Biosketch, Z. Jia and letter).
Antibody performance will be judged by conventional operating characteristics (accuracy, sensitivity, and specificity) but also by criteria that produce the smallest panels that maximizes the percent of cases of the TMA accurately discriminated as aggressive or nonagressive by survival and other criteria. This is an important consideration, as a true classifier panel should contain biomarkers effective with cases that other biomarkers may be insensitive to, i.e. cover the diversity of prostate cancer. Thus, individual antibodies will be scored by the number of cases unique classified with very large or very small odds ratios that other antibodies fail to distinguish (i.e. the number of unique cases accurately classified). These criteria further insure that the minimum number of antibodies to discriminate all amendable cases of the TMA will be formed.
The discriminatory power and the rate of characterization of the prognostic antibodies identified in Specific Aim 2 may be improved using image analysis that provides for quantitative determination of antibody labeling intensity. Rapid scanning, digitization, and the use of a newly developed algorithm for two-color separation are established at the BIMR largely as the developmental work of one of the applicants (SK). Digitized IHC labeled prostate TMA are maintain on a server located at the BIMR and accessible by all participants via a secure portal (https://scanscope.burnham.org/Login.php). This greatly facilitates the monitoring of IHC results and planning of next steps and immunoscoring sessions. UCI SPECS pathologists utilize high resolution line scanned H and E and IHC images of this site for immunoscoring of other projects and confirmed the histological features of the TMAs such as Gleason scores, presence of PIN, etc. This technology allows for automated quantification of cell-specific antibody staining of TMA samples without reliance on “shape recognition” or manual inspection to determine cell-type. This technology will be tested using the panel of prognostic antibodies developed in the first two specific aims.
D.3.a. Double Labeling.
Double labeling places constraints on the combination of standard (anti-PSMA, anti-AMACR, and anti-cytokeratin) and candidate antibody combinations owing to the need to use secondary antibodies for the development of two different chromagens. The methods that we have previously used for double labeling (Krajewski 2007; Krajewska 2008) will be followed closely. In general candidate antibodies will be derived from rabbit sera. Indirect IHC using biotin labeled anti-rabbit IgG will be applied for development of DAB (3,3¶-diaminobenzidine chromagen, DAKOCytomation; brown). Mouse monoclonal antibodies to AMACR, PSMA, or cytokeratin will be identified by addition of biotin-labeled anti-mouse for development of the black SG precipitate (Serotec; SG chromagen, Vector Lab., Inc.; black). No or very light counter staining with Nuclear Red (DAKOCytomation) will be applied
D.3.b. Validation of Prostate Cancer Predictive Antibodies on Tissue Microarrays (TMAs).
Color unmixing has been validated for sections labeled with hematoxyln and DAB (Preliminary Data). As noted, actual isolation of subsets of pixels that co-localize with epithelial or tumor cells is a milestone of Phase I. Validation will be extended to DAB and SG double labeled sections and to colocalized integrated and normalize pixel values. For this purpose it is important to note that visual scores are traditional obtained as the product of the intensity of labeling (on a 0 to 3+ scale) times the percent of tumor or epithelial cells that exhibit positive labeling. Here both factors will be used to validate co-localization. A test system utilizing a polyclonal anti-AMACR (DAB) and monoclonal anti-cytokeratin (SG) alone and in combination will be applied to both the tumor TMA and to the BPH TMA. First, analogous to the hematoxyln-DAB system, deconvolution results (reconstructed DAB image and reconstructed SG image) for the combination labeling will be compared to individual labeling (ground truth). These tests will define the accuracy as percent error +/−standard deviation for each chromagen. Second, colocalized pixel sums for AMACR labeling as a “standard” for binding to a high percentage of tumor cells will be determined. This is the sum of pixel intensity for DAB at pixels positive for SG. The pixel sum for DAB will be normalized to SG for all cases to correct for the variable amount of total epithelium on each core. The normalized sums are expected to be maximal for tumor sections where AMACR expression is commonly positive in most cells of most tumors but to exhibit minimum overlap in cases of BPH. Indeed simple thresholding may succeed defining a single value that best separates average tumor from average BPH. This may be expected since AMACR labeling will be applied based on optimization of tumor sections. Third, visual score by two pathologists (S. Krajewski and D. Mercola) will be acquired for all the single-antibody (DAB or SG) labeled TMAs. The results of spectral unmixing for DAB and SG will be compared to visual scoring for these chromagens as for the previous studies. Finally, the normalized DAB pixel sum is expected accurately correlate with the percent tumor cell component determined by the pathology and especially to correlate with the ration of percent DAB positive tumor cells over percent positive SG cytokeratin cells Thus, globally we predict:
On a case by case basis plots of normalize DAB/SG vs. percent DAB positive/percent SG are predicted to have a high Pearson correlation with a slope ˜1 and error similar to the preliminary Results of <10%. Validation of spectral unmizing for this chromaphore system will provide a major milestone of Phase I and means of automated antibody biomarker screening of Phase II.
