Patent Application
20050142573
The present invention relates to methods for gene segregation to identify clusters of genes associated with biological sample phenotypes and for classifying biological samples on the basis of gene expression patterns derived from those samples.
For many years established human cancer cell lines have been used as models to study human cancers because, to a large degree, they faithfully recapitulate many biological features of human tumors. Established human cancer cell lines maintained in vitro are not expected to fully recapitulate the gene expression patterns of human clinical cancers. This essentially precludes their use as model systems for global gene expression analysis of human tumors. It is likely that the longer that cancer cell lines are maintained in vitro, the more they degrade as models for transcription changes in human clinical cancers.
Recent experiments using established human prostate and breast cancer cell line models indicate that this degradation may be at least partly reversed by using established cancer cell lines to generate experimental tumors in mice and to develop xenograft-derived cell lines from these experimental tumors (Glinsky, G. V., Glinskii, A. B., McClelland, M., Krones-Herzig, A., Mercola, D., Welsh, J. 2002. Microarray gene expression analysis of tumor progression in the nude mouse model of human prostate cancer. In Proceedings of the 93rd Annual Meeting of the American Association for Cancer Research, April 6-10, San Francisco, Calif., 43: 462 (Abstract#4480), incorporated herein by reference). Furthermore, the study of differential gene expression observed using cell lines maintained in vitro and in cell line-induced experimental tumors in mice avoids many of the problems associated with cellular heterogeneity and experimental manipulation of clinical samples. It appears that the in vitro and in vivo human prostate cancer progression models partially recapitulate gene expression behavior of clinical prostate tumor samples, at least with respect to the consensus differentially regulated gene class that has been recently defined for multiple xenograft-derived human prostate cancer cell lines (Glinsky, G. V., Glinskii, A. B., McClelland, M., Krones-Herzig, A., Mercola, D., Welsh, J. 2002. Microarray gene expression analysis of tumor progression in the nude mouse model of human prostate cancer. In Proceedings of the 93rd Annual Meeting of the American Association for Cancer Research, April 6-10, San Francisco, Calif., 43: 462 (Abstract#4480), incorporated herein by reference).
While several useful methods of classification of human and other tumors are known, these methods tend to be a highly subjective in nature and at best semi-quantitative. Recent advances in global gene expression analysis of human tumors using cDNA or oligonucleotide microarray technologies set the stage for the development of improved quantitative methods for human tumor classification (see, e.g., Magee, J. A., Araki, T., Patil, S., Ehrig, T., True, L., Humphrey, P. A., Catalona, W. J., Watson, M. A., Milbrandt, J. Expression profiling reveals hepsin overexpression in prostate cancer. Cancer Res., 61: 5692-5696, 2001; Dhanasekaran, S. M., Barrette, T. R., Ghosh, D., Shah, R., Varambally, S., Kurachi, K., Pienta, K. J., Rubin, M. A., Chinnalyan, A. M. Delineation of prognostic biomarkers in prostate cancer. Nature, 412:822-826, 2001; Welsh, J. B., Sapinoso, L. M., Su, A. I., Kern, S. G., Wang-Rodriguez, J., Moskaluk, C. A., Frierson, H. F., Jr., Hampton, G. M. Analysis of gene expression identifies candidate markers and pharmacological targets in prostate cancer. Cancer Res., 61: 5974-5978, 2001; Luo, J., Duggan, D. J., Chen, Y., Sauvageot, J., Ewing, C. M., Bittner, M. L., Trent, J. M., Isaacs, W. B. Human prostate cancer and benign prostatic hyperplasia: molecular dissection by gene expression profiling. Cancer Res., 61: 4683-4688, 2001; Stamey, T A, Warrington, J A, Caldwell, M C, Chen, Z, Fan, Z, Mahadevappa, M, McNeal, J E, Nolley, R, Zhang, Z. Molecular genetic profiling of Gleason grade 4/5 prostate cancers compared to benign prostatic hyperplasia. J. Urol., 166: 2171-2177, 2001; Luo, J., Dunn, T, Ewing, C, Sauvageot, J., Chen, Y, Trent, J, Isaacs, W. Gene expression signature of benign prostatic hyperplasia revealed by cDNA microarray analysis. Prostate, 51: 189-200, 2002; Singh, D., Febbo, P. G., Ross, K., Jackson, D. G., Manola, C. L., Tamayo, P., Renshaw, A. A., D'Amico, A. V., Richie, J. P., Lander, E. S., Loda, M., Kantoff, P. W., Golub, T. R., Sellers, W. R. Gene expression correlates of clinical prostate cancer behavior. Cancer Cell, 1: 203-209, 2002; Rhodes, D. R., Barrette, T. R., Rubin, M. A., Ghosh, D., Chinnaiyan, A. M. Meta-analysis of microarrays: interstudy validation of gene expression profiles reveals pathways dysregulation in prostate cancer. Cancer Res., 62: 4427-4433, 2002; Pollack, J. R., Perou, C. M., Alizadeh, A. A., Eisen, M. B., Pergamenschikov, A., Williams, C. F., Jeffrey, S. S., Botstein, D., Brown, P. O. Genome-wide analysis of DNA-copy number changes using cDNA microarrays. Nature Genetics. 1999. 23: 41-46; Forozan, F., Mahlamaki, E. H., Monni, O., Chen, Y., Veldman, R., Jiang, Y., Gooden, G. C., Ethier, S. P., Kallioniemi, A., Kallioniemi, O-P. Comparative genomic hybridization analysis of 38 breast cancer cell lines: a basis for interpreting complementary DNA microarray data. Cancer Res. 2000. 60: 4519-4525; Perou C M, Jeffrey S S, van de Rijn M, et al. Distinctive gene expression patterns in human mammary epithelial cells and breast cancers. Proc Natl Acad Sci USA. 1999. 96:9212-9217; Perou C M, Sorlie T, Eisen M B, et al. Molecular portrait of human breast tumors. Nature. 2000. 406:747-752; Clark, EA, Golub T R, Lander E S, Hynes R O. Genomic analysis of metastasis reveals an essential role for RhoC. Nature 2000. 406:532-535; Welsh, J. B., Zarrinkar, P. P., Sapinoso, L. M., Kern, S. G., Behling, C. A., Monk, B. J., Lockhart, D. J., Burger, R. A., Hampton, G. M. Analysis of gene expression profiles in normal and neoplastic ovarian tissue samples identifies candidate molecular markers of epithelial ovarian cancer. Proc Natl Acad Sci USA. 2001. 98:1176-1181, incorporated herein by reference). However, direct attempts to identify genes differentially regulated in tumors that are useful for tumor classification, clinical management and prognosis have produced limited success, in part, because of intrinsic cellular heterogeneity and variability in cellular composition of clinical samples, the statistically underdetermined nature of the problem in which the number of variables (e.g., expression data points) exceeds the number of observations (i.e., independent samples from which the data are gathered), and the absence of a uniform, readily accessible and reproducible reference standard against which differential expression can be evaluated.