Candidate stroma biomarker antibodies will be treated in a converse fashion. Mutually exclusive pixel sums (all pixels other than cytokeratin-positive pixels) will be integrated. This guarantees that epithelial components. These values will be normalized to the nonepithelial pixel sum intensity for a trichrome stain of the TMA using a second spectral unmixing calculation to identify connective tissue component (blue).
We are aware that the quantification method being developed here has numerous additional standardization issues. It is entirely dependent on the properties of reference antibodies to define “cell-type”. Antiamacr is in wide clinical use for the identification of prostate tumor cells in non prostate tissue in the presence of other components including glands. Nevertheless it is not unchallenged and “negative” results have been noted to occur for up to 30% of prostate cancer cells [76-81]. Thus pixels identified by these criteria may only “sample” a large proportion of tumor cells. This may be acceptable unless particular classes of tumor cells such as those expressing genes correlating with, say, rercurrence, are preferentially negative. It will be important to utilize other criteria such as visual inspection by trained pathologist and the use of other faithful tumor cell markers reveal significant bias. We have identified a large panel of genes that are preferentially expressed by prostate tumor cells [18]. In addition, standard alternatives such as antiPSA and antiPSMA may be compared to determine labeling deficiency by antiAmacr.
We have chosen to concentrate on the use of monoclonal antibodies for these studies as they generally display higher specificity and consistency compared to polyclonals and are therefore better adapted to commercialization into clinical development. Polyclonal antibodies are commercially available and might prove to be more sensitive in FFPE tissues, and therefore may be explored. Commissioned monoclonal antibodies are amenable to clear definition of ownership and path to market.
Many antibodies against prostate cancer tissues are commercially available. However, antibodies against important biomarkers that are not currently commercially available or that fail to meet quality control specified in specific aim 1 will be made using peptide antigens (Lampire Biologicals, San Diego, Calif.) as for previous studies [50, 68-74].
Finally an important challenge in Phase II will be the combining of multiple antibodies with possible individual optimization protocols to a single tissue section. If this can not be achieved conveniently, i.e. without serial application, the panel will be applied on multiple slides using 2-3 different antibodies of the panel per slide. Although less convenient, the use of two or possible three serial sections of patient biopsy tissue does materially effect the ability to derive prognosis from our predictive antibody panel.
A. Specific Aims.
Nomograms are sets of clinical parameters that are used to estimate the risk of prostate cancer recurrence [1, 2]. We propose to improve on the current nomograms by including predictions based on gene expression.
We have used a novel strategy to identify and validate genes whose expression correlates with prostate cancer progression in either tumor tissue or in stroma near to tumor, across multiple independent microarray datasets. We will convert this set of expression differences into a clinical assay. Our proposed strategy involves monitoring a panel of RNAs, including some RNAs that predict the risk of disease recurrence, some RNAs for housekeeping genes (internal controls), and some RNAs that are used to determine the tissue composition of a prostate sample (tumor, stroma, BPH). The inclusion of RNAs to monitor tissue percentage allows only suitable prognostic markers to be monitored in each sample; those prognostic markers that are directed towards the primary tissue in that particular sample.
We will use an RNA detection strategy (QuantiGene Plex 2.0) that works on both fresh frozen and FFPE samples, and that can accurately monitor up to 36 different RNAs, simultaneously. The assay runs on the FDA-approved Luminex platform, already used in clinical labs. We will first screen our candidate RNAs for those that perform well on this platform using RNA from fresh frozen samples with known microarray expression patterns. Panels will then be applied to 150 tumor-enriched FFPE samples and 150 stroma-enriched (near to tumor), from prostate cancer patients, with up to two decades of clinical history. The best performing subset of genes will be assembled into two panels for clinical use, one for use in stroma-enriched samples, and the other to be used in tumor-enriched samples.
The long-term goal is to validate the classifiers in a prospective study on newly recruited prostatectomy samples.
B. Background and Significance.
Cancer and the Need for Prognostic Markers.
Prostate cancer is the most common malignancy of males in the United States [3]. Patients newly diagnosed with advanced prostate cancer that do not yet have evidence of metastases are generally advised to submit to invasive therapies such as radical prostatectomy or radiation treatment. However, the majority of prostate cancers are a slow growing indolent form with a low risk of mortality. Patients with early stage disease and extremely favorable nomogram scores, suggesting indolence of the cancer, can instead opt for intensive vigilance. We propose the development of a gene-expression-based clinical test that makes a differential prognostic prediction between indolent and aggressive forms of prostate cancer. This test would provide an additional key aid to prostate cancer patients, and doctors, in making their treatment decisions, and will be particularly useful for those patients that are not at the extremes of the current nomogram scoring systems [1, 2].
While other studies to detect RNA-based prognosticators for prostate cancer have been performed, they have limited agreement with each other, and very limited overlap with prognosticators found by other methods [4-7]. We have developed a different method that identifies prognostic markers and we have cross-validated them across different data sets (detailed below). We now propose to convert a panel of these prognosticators into a useful clinical assay. We will use the QuantiGene Plex 2.0 Assay (Panomics, Inc., Fremont, Calif.), which is as sensitive as real time PCR but can be much more extensively multiplexed [8, 9]. The assay can detect up to 36 targets per well. The assay is based on the branched DNA (bDNA) technology, which amplifies signal directly from captured target RNA without purification or reverse transcription. RNA quantitation is performed directly from fresh frozen tissue or from formalin-fixed, paraffin-embedded (FFPE) tissue homogenates, and is relatively insensitive to RNA degradation and to chemical modifications introduced by formalin-fixation [10, 11]. The method is already in the FDA-approved clinical diagnostic VERSANT 3.0 assays for HIV, HCV and HBV viral load [12] and has been used in biomarker discovery, secondary screening, microarray validation, quantification of RNAi knockdowns and predictive toxicology [11, 13-15].
C. Preliminary Studies.
The key to this project is the set of genes that we will put into the prognostic assay. We describe how we obtained these genes in some detail here.
We previously developed methods to determine the genes preferentially expressed by the three major cell types of tumor-bearing prostate tissue: tumor epithelial cells, benign epithelial cells (BPH) and stromal cells [16]. We have now extended this method so that we can now identify transcription changes that correlate with early cancer recurrence in one or more of these three cell types. In addition to transcription changes in tumor cells that correlate with recurrence, we find that prognostic changes also occur in stroma near to tumor but not in BPH. We have validated a subset of these new recurrence-related genes using independent publicly available microarray data sets. Table 31 summarizes the data sets we have analyzed from various sources, including our own prostatectomy samples.
Identification of Cell-Specific Genes.
Most previous experiments to determine expression profiles of solid tumors using microarrays involved “enriched” tumor fractions. There are three limitations of this strategy. First, samples vary in purity, introducing an error due to various amounts of accompanying tissue types. Second, the change in gene expression of other cell types is subsumed in a single number, obscuring the unique profiles of these accompanying cell types. Third, substantial amounts of stroma are intrinsic to the structure of nearly all prostate tumors. We devised a method for the deconvolution of average cell-specific gene expression from a set of samples containing different mixtures of cell types [16]. Estimates of the amount of three major cell types were made: tumor epithelial cells (tumor, T), epithelium of benign prostatic hyperplasia (BPH, B), and stromal cells (S, including pooled smooth muscle, connective tissue, infiltrating immune cells, and vascular elements). The amount of mRNA (Affymetrix signal intensity, Gij) from a given gene is the sum of the amount of each cell type multiplied by the intrinsic expression, A, of that gene by the given cell type:
G
ij=βBPH,jxBPH,i+βT,jxT,i+βS,jxS,i+εij (1)
where Xi is the proportion of each cell type and ε is the error. The model identified hundreds of genes significantly more expressed in only one tissue and examples were validated by laser capture micro-dissection and immunohistochemistry [16].