In the context of clinical tumor samples, an acceptable reference standard against which differential gene expression can be evaluated should meet the following requirements:
In a more general sense, gene expression drives the acquisition of cellular phenotypes during differentiation of precursor or stem cells. Identification of genes that are differentially expressed between precursor cells and differentiated cells, or between different types of differentiated cells is an important step for understanding the molecular processes underlying differentiation. The ability to control differentiation of precursor or stem cells so as to direct the cells down a desired differentiation pathway is an important goal, as it represents a tissue engineering solution to the problem of alleviating the shortage of tissue and organs useful for grafting and transplantation. Furthermore, normal and transformed cell-type specific markers, useful for, e.g., molecular-recognition-based targeting of therapeutics such as e.g., rituximab and other recognition based therapeutics, can be identified from sets of genes concordantly regulated in particular normal and transformed cell types.
Attempts to identify directly genes that are differentially regulated in various cell lines suffer from some of the same difficulties referenced above for tumor samples. One of the most common problems for the array-based study is that they usually generate vast data sets. For example, gene expression analysis of a single tumor cell line and a single normal epithelial counterpart typically identifies many thousands of transcripts as differentially expressed at a statistically significant level. Up to 40-50% of the surveyed genes will be identified as differentially expressed when one compares gene expression profiles of normal epithelial and stromal cells. Obviously, any meaningful design of follow-up clinical and/or experimental validation experiments would require an application of further data reduction steps. Our work makes contribution to the solution of this problem by providing a convenient and simple data reduction technique. Two possible approaches seem to be appropriate: one can narrow a set of candidate genes identified in cell lines to those that maintain similar transcript abundance (or other type of gene expression) behavior in a relevant set of clinical tumor samples and design a hypothesis-driven study aimed at identifying potential biologically important genes and/or pathways using cell lines as a model system. Alternatively, one can identify or design cell lines that recapitulate gene expression behavior identified in clinical samples and again use the model system for the assessment of the biological relevance of the gene expression changes. During the last two years or so a third approach is rapidly emerging. It is based on simultaneous analysis of gene expression and DNA copy number changes with an aim to identify the genes that acquired mRNA abundance changes due to the amplification or deletion of the corresponding genes. The cancer cell lines are certainly attractive model systems to undertake such validation study. Suitable reference standards also are needed against which gene expression patterns can be evaluated in normal (i.e., not tumor) cells and/or tissues. Here again, acceptable reference standards would be expected to have the following properties:
Different types of normal cells and/or tissues should display different degrees of resemblance between their gene expression patterns as compared to the gene expression pattern exhibited by the reference standard samples;
There thus exist in the art a need for improved methods of biological sample classification, for improved methods of identifying genes that are differentially expressed or regulated in biological samples such as tumors and normal cells, for reference standards that can be used in accordance with these methods, and for identified sets of coordinately regulated genes, the expression patterns of which can be used for classifying samples and for developing cell- or tissue-specific markers. The present invention addresses these and other shortcomings of the art.
Broadly, it is an object of the invention to provide improved quantitative methods for classifying tumor and normal samples.
It is a further object of the invention to provide useful reference standards for classifying tumors and normal samples.
It is a still further object of the invention to provide methods for classifying tumor and normal samples on the basis of gene expression data.
Thus, in one aspect, the invention provides a method for classifying a sample in which a first reference set of expressed genes is identified, the first reference set consisting of genes that are differentially expressed between a first set of tumor cell lines and a set of control cell lines, a second reference set of expressed genes is identified, the second reference set consisting of genes that are differentially expressed between a first set of samples and a second set of samples, wherein the first and second samples differ with respect to a sample classification, a concordance set of expressed genes is identified, the concordance set consisting of genes that are common to the first and second reference sets and wherein, preferably, the direction of the differential expression is the same in the first and second reference sets, identifying a minimum segregation set of expressed genes within the concordance set, the minimum segregation set consisting of a subset of expressed genes within the concordance set selected so that a first correlation coefficient between an average fold-change or difference of the gene expression data from the lines and an average fold-change or difference of the gene expression data from the samples exceeds a pre-determined value, calculating for the expressed genes within the minimum segregation set a second correlation coefficient between the average fold-change or difference of the gene, expression data from the cell lines and a fold-change or difference of the gene expression data from an unclassified sample, and classifying the unclassified sample as belonging to the first set of samples or the second set of samples according to the sign of the second correlation coefficient.
In a preferred embodiment, the first set of samples and the second set of samples comprise tumor cells and/or tissues containing tumor cells, that differ with respect to a tumor classification such as, e.g., benign versus malignant growth, local and/or systemic recurrence, invasiveness, metastatic propensity, metastatic tumors versus localized primary tumors, degree of dedifferentiation (poor, moderate, or well differentiated tumors), tumor grade, Gleason score, survival prognosis, disease free survival, lymph node status, patient age, hormone receptor status, PSA level, and histologic type.
In another embodiment, reference sets are obtained without the use of cell lines, but instead rely solely on the use of clinical samples. In this embodiment, a first reference set is obtained by looking at differential expression among two or more sets of clinical samples, preferably using average expression values, wherein the two or more sets differ with respect to a known phenotype. A concordance set is then obtained by determining concordance between the differentially expressed genes established using the two or more clinical sample groups and one or more individual samples within the group that demonstrate the best fit (highest correlation coefficient) between the individual sample(s) and the average group measurements.