In Silico Estimates of Tissue Percentages.
Estimates of tissue percentages made by pathologists for all the samples in data set 1, 2 and 3 allowed identification of individual transcript levels that correlated best with tissue percentage. The expression levels of each of these overlapping genes were fitted to a simple linear model for each tissue type and were ranked by their correlation coefficient. A subset of the top genes from one data set was subsequently used to predict tissue percentage in the other data set. The Pearson correlation coefficients between predicted cell type percentage (tumor, stroma and BPH cells) and pathologist's estimates for all pairwise predictions of the three data sets range from 0.45-0.87 (p<0.001 in all comparisons).
Estimation of cell type percentage proved to be highly relevant. In data set 4, recurrent cases had a systematically higher percentage of tumor tissue than non-recurrent cases. Unless recognized and taken into account, this skew would generate false expression-derived estimates regarding recurrence.
Identification of Cell-Specific Biomarkers of Aggressive Prostate Cancer.
We have now extended equation 1 to identify genes specific to cell-type and aggression, for cases with known follow-up history. To obtain cell-specific gene expression for both recurrent and non-recurrent cases, the summation of equation 1 is simply segregated to reserve terms with βj coefficients for non-recurrent cases and denoting recurrent cases (rs) at the end with a separate coefficient, γ
G
ij=(βBPH,jxBPH,i+βT,jxT,i+βS,jxS,i)+rs(γBPH,jxBPH,i+γT,jxT,i+γS,jxS,i)+εij (2)
Multiple linear regression (MLR) analysis was carried out leading to the calculation of all βj, all γj, and their associated t-statistic values. Thus, estimates of the intrinsic expression of three cell types (T, S and BPH) for non-recurrent and recurrent prostate cancer were derived.
In data set 1 (U133Plus2.0 array), for example, 928 differentially regulated genes were identified in early recurrent cancer types at an adjusted p value of less than 0.05, including 405 tumor- and 561 stroma-related prognostic genes. In both data sets 1 and 2, the most significant changes were observed in the stromal tissue portion of specimens that were from near tumor (reactive stroma). The ability to look for changes in expression in stroma during recurrence is one of the major advantages of our approach.
Confirmation of Prognostic Genes using Independent Data Sets (Cross-Validation).
The six available expression microarray data sets with information on prostate cancer recurrence (Table 31) allowed identification of that subset of candidate prognosticators that could be validated. We filtered all sets for γ with p<0.05; then mapped identical Affymetrix probes (data set 1, 2, 4, 5 and 6) or gene symbol (data set 2). Finally, we identified genes that occurred in both compared data sets, and showed the same direction of change in differential expression between recurrent and non-recurring samples. Overall, 152 of 185 (82.2%) genes were concordant across pairs of data sets (p<10−18). About one third of the 152 concordant genes correspond to those previously reported by others as related to outcome in prostate cancer. About a quarter may be in error (false discovery rate given that 31 of 185 were not concordant). Some sets of genes are functionally related to biological processes considered important in the progression of prostate cancer, exemplified by several members of the Wnt signal transduction pathway.
The enormous tissue percentage diversity among published data sets (all “tumor enriched” sets had some samples with less than 30% tumor, according to our in silico analysis) and a frequent bias in tumor percentages between recurrent and non-recurrent cases (leading to any tumor-specific gene being erroneously associated with recurrence) provides two explanations for the previous struggle of the community to find a valid recurrence-specific signature in any one data set.
Gene Expression Quantification Using the QuantiGene Plex 2.0 Assay.
We have tested the sensitivity and the technical and biological accuracy of the assay using a panel of genes in a 10-Plex. The ten-gene panel included two housekeeping genes and eight genes with cell type percentage predictive power for prostate tumor, stroma, and BPH. The assay was performed on 12 fresh frozen prostate cancer samples and 9 FPEE samples with various amounts of tumor, stroma, and BPH.
A standard curve for the housekeeping gene ribosomal protein S20 proved that the Plex 2.0 assay is highly reproducible and sensitive with a wide dynamic range (not shown).
Transcripts for all ten genes were accurately measured over a wide dynamic range when the template amount was over 33 ng. The gene expression levels for all eight tissue-specific genes detected by either the P1ex 2.0 assay, or the Affymetrix U133P2 array using the same RNA samples, had correlation coefficients ranging from 0.64 to 0.89. Moreover, all eight tissue-enriched genes showed good correlations with their respective cell type percentages in FFPE samples. These preliminary experiments demonstrate that the Plex 2.0 assay is a very sensitive and reproducible method, consistent with microarray data.
D. Research Design and Methods.
The thousands of tissue specific genes and over 150 candidate prognostic genes that we have identified will vary in their practical usefulness. Furthermore, not all of these genes will translate to a particular assay platform, due to circumstances such as splicing variants that may not behave identically. This project will find a subset of high performance genes for our chosen assay strategy, gleaned from among the many high-confidence candidate genes we have identified.
We will convert the gene markers into an assay that can be easily adapted in a clinical lab, using the Plex 2.0 assay on FFPE samples (no RNA extraction or reverse transcription required). For probe validation, assays will be performed on 24 total RNA samples which already have previously reported microarray data. Probes that correlate best with the microarray data will be used to analyze 150 FFPE samples with annotated recurrence status (over a decade of post-surgery follow-up in most cases). A classifier that can distinguish indolent and/or aggressive cases will be developed and outcome prediction accuracy will be estimated by cross-validation.
Step 1. Select Candidate Genes for Further Validation.
We have selected a list of gene biomarkers for further analysis, including 75 prognostic marker genes from our studies and 25 that are found in at least one of our datasets and in the literature, 30 tissue component prediction genes, and 4 housekeeping genes which represent relatively low, medium and high expression levels.
Step 2. QuantiGene Plex Assay Probe Design and Validation.
Frozen Tissue Samples.
24 total RNA samples that already have Affymetrix gene expression data will be used in the Plex 2.0 assay. The RNA samples will be selected to encompass a wide range of tissue percentages and equal numbers of non-recurrent and recurrent cases. Probes of the Plex 2.0 assay will be designed by Panomics. Each panel of the Plex 2.0 assay will contain up to 36 genes. We will test four panels, totaling 130 or more candidate genes. The assay will be performed using our Bio-Plex system which relies on FACS sorting of fluorescently encoded beads.
Selection of Genes for Future Use.
Genes that show significant correlation between the Plex assay and Affymetrix assay will be kept for further analysis. Genes with very low signal or low variance in these assays will be eliminated from further analysis. We will combine the top performing genes into three panels (36 genes per panel) for further study. If necessary, more potentially useful prognostic or tissue-enriched transcripts will be screened.
Step 3. Develop Classifiers for Recurrence Prediction.
FFPE Samples.
We will acquire a set of 150 archived prostate cancer samples from the SPECS study for validation. Two samples will be selected from each block. One will be tumor-enriched (>70% tumor cells) and the other stroma-enriched (>70% stroma cells near to tumor: “Reactive stroma”) as estimated by pathologists. These blocks have 8-20 years of associated clinical data and represent a range of overall survival and time to recurrence. Gleason scores range from 5-8. Samples will be coded for blind analysis. Plex 2.0 Assays will be performed on the three panels of above selected genes.
Outcome Prediction.
We will first use a subset of the samples with the pathologists' estimates of cell type percentages to develop linear models of cell type component prediction. Cell type percentages of the remaining samples will be estimated using these linear models and the most predictive markers will be identified to be retained in the ultimate clinical assay.
Samples will be divided into tumor-enriched samples, stroma-enriched samples. Those samples that prove not to be suitably enriched will be set aside. We will use the appropriate tissue-enriched samples to develop classifiers that distinguish aggressive and indolent cancers using Prediction Analysis for Microarrays (PAM) [17] and Support Vector Machine (SVM) [18, 19] approaches. Misclassification error will be estimated by the 10-fold cross-validation or the leave one out strategy. These tools will be implemented in R (http://www.r-project.org/). Two classifiers will be developed, one for tumor-enriched samples and one for stroma-enriched samples.