In other preferred embodiments, the gene expression data is selected from the group consisting of mRNA quantification data, cDNA quantification data, cRNA quantification data, and protein quantification data.
In another aspect, the invention provides for a method for identifying a set of genes in which a first reference set of expressed genes is identified, the first reference set consisting of genes that are differentially expressed between a first set of tumor cell lines and a set of control cell lines, a second reference set of expressed genes is identified, the second reference set consisting of genes that are differentially expressed between a first set of samples and a second set of samples, wherein the first and second samples differ with respect to a sample classification, a concordance set of expressed genes is identified, the concordance set consisting of genes that are common to the first and second reference sets and wherein, preferably, the direction of the differential expression is the same in the first and second reference sets, and identifying a minimum segregation set of expressed genes within the concordance set, the minimum segregation set consisting of a subset of expressed genes within the concordance set selected so that a first correlation coefficient between an average fold-change or difference of the gene expression data from the lines and an average fold-change or difference of the gene expression data from the samples exceeds a pre-determined value.
In another embodiment, the minimum segregation set is determined without use of cell line data. This embodiment is preferred when no appropriate cell lines are available. In this embodiment, two or more groups of clinical samples, differing with respect to a known phenotype are used to generate a first reference set. Preferably, this is accomplished by determining average fold expression changes (optionally log transformed), and identifying a set of differentially expressed genes that are consistently (i.e., up- or down-regulated) in one group as compared to another group. The second reference set is obtained by determining for individual sample(s) within a group, fold-expression changes for genes within the first reference set, and finding those genes concordantly over- or under-expressed, in the individual sample(s) cf. the first reference set, and identifying those individual samples for which the individual gene expression values are most highly correlated with the expression of the genes in the first reference set. This essentially consists of calculating phenotype association indices for the individual gene expression measurements within the sample, and selecting as the second reference those genes identified as being concordantly expressed in the most highly correlated individual sample(s).
In yet another preferred embodiment, the invention provides minimum segregation sets of expressed genes. Such sets have utility as tools for, e.g., sample classification or prognostication, and as sources of cell- or tissue-specific markers. The markers can be used as, e.g., targets for delivery of cell- or tissue-specific reagents or drugs, or to monitor drug effects on a molecular scale.
In yet another preferred embodiment, the invention provides a kit comprising a set of reagents useful for determining the expression of a subset of genes identified using the methods of the invention, along with instructions for their use. The reagents can be affixed to a solid support and used in a hybridization reaction, or alternatively can be primers for use in nucleic acid amplification reactions.
Additional advantages and aspects of the present invention are now described with reference to the detailed description and drawings, below.
Definitions
All terms, unless specifically defined below, are intended to have their ordinary meanings as understood by those of skill in the art. Claimed masses and volumes are intended to encompass variations in the sated quantities compatible with the practice of the invention. Such variations are contemplated to be within, e.g. about +10-20 percent of the stated quantities. In case of conflict between the specific definitions contained in this section and the ordinary meanings as understood by those of skill in the art, the definitions supplied below are to control.
“Identifying a set of expressed genes” refers to any method now known or later developed to assess gene expression, including but not limited to measurements relating to the biological processes of nucleic acid amplification, transcription, RNA splicing, and translation. Thus, direct and indirect measures of gene copy number (e.g., as by fluorescence in situ hybridization or other type of quantitative hybridization measurement, or by quantitative PCR), transcript concentration (e.g., as by Northern blotting, expression array measurements or quantitative RT-PCR), and protein concentration (e.g., by quantitative 2-D gel electrophoresis, mass spectrometry, Western blotting, ELISA, or other method for determining protein concentration) are intended to be encompassed within the scope of the definition.
“Differentially expressed” refers to the existence of a difference in the expression level of a gene as compared between two sample classes. Differences in the expression levels of “differentially expressed” genes preferably are statistically significant.
“Tumor” is to be construed broadly to refer to any and all types of solid and diffuse malignant neoplasias including but not limited to sarcomas, carcinomas, leukaemias, lymphomas, etc., and includes by way of example, but not limitation, tumors found within prostate, breast, colon, lung, and ovarian tissues.
A “tumor cell line” refers to a transformed cell line derived from a tumor sample. Usually, a “tumor cell line” is capable of generating a tumor upon explant into an appropriate host. A “tumor cell line” line usually retains, in vitro, properties in common with the tumor from which it is derived, including, e.g., loss of differentiation, loss of contact inhibition, and will undergo essentially unlimited cell divisions in vitro.
A “control cell line” refers to a non-transformed, usually primary culture of a normally differentiated cell type. In the practice of the invention, it is preferable to use a “control cell line” and a “tumor cell line” that are related with respect to the tissue of origin, to improve the likelihood that observed gene expression differences are related to gene expression changes underlying the transformation from control cell to tumor.
An “unclassified sample” refers to a sample for which classification is obtained by applying the methods of the present invention. An “unclassified sample” may be one that has been classified previously using the methods of the present invention, or through the use of other molecular biological or pathohistological analyses. Alternatively, an “unclassified sample” may be one on which no classification has been carried out prior to the use of the sample for classification by the methods of the present invention.
“According to the sign of” a correlation coefficient refers to a determination based on the sign, i.e., positive or negative, of the referenced correlation coefficient. For example, a sample may be classified as belonging to a first set of samples if the sign of the correlation coefficient is positive, or as belonging to a second set of samples if the correlation coefficient is negative.
“Orthotopic” refers to the placement of cells in an organ or tissue of origin, and is intended to encompass placement within the same species or in a different species from which the cells are originally derived.
“Ectopic” refers to the placement of cells in an organ or tissue other than the organ or tissue of origin, and is intended to encompass placement within the same species or in a different species from which the cells are originally derived.
Introduction
Completion of the draft sequence of the human genome offers an unprecedented opportunity to study the genetic basis of human cancer progression. During malignant progression, genomic instability leads to continuously emerging phenotypic diversity, clonal evolution, and clonal selection resulting in the remarkable cellular heterogeneity of tumors. The phenotypic diversity of cancer cells is associated with significant mutation-driven changes in gene expression, although not all mutations and differences in gene expression are crucial or even relevant to the malignant phenotype. It therefore is important to identify expression changes that are highly relevant and characteristic of malignant phenotypes and progression pathways, more than one of which may exist (Hanahan, D., Weinberg, R. A. The hallmarks of cancer. Cell. 2000. 100: 57-70, incorporated herein by reference.). The methods of the present invention address this goal by providing analytical techniques to identify those expression changes highly correlated with and indeed predictive of certain clinically relevant features of malignant phenotypes and progression pathways.
In a broad and general sense, as applied to the analysis of tumor samples, the methods of the invention use gene expression data from a set of tumor cell lines and compare those data with gene expression data from a set of control cell lines to identify those genes that are differentially expressed in the tumor cell lines as compared to the control cell lines. In preferred embodiments, each of these sets includes more than a single member, although it is contemplated to be within the scope of the present invention to practice embodiments in which either or both of the set of tumor cell lines and the set of control cell lines includes only one member. The identified genes are referred to as a first reference set of expressed genes. Preferably, the control cell line and the tumor cell lines are related insofar as the control cell lines represent physiologically normal cells from the tissue or organ from which the tumor represented by the tumor cell lines arose. For example, if the tumor cell lines are derived from a prostate tumor, the control cell lines preferably are primary cultures of normal prostate epithelial cells. In the preferred embodiments, more than one tumor cell line and more than one control cell line is used to generate the reference set so as to reduce the number of genes in the first reference set by eliminating those genes that are not consistently differentially expressed between the tumor and control cell lines.
In other embodiments, the method may be practiced using only one tumor cell line and one control cell line, and identifying the set of genes differentially expressed between the tumor cell line and the control cell line. However, by carrying out a series of comparison between multiple control cell lines and multiple tumor cell lines the first reference set is more likely to contain only those genes that are consistently differentially expressed between the normal and tumor classes of cell lines (i.e., a gene is included within the first reference set if its expression level is always higher in each of the tumor cell lines examined as compared to each of the control cell lines examined, or if its expression level is always lower in each of the tumor cell lines examined as compared to each of the control cell lines examined).
In yet another embodiment, exemplified below as Example 6, the methods of the invention may be practiced without the use of cell lines, using instead data derived only from clinical samples. In a similar manner, the methods of the invention may be practiced using only data derived from cell lines.
For example, consider an embodiment in which the first reference set is derived using data obtained from three separate control cell lines and six separate tumor cell lines. For each gene considered for inclusion within the first reference set, pairwise comparisons are carried out for each of the 3×6 or 18 pairwise combinations between control cell lines and tumor cell lines. A candidate gene will be included in the first reference set if each of the 18 pairwise comparisons reveals the gene to be consistently differentially expressed (i.e., gene expression always is higher in the control cell line or always higher in the tumor cell line for each of the 18 pairwise comparisons). As one of ordinary skill readily will appreciate, it may sometimes be necessary to scale the datasets prior to carrying out the pairwise comparisons. Such scaling may be routinely implemented in the analysis software provided by commercial suppliers of expression arrays or array readers (such as, e.g., Affymetrix, Santa Clara, Calif.). For a general discussion of data scaling for and differential gene expression analysis, see, e.g., Affymetrix Microarray Suite 4.0 User Guide, Affymetrix, Santa Clara, Calif., incorporated herein by reference.
The first reference set therefore is a set of genes that have met a screening criterion requiring that the genes be differentially expressed between tumor and control cell lines. This criterion reflects the hypothesis that differences in the tumor and control cell phenotypes are driven, at least in part, by differences in gene expression patterns in the tumor and control cells. In the practice of the invention, generating a first reference set typically results in an order of magnitude or greater reduction in the number of genes that remain under consideration for inclusion in a cluster or for use in the sample classification methods.
Because the tumor and control cell lines have at some point been cultured in vitro, their gene expression patterns likely will not exactly correspond with the expression patterns of their counterparts grown in vivo. Consequently, the methods of the invention use additional steps to establish a second reference set of expressed genes that are differentially expressed in cells of biological samples that differ with respect to a classification. The classification may be an outcome predictor or cellular phenotype or any type of classification that may be used for classifying biological samples. The classification may be binary (i.e., for two mutually exclusive classes such as, e.g., invasive/non-invasive, metastatic/non-metastatic, etc.), or may be continuously or discretely variable (i.e., a classification that can assume more than two values such as, e.g., Gleason scores, survival odds, etc.) The only requirement is that the classified trait must be something that can be observed and characterized by the assignment of a variable or other type of identifier so that samples belonging to the same class may be grouped together during the analysis.
The second reference set of expressed genes may be obtained following essentially the same techniques described above for the first reference set, except sets of samples obtained from in vivo sources are used instead of sets of cell lines. In embodiments of the invention directed toward tumor analysis, classification or prognostication, the sample sets preferably consist of tumor samples obtained from patients that are analyzed without any intervening tissue culturing steps so that the gene expression patterns reflect as closely as possible the pattern within cells growing in their undisturbed, in vivo environment. Here again, the goal is to obtain a reference set that includes genes differentially expressed between samples belonging to different classifications. As is the case with the first reference set, it is preferable to include several independent samples within a classified set and to carry out a plurality of pairwise comparisons to identify differentially expressed genes for inclusion into the second reference set.
For example, assume the classification of interest is invasiveness (e.g., turning on whether tumor-free surgical margins are observed). It is preferable to use as the sample sets a number of invasive samples and a number of non-invasive samples. The number of pairwise comparisons that can be carried out is of course equal to the product of the numbers of independent samples in each category. Ideally, each of these pairwise comparisons is carried out and the same consistently differentially expressed criterion described above is used to select genes for inclusion into the second reference set.