We will also attempt in silico correction of transcript levels based on the tissue percentage markers present in each multiplex. We will attempt to adjust signals to reflect the tissue percentages by simple linear regression and determine if this variable improves disease outcome prediction.
Pre- and post operation PSA, pathology T stage, and Gleason scores are available for all cases. Thus, using these parameters plus our RNA-based classifier, the nomogram-predicted disease free survival can be calculated.
Final Predictive Set.
The initial four panels of up to 36 genes, each, will be reduced to three panels after initial screening. Then these three panels used in the FFPE study will be further condensed into just two panels that contain only useful genes for tissue percentage estimation and for prognosis: one panel for stroma-enriched samples and one for tumor-enriched samples. Both panels will measure up to 10 RNAs for estimating tissue percentage, 25 RNAs for prognosis, and 3 or more housekeeping controls.
Further Studies.
Application to Biopsies.
We have found biopsies to be an excellent source of RNA. If any stroma biomarkers are associated with recurrence, we will test the Plex 2.0 assay on 10 of our hundreds of snap frozen biopsy samples to determine technical feasibility. It is possible that biopsies that are negative for cancer may still have regions that are close enough to the missed tumor that they show “reactive” gene changes. This would revolutionize the assessment of patients that are negative for cancer upon biopsy.
More Sophisticated Class Prediction Algorithms.
In this project, we propose to use in silico cell type composition prediction to estimate tumor percentages only for sample quality control. However, knowledge of tissue composition opens up opportunities for many intellectual advances in data analysis. We are developing a new classification method which takes advantage of cell composition information without rejecting any high quality data, and results in better performance than PAM and SVM-based predictions [20].
Signaling Pathway Analysis for Understanding Prostate Cancer Progression.
Our preliminary study on pathway analysis shows that our newly identified predictive markers for recurrence are significantly enriched for elements involved in cancer related pathways, exemplified by the Wnt signaling pathway. One of our long term goals is to explore the mechanisms of cancer-related pathways that are cross-validated in multiple data sets using tools such as DAVID (The Database for Annotation, Visualization and Integrated Discovery) [21, 22]. These pathways are potential targets for novel therapeutic treatment.
1. Unique in Silico Tissue Composition Prediction Strategy Based on Gene Expression Profiling.
Large variations in the proportion of tissue components in prostate cancer tissue samples lead to considerable noise and even misleading results in mining microarrays data for prognosticators. We have generated and validated linear models for tissue component estimations based on gene expression levels. Lists of 10˜20 genes that define tumor, stroma and BPH tissue, allow the proportion of each of these tissues to be determined from gene expression profiles, alone. This novel approach of in silico tissue component prediction will be used for quality control by determining the major cell components in each clinical RNA sample.
2. Unique Prognostic Gene Biomarkers.
Using a multiple linear regression model which integrates tissue component percentages, we have identified a list of tumor- and reactive stroma-associated prognostic biomarkers, which can distinguish indolent and aggressive prostate cancer. Markers were then cross-validated between different microarray data sets produced by different research groups. Most of these prognostic markers were not previously identified by other studies. This is a simple and yet novel approach to find better, more precise, prognosticators for disease progression.
3. Accurate and Sensitive Multiple Gene Expression Quantitation.
A single prostate cancer prognostic marker is unlikely to be able to classify patients. Instead, a group of markers will be needed to account for the genetic variability of patients and the variability in cancer progression. The QuantiGene Plex 2.0 assay (Panomics, Inc) allows simultaneous quantification of multiple RNA targets directly from tissue homogenates. The assay does not require RNA purification, reverse transcription, or target amplification, because it combines branched DNA (bDNA) signal amplification technology and xMAP® (multi-analyte profiling) beads. The assay uses the FDA approved Luminex system already found in clinical labs.