It is contemplated, that in certain instances, especially, e.g., when the variance within a sample set is low, it will not be necessary to carry out all pairwise comparisons to select genes for inclusion into the first or second reference set. In practice, one of ordinary skill can readily determine whether it is advantageous to carry out all pairwise comparisons, or fewer than all pairwise comparisons by examining the convergence behavior of the reference sets as additional comparisons are carried out. If the sets apparently converge prior to completion of all possible pairwise comparisons, then the added benefit of exhaustive comparison may be small and so can be avoided.
Similar principles drive the selection of the numbers of cell lines and cell samples used to derive the first and second reference sets as apply to the study of other cell and molecular biological phenomena. One of ordinary skill readily will appreciate that the accuracy of the reference sets can increase as more cell lines and samples are used so that statistical noise is minimized. It currently is contemplated that preferred numbers of different cell lines and samples per set used for calculating reference sets be in the range of 2 to 50 per set, or in the range of 2 to 25, or in the range of 2 to 10, or in the range of 3 to 5 per set. While not preferred, it also is contemplated to be within the scope of the present invention to use sets consisting of a single type of cell in one or more of the four sets of input cells used to calculate the first and second reference sets (i.e., tumor cell lines, control cell lines, first sample, and second sample). Direct statistical analysis using T-test and/or Mann-Whitney test for identification of genes differentially expressed in sets of biological samples that differ with respect to a classification is also applicable to the methods of the present invention. The average expression values for genes across the first and second sets of biological samples that differ with respect to a classification are used for calculation of fold expression changes (see below).
After the first and second reference sets of differentially expressed genes are identified, a concordance set of expressed genes is identified. The concordance set is obtained by comparing the first and second reference sets. Two criteria preferably are used to identify genes for inclusion into the concordance set: 1) the candidate gene is present in first and second reference sets; 2) the direction of the candidate gene's differential is the same in the first and second reference sets. Again, as one of ordinary skill readily will recognize, there is a certain degree of arbitrariness to the sign of the differential, as it is determined by, e.g., the direction of the comparison between samples [sample 1/sample 2, cf. sample 2/sample 1, or alternatively, sample 1-sample 2, cf. sample 2-sample 1]. In any event, the arbitrariness does not affect the results because the direction of the comparison is the same across the entire set of expressed genes. The first criterion is, in general, required for inclusion of a gene within the concordance set, while the second criterion is preferred, but optional. In practical terms, identification of a single reference set of differentially expressed genes could serve as a starting point for identification of a concordant set of transcripts. For example, one can identify a reference set of differentially regulated genes in a panel of biological samples subject to a classification and proceed directly to identification of a concordant set of differentially regulated genes in cell lines.
Once the concordance set has been established, information about the rank order of expression differences is used to establish another subset of genes. This subset is referred to as the minimum segregation set. The minimum segregation set may conveniently be selected by generating a scatter plot from which may be determined correlations between the −fold expression change or difference in the cell lines and the samples. In preferred embodiments, the −fold expression change is used, and is calculated by obtaining for gene x the ratio of the average expression value obtained across all tumor cell lines and across all control cell lines, and across the first and in the second sample sets, i.e.,
In another preferred embodiment, differences in expression values are used and are calculated as:
difference =<expression>1−<expression>2,
In other embodiments, preferred if the number of observations of gene x expression in each set is small, (i.e., on the order of one or two), a modified average fold change across all observations, <expression>m, can be used in lieu of <expression>1/<expression>2 to improve the performance of the method. The modified average fold change <expression>m explicitly is defined as:
<expression>m=<expression>1/<expression1+expression2>
where there are observations of expression value E for gene x from set 1 and M observation of expression value E for gene x from set 2. Improvement in the method performance can be determined using samples of known classification, and assessing the overall accuracy of the method in classifying known samples using <expression>m in lieu of <expression>1/<expression>2.
Consider the following observations of expression values E for gene x in which N=M=5:
A scatter plot can be generated for genes within the concordance set in which each gene is assigned a point in the scatter plot. The (x,y) location of that point will be, or will be proportional to, the −fold expression change or difference in the cell line data (e.g., x), and the −fold expression change or difference in the sample data (e.g., y). Of course, the selection of the data assigned to be plotted on the abscissa and that to be plotted on the ordinate is arbitrary, so that one could have the x value correspond to the sample data and the y value correspond to the cell line data. In preferred embodiments, the −fold expression change or difference data is logarithmically transformed prior to plotting said data on the scatter plot.
The scatter plot potentially will be populated by data points that fall within any of the four quadrants of a graph in which the axes intersect at (0,0). Define quadrant I as negative x, positive y, quadrant II as positive x, positive y, quadrant III as positive x, negative y, and quadrant IV as negative x, negative y. The minimum segregation class is selected so as to include genes that fall within quadrants II and IV, and preferably to include only those genes within quadrants II and IV whose −fold expression changes or differences are highly positively correlated between the cell line and sample data. Alternatively, the minimum segregation class may be selected so as to include genes that fall within quadrants I and III, and preferably to include only those genes within quadrants I and III whose −fold expression changes or differences are highly negatively correlated between the cell line and sample data.
The scatter plots described above provide a convenient graphical representation of the data used in the clustering and classification methods of the present invention, although it is not necessary to generate such plots in the practice of the invention. Correlation coefficients can be generated for arrays of data without first plotting the data as described above. The expression data can be sorted by the values of the fold expression changes or differences and subsets of highly correlated data can be selected visually or with the aid of, e.g., regression analysis. Correlation coefficients may then be calculated on the subset of data.
Genes whose expression changes are highly correlated (positively or negatively) between the cell line and sample data may be identified by calculating a correlation coefficient for one or more subsets of genes that fall within quadrants II and IV (or alternatively for those that fall within quadrants I and III) of a scatter plot, and selecting as the minimum segregation set, those genes for which the correlation coefficient exceeds a predetermined value. Any one of a number of commonly used correlation coefficients may be used, including correlation coefficients generated for linear and non-linear regression lines through the data. Representative correlation coefficients include the correlation coefficient, px,y, that ranges between −1 and +1, such as is generated by Microsoft Excel's CORREL function, the Pearson product moment correlation coefficient, r, that also ranges between −1 and +1, that that reflects the extent of a linear relationship between two data sets, such as is generated by Microsoft Excel's PEARSON function, or the square of the Pearson product moment correlation coefficient, r2, through data points in known y's and known x's, such as is generated by Microsoft Excel's RSQ function. The r2 value can be interpreted as the proportion of the variance in y attributable to the variance in x.