Our data prove the accuracy and sensitivity of the assay, and the ability to predict tissue proportions in FFPE samples. We will convert a large number of previously identified and successfully cross-validated prognostic genes into the QuantiGene assay system that can then be easily adopted by clinical labs. The QuantiGene assay gene panel will be tested on our large collection of FFPE samples that have up to decades of patient data after surgery.
We recently published a dataset for prostate cancer study (publicly available at GEO database with access number GSE8218) [3]. This dataset consists of 136 samples from 82 patients who went through prostatectomy. Of these 82 patients, 45 underwent disease relapse, 33 did not and the remaining 4 were unknown. Here we used the 130 samples with definitive relapse status for this study. In some cases, more than one sample was collected from different regions of prostate of the same patient, for example, from tumor-enriched microdissected tissue and from nontumor tissue from ≧1.5 cm from tumor (usually the contralateral lobe). For each sample which was used for microarray assay, four pathologists independently reviewed the hematoxylin and eosin (H&E) stained sections and estimated the percentages of three major cell components, i.e., tumor, stroma and BPH. The goal of this study is to identify genes that are associated with disease progression in tumor cells or maybe in other types of cells which indicate gene expression changes in the tumor micro-environment [16].
At first, we did differential analysis on all the 130 samples using the LIMMA package (http://www.bioconductor.org) in R [5]. We identified 602 altered genes between relapse and non-relapse groups by the criterion of B>0, where B represents log-likelihood-ratio of being differentially expressed versus being equivalently expressed. Thus, B>0 indicates that the gene under consideration has altered expression between relapse and non-relapse groups. The same criterion applied to the gene selection in the subsequent analyses. We then randomly selected a subset of 40, 45, . . . , 120, 125 samples from the data and carried out differential expression analysis respectively. If increase of sample size boosts power, we expect to see that more genes are detected when sample size becomes larger and the overlap of the signatures detected at different sample sizes is large, i.e., the circles and squares in
Next, we selected samples by stepwise enriching the tumor or stroma components which are two major types of cells in prostate tissue. Specifically, we used T, k % (k=0, 5, . . . , 70, 75) as cutoff for sample selection, where T stands for the percentage for tumor component. The number of genes identified in each case were summarized in
602 genes were false discoveries due to the diversity in terms of cell composition of samples. This suggested that employing all the 130 samples available is not the optimal strategy. However, there was another peak for the curve indicated by the circles when 40 samples (with tumor component greater than 35%) were used. The overlap between the detected genes at this point (as new reference gene list) with other gene lists near this point (sample size 22 to 49) was plotted in
A similar phenomenon was observed when we investigate relapse-associated stromal genes. There were two peaks for the genes predicted to associated with recurrence (circles) at sample size 70 and 92 in the right half of the plot (stroma enriched samples). The overlap between the genes identified at these two points and gene lists around these two points (24 to 106) were fairly high (≧76%, see
The original paper dealt with the heterogeneous samples via using a multiple-linear-regression (MLR) model by which the observed Affymetrix gene expression values are described as linear combination of the contribution from different types of cells [3] [17]. Specifically, the following model was applied to the expression data for each gene,
where g is the observed expression for a gene, b0 is the grand mean, C=3 indicating 3 types of cell component, pj is the percentage of cell type j, b1 represent the expression of this gene in cell type j when the case is non-relapse, γj is the extra expression (either up- or down-regulated) in cell type j when the case relapses, and finally I(RS=1) is an indicator variable with I=1 if the case relapses (denoted by RS=1) and I=0 if the case does not recur (denoted by RS=0). We reanalyzed the data with exactly the same method and detected 119 relapse-associated genes in tumor and 247 relapse-associated gene in stroma. These two gene lists have 36 and 169 genes in common respectively with the 247 genes identified for tumor (sample size=40 in
MLR are more reliable based on above reasoning; thus, we try to validate results of t-test by MLR results. For random experiment (
Suppose we calculate significance level for overlap of two tumor gene lists, i.e., 119 genes by MLR and 247 genes by t-test. Let count=0. From ˜22,000 genes, we randomly selected two gene lists of length 119 and 247, respectively. Not that 119 and 247 are the lengths of genes identified separately by t-test and MLR. If the overlap of the two randomly selected gene lists is equal or greater than 36 (observed overlap between these two tumor gene lists), we let count increase by 1. We repeated this process 10,000 times and the p-value of the observed overlap of tumor genes is calculated as
p=count/10000.