In a preferred embodiment, the −fold expression change or difference data are logarithmically transformed (e.g., log10 transformed), and the minimum segregation set is selected so that the correlation coefficient, Px,y, is greater than or equal to 0.8, or is greater than or equal to 0.9, or is greater than or equal to 0.95, or is greater than or equal to 0.995. One of ordinary skill can readily work out equivalent values for other types of transformations (e.g. natural log transformations) and other types of correlation coefficients either mathematically, or empirically using samples of known classification.
The method can be terminated at the step of selecting the minimum segregation set. This set will consist of a collection or cluster of genes that is coordinately regulated during processes that result in phenotypic changes between the types of samples that comprise the sample sets.
The method may be continued, as described immediately below, to classify a sample as belonging to the first sample set or to the second sample set. The classification method uses a minimum segregation set of expressed genes to calculate a second correlation coefficient referred to as a “phenotype association index.” The method contemplates several different embodiments for calculating the second correlation coefficient. In a preferred embodiment, the second correlation coefficient is calculated by determining for an individual sample for which classification is sought, the −fold expression change for each gene x within the minimum segregation set. Preferably, the − fold expression change is determined with respect to the average value of expression for gene x across all samples used to identify the minimum segregation set. In the table above, assume set 1 data correspond to a first set of samples and that set 2 data correspond to a second set of samples. The average expression value for gene x across these samples is equal to 3.7. In this preferred embodiment, the −fold expression change is determined by computing the ratio of the expression value for gene x in the individual sample to the 3.7 average value across all the samples used to identify the minimum segregation set. For example, if the observed gene x expression value in the sample is 7, then the −fold expression change calculated according to this embodiment is 7/3.7=1.9. If the data were logarithmically transformed prior to identifying the minimum segregation set, then the same logarithmic transformation is carried out on the individual sample data prior to calculating the correlation coefficient.
In this preferred embodiment the classification is made according to the sign of this second correlation coefficient (phenotype association index). Given the setup outlined above, using −fold expression changes <expression>1/<expression1+expression2> for the sample sets to calculate the minimum segregation set, a positive correlation coefficient obtained for the classified sample indicates that the sample is a member of sample set 1, while a negative correlation coefficient indicates the sample belongs to sample set 2.
In a refinement of this preferred embodiment, the magnitude of the correlation coefficient can be used as a threshold for classification. The larger the magnitude of the correlation coefficient, the greater the confidence that the classification is accurate. As one of ordinary skill readily will appreciate, the appropriate threshold can be determined through the use of test data that seek to classify samples of known classification using the methods of the present invention. The threshold is adjusted so that a desired level of accuracy (e.g., greater than about 70% or greater than about 80%, or greater than about 90% or greater than about 95% or greater than about 99% accuracy is obtained). This accuracy refers to the likelihood that an assigned classification is correct. Of course, the tradeoff for the higher confidence is an increase in the fraction of samples that are unable to be classified according to the method. That is, the increase in confidence comes at the cost of a loss in sensitivity.
In another preferred embodiment, multiple minimum segregation sets can be identified and used to increase the sensitivity of the method. Here again, test data from samples of known classification are used to identify the minimum segregation sets and classify the individual samples. In a preferred embodiment, successive minimum segregation classes are identified using expression data from true positive and false positive samples. The expression data from these samples is again broken down into two sample sets, with the true positives assigned to, e.g., sample set 1, and the false positives assigned to sample set 2. The re-apportioned expression data are used to identify another concordance set and another minimum segregation set. This additional minimum segregation set is used to re-score the samples with particular attention paid to the ability of the set to properly classify the false positives.
Several such iterations can be done, and criteria developed to improve the accuracy of the method by evaluating the behavior of known samples against a number of minimum segregation sets. Such analysis can be used to show, e.g., that true positives score with the correct phenotype association index in, e.g., 3 of 3 minimum segregation sets.
As one of ordinary skill will recognize, a similar approach can be used with false negatives, wherein the true negatives and the false negatives are used in an iterative embodiment of the invention, with the false negatives re-assigned to sample set 1 and the true negatives assigned to sample set 2. Blended methods also may be used in which, e.g., the true positives and false negatives are assigned to sample set 1 and the true negatives and false positives assigned to sample set 2, or any other logical combination that uses mis-classified samples to iteratively obtain minimum segregation sets that are used either alone or in conjunction with other sets to improve the accuracy of the classification methods of the present invention.
While the clustering and classification methods have been described primarily with reference to tumor samples, they are readily applicable to any biological analysis for which appropriate cell lines and samples can be obtained. These include by way of example, but not limitation, omnipotent stem cells, pluripotent precursor cells, various terminally differentiated cells, etc. The clustering methods applied to cell differentiation analyses will identify gene clusters that are coordinately regulated in differentiation programs. These genes are useful not only from a basic research point of view (e.g., to identify novel transcription factors or response elements), but also to identify gene products specifically expressed in one but not another cell type. Such gene products are useful for, e.g., targeting of therapeutic molecules using reagents that have affinity for the specifically expressed gene products.
Application of the methods of the present invention to the study and classification of cancers represents an important advance made possible in large part by the ready availability of gene expression data. Recent gene expression analysis data revealed that direct comparison of expression profiles for individual tumors to identify the transcriptome of human cancer progression is extremely challenging. Continuous phenotypic changes in cancer cells during tumor progression, individual phenotypic variations, intrinsic cellular heterogeneity, and variability in cellular composition of the primary and metastatic tumors render extremely problematic the selection of the gene expression changes relevant to tumor progression and metastasis. Furthermore, the use of human tumors and metastatic material, itself, limits the direct manipulation of variables that might otherwise reveal regulatory defects that are not apparent in the ground state expression patterns of in vivo tumors.