By the same means, we calculated the significance level for overlap of two stroma gene lists as well. Both p-values for tumor overlapping genes and stroma overlapping genes were ≦0.0001. This again verified the discoveries by t-test with stepwise enriched samples.
Simulated Study
In this section, we generated a dataset consisting of 200 samples each of which is composed of three types of cells. This is to mimic the situation we are facing for prostate cancer study. We randomly assigned the 200 samples into either case group (denoted by 1) or control group (denoted by 0). Here case means aggressive prostate cancers which will progress even after surgical removal prostate gland; while control denotes indolent prostate cancer which will not recur after prostatectomy. For each sample, the percentages of three cell types were simulated as follows. We let cell type 3 (BPH) be the minority cell which takes up to 10% volume in tissues; thus, we first generated the percentage of cell type 3 (x3) from uniform distribution U(0, 0.1). We then generated the percentage of cell type 1 (x1 for tumor) from U(0, 1-x3), and the percentage of cell type 2 (x2 for stroma) is therefore 1-x1-x3. For each sample, we simulated expression data for 1000 gene as follows. We let gene 1 to 60 have altered expression in cell type 1 between case and control. The differences in terms of expression for gene 1 to 20, gene 21 to 40 and gene 41 to 60 are set to 0.5, 1.0 and 2.0, respectively. The same setting was used for generating differentially expressed genes for cell type 2 (gene 61 to 120). Due to the small load for cell type 3, we assume that the difference in cell type 3 between case and control is undetectable, so we did not simulate differentially expressed genes for cell type 3.
First, we randomly selected a subset of 40, 50, . . . , 190, 200 samples from the data and carried out differential expression analysis using LIMMA. The sensitivity, specificity and false discovery rate had been logged in each situation. Such analysis was repeated 100 times and the average operating characteristic is summarized in
Considering the heterogeneity in cell composition, we then selected samples by stepwise enriching one type of cell. Specifically, we included samples with x1, k % (k=0, 5, . . . , 85, 90) in expression comparison procedure, and then identified genes that are differentially expressed in cell type 1 between case and control. With varying cutoff, the number of samples included in analysis and the sensitivity or power achieved by these samples are summarized in Table 32. Obviously, the maximum sensitivity or power is 73.3% which is much higher than any figures attained by randomly selected sample in
Finally, we applied MLR to the simulated data and the results were much improved compared to the regular t-test with enriched samples (Table 32). This is what we expected and attested plausibility of validating results of t-test by using results of MLR analysis.
indicates data missing or illegible when filed
Drosophila)
Drosophila)
The multiple linear regression method was extended to divide tumor cases into those with good outcome (never relapsed following surgery, i.e. appear to be cured) from bad outcome, i.e. in several months or years following surgery their tumor reappeared. The genes that are specifically differentially expressed in the bad outcome cases were identified (the list). These genes or a subset of them may be measure in a new patient to determine whether he matches a good or bad outcome profile. In summary, differences in RNA levels that correlated with relapse versus non-relapse were calculated for four expression microarray data sets (data set 1, 2, 3 and 4) using multiple linear regression models which used these percentages in a linear model. Many of these relapse-associated changes in transcript levels occurred in adjacent stroma. Data set 3 does not have pathologist's estimation of tissue percentage and in silico tissue prediction model was used to predict tissue percentages. The identified genes are listed in Tables 35-42.
It is to be understood that while the invention has been described in conjunction with the detailed description thereof, the foregoing description is intended to illustrate and not limit the scope of the invention, which is defined by the scope of the appended claims. Other aspects, advantages, and modifications are within the scope of the following claims.
This application claims benefit of priority from U.S. Provisional Application Ser. No. 61/119,996, filed on Dec. 4, 2008.
This invention was made with government support under grant no. CA114810 awarded by the National Institutes of Health. The government has certain rights in the invention.
| Number | Date | Country | |
|---|---|---|---|
| 61119996 | Dec 2008 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 13132878 | Jun 2011 | US |
| Child | 13857060 | US |