A complementary experimental approach to the extensive clinical sampling was developed employing gene expression analysis of selected cancer cell lines representing divergent clinically relevant variants of cancer progression (Table 1). These cell lines were surveyed under various in vitro and in vivo conditions that model microenvironments favorable to the malignant phenotype, including differential serum withdrawal responsiveness in vitro and induction of experimental tumors in nude mice, ultimately to identify expression changes characteristic of human cancer progression. These cell lines provide a representative group of tumor cell lines that can be used in the practice of the methods of the invention (although other transformed cell lines, such as are readily available from depositories such as ATCC or commercial suppliers also can be used). The methods of the invention also may be practiced using, e.g., one or more of the 38 human breast cancer cell lines described in Forozan, F., Mahlamaki, E. H., Monni, O., Chen, Y., Veldman, R., Jiang, Y., Gooden, G. C., Ethier, S. P., Kallioniemi, A., Kallioniemi, O—P. Comparative genomic hybridization analysis of 38 breast cancer cell lines: a basis for interpreting complementary DNA microarray data. Cancer Res. 2000. 60: 4519-4525, incorporated herein by reference. The methods of the invention also may be practiced using one or more of the 60 human cancer cell lines representing multiple forms of human cancer and utilized in the National Cancer Institute's screen for anti-cancer drug was described in Ross, T D, Scherf, U, Eisen, M B, Perou, C M, Rees, C, Spellman, P, Iyer, V, Jeffrey, SS, Van de Rijn, M, Waltham, M, Pergamenschikov, A, Lee, J C F, Lashkari, D, Shalon, D, Myers, TG, Weinstein, J N, Botstein, D, Brown, P O. Systematic variation in gene expression patterns in human cancer cell lines. Nature Genetics, 24: 227-235, 2000, incorporated herein by reference. Classification of the human cancer cell lines based on the observed gene expression profiles revealed a correspondence to the tissue of origins of the corresponding tumors from which the cell lines were derived (Ross, D T, et al, 2000).
Each cell line and experimental condition provided a criterion that a gene met in order to be retained in the next step of analysis. Thus, the cancer cell lines represented in Table 1 are especially useful for the practice of the clustering and classification methods of the invention. Each step in the gene selection process (i.e., identification of a first and a second reference set, identification of a concordance set and finally, identification of a minimum segregation set) can be thought of as a cut-off criterion that allows genes to pass to the next stage in the analysis. The identified set of candidate genes that satisfies these criteria comprises genes, the differential expression of which is associated with certain features of the malignant phenotype and that is relatively insensitive to significant alterations in cell type and environmental context. Consequently, these genes represent reliable starting points for identifying genes that are commonly altered in human cancer and represent a consensus transcriptome of cancer progression. Other cell line combinations suitable for practicing the methods of the present invention are set forth in Tables 2-4. Table 2 lists representative cell line combinations for normal cells and certain cancers (e.g., breast, prostate, lung). These combinations are especially useful for identifying genetic markers that serve as diagnostics for a malignant phenotype. Such markers, in addition to providing diagnostic information, can also provide drug discovery targets. Table 2 also lists representative cell line combinations for precursor and differentiated cells, useful for identifying differentiation markers. Such markers can be used to screen for agents that activate differentiation programs to further basic research, as well as tissue engineering work. Table 3 lists additional tumor cell/control cell line combinations useful for practicing the methods of the invention to identify markers of malignant phenotype for diagnostic as well as drug discovery purposes. Table 4 provides additional primary tumor/metastatic tumor cell line combinations useful for practicing the methods of the invention to identify markers of metastatic potential for diagnostic, prognostic and therapeutic applications.
References: Pettaway, C. et al. Clin. Cancer Res., 2: 1627, 1996; Bae, V. et al. Int. J. Cancer, 58:721, 1994; Plymate, et al. J. Clin. Endocrinol., Met. 81: 3709, 1996; Morikawa et al. Cancer Res., 48: 1943, 1988; Morikawa et al. Cancer Res., 48: 6863, 1988; Schackert et al. Am. J. Pathol., 136: 95, 1990; Zhang et al. Cancer Res., 51: 2029, 1991; Zhang et al. Invasion Metastasis, 11: 204, 1991; Price et al. Cancer Res., 50: 717, 1990; Mukhopadhyay et al. Clin Exp Met., 17: 325, 1999; Glinsky et al. Clin. Exper. Metastasis, 14: 253, 1996; Glinsky et al. Cancer Res., 56: 5319, 1996; Glinsky et al. Cancer Lett., 115: 185, 1997; McConkey et al. Cancer Res., 56: 5594, 1996; Glinsky et al. Transf Med Rev 14: 326, 2000 (incorporated herein by reference).
Application of the methods of the invention to the study of particular cancers is described generally below, and is followed by specific working examples demonstrating aspects of the invention.
Prostate Cancer
As many as 50% of men, aged 70 years and over have microscopic foci of prostate cancer without clinical evidence of disease (Trump, D. L., Robertson, C. N., Holland, J. F., Frei, E., Bast, R. C., Kufe, D. W., Morton, D. L., and Weishselbaum, R. R. Neoplasms of the prostate. In: D. L. Trump, C. N. Robertson, J. F. Holland, E. Frei, R. C. Bast, D. W. Kufe, D. L. Morton, and R. R. Weishselbaum (eds.), Cancer Med, Vol. 3, pp. 1562-86. Philadelphia: Lea & Febiger, 1993.). Although some prostate cancers remain indolent and confined to the gland, other prostate cancers behave more aggressively and metastasize if not adequately treated. Prostate cancer is the second most lethal neoplasia in males after lung cancer. Because of widespread screening programs utilizing serum PSA values, many more cases of early stage disease are being diagnosed. In 1988 approximately 50% of patients were diagnosed with early stage disease (stage I and II). Today, about 75% of patients have early stage disease that is potentially curable.
Unfortunately, the only potentially curative therapy for prostate cancer consists of radical prostatectomy or other local therapies such as external irradiation, implanted irradiation seeds, or cryotherapy. The use of prostatectomy has increased in step with the amount of diagnosed early stage prostate cancer. SEER data indicates an increase in prostatectomies from 17.4 per 100,000 in 1988 to 54.6 per 100,000 in 1992. Insufficient treatment leads to local disease extension and metastasis. Current methods, such as Gleason scores are not perfectly reliably correlated with whether a tumor is aggressive or indolent. Thus, developing a treatment strategy appropriate for any individual is difficult. The recognition of those genetic changes that portend metastatic prostate cancer would, therefore, be a breakthrough. The methods of the present invention readily identify such genetic changes.
Breast Cancer
Breast cancer is the most common cancer among women in North America and Western Europe and is the second leading cause of female cancer death in the United States. In the United States, age-adjusted breast cancer incidence rates have considerably increased during last century. Approximately 40% of patients diagnosed with breast cancer have disease that has regional or distant metastases and, at present, there is no efficient curative therapy for breast cancer patients with advanced metastatic disease. Thus, developing a treatment strategy appropriate for any individual with early stage disease is difficult and insufficient treatment leads to local disease extension and metastasis. Therefore, there is an urgent clinical need for novel diagnostic methods that would allow early identification of those breast cancer patients who are likely to develop metastatic disease and would require the most aggressive and advanced forms of therapy for increased chance of survival. The identification of those genetic changes that distinguish aggressive metastatic disease and predict metastatic behavior would, therefore, be a breakthrough. The methods of the present invention provide information that allows prognostication of aggressive metastatic disease.
Recent gene expression analysis of human tumor samples employing cDNA microarray technology underscores the difficulties in identification of the cellular origin of differentially expressed transcripts in clinical samples due to the remarkable cellular heterogeneity and variability in cellular compositions of human tumors (Golub, T. R., Slonim, D. K., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J. P., Coller, H., Loh, M. L., Downing, J. R., Caliguri, M. A., Bloomfield, C. D., Lander, E. S. 1999. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science, 286: 531-537; Perou C M, Jeffrey S S, van de Rijn M, et al. Distinctive gene expression patterns in human mammary epithelial cells and breast cancers. Proc Natl Acad Sci USA. 1999. 96:9212-9217; Perou C M, Sorlie T, Eisen M B, et al. Molecular portrait of human breast tumors. Nature. 2000. 406:747-752, incorporated herein by reference). However, a cDNA microarray analysis of gene expression in melanoma cell lines of distinct metastatic potential, was successfully employed for identification of RhoC as an essential gene for the acquisition of metastatic phenotype by melanoma cells (Clark, E A, Golub T R, Lander E S, Hynes R O. Genomic analysis of metastasis reveals an essential role for RhoC. Nature 2000. 406:532-535, incorporated herein by reference). Established human cancer cell lines were utilized for parallel comparisons of the alterations in DNA copy number and gene expression associated with human breast cancer (Pollack, J. R., Perou, C. M., Alizadeh, A. A., Eisen, M. B., Pergamenschikov, A., Williams, C. F., Jeffrey, S. S., Botstein, D., Brown, P. O. Genome-wide analysis of DNA-copy number changes using cDNA microarrays. Nature Genetics. 1999. 23: 41-46; Forozan, F., Mahlamaki, E. H., Monni, O., Chen, Y., Veldman, R., Jiang, Y., Gooden, G. C., Ethier, S. P., Kallioniemi, A., Kallioniemi, O-P. Comparative genomic hybridization analysis of 38 breast cancer cell lines: a basis for interpreting complementary DNA microarray data. Cancer Res. 2000. 60: 4519-4525, incorporated herein by reference). Thus, model systems are a reasonable source of gene candidates to be studied in the much more heterogeneous environment of real human tumors.
Analysis of gene expression in normal and neoplastic ovarian human tissues using methods of the present invention revealed that high malignant potential ovarian cancers exhibited gene expression profile somewhat similar to the ovarian cancer cell lines (Welsh, J. B., Zarrinkar, P. P., Sapinoso, L. M., Kern, S. G., Behling, C. A., Monk, B. J., Lockhart, D. J., Burger, R. A., Hampton, G. M. Analysis of gene expression profiles in normal and neoplastic ovarian tissue samples identifies candidate molecular markers of epithelial ovarian cancer. Proc Natl Acad Sci USA. 2001. 98:1176-1181, incorporated herein by reference), further validating the complementary gene expression analysis approach utilizing selected established cancer cell lines and clinical samples.
Metastasis
Cancer cells have exceedingly low survival rates in the circulation (reviewed in [Glinsky, G. V. 1993. Cell adhesion and metastasis: is the site specificity of cancer metastasis determined by leukocyte-endothelial cell recognition and adhesion? Crit. Rev. Onc./Hemat., 14: 229-278, incorporated herein by reference). Even if the bloodstream contains many cancer cells, there may be no clinical or pathohistological evidence of metastatic dissemination into the target organs (Williams, W. R. The theory of Metastasis. In The Natural History of Cancer. 1908; 442-448; Goldmann, E. 1907. The growth of malignant disease in man and the lower animals, with special reference to the vascular system. Proc. R. Soc. Med., 1:1-13; Schmidt, M. B. In Die Verbreitungswege der Karzinome und die bezienhung generalisiertes sarkome su den leukamischen neubildungen. Fischer, Jena, 1903, incorporated herein by reference). The levels of metastatic efficiency at the intramicrovascular (postintravasation) phase of metastatic dissemination were shown to be only 0.2% and 0.003% in high and low metastatic variants of B16 melanoma cells, respectively, injected at a concentration of 105 cells into the tail veins of laboratory mice (Weiss, L. 1990. Metastatic inefficiency. Adv. Cancer Res., 54: 159-211; Weiss, L., Mayhew, E., Glaves-Rapp, D., Holmes, J. C. 1982. Metastatic inefficiency in mice bearing B16 melanomas. Br. J. Cancer, 45: 44-53, incorporated herein by reference). The fate of cancer cells in the circulation is a rapid phase of intramicrovascular cancer cell death, which is completed in <5 minutes and accounts for 85% of arrested cancer cells. This is followed by a slow phase of cell death, which accounts for the vast majority of the remainder (Weiss, L. 1988. Biomechanical destruction of cancer cells in the hart: a rate regulator of hematogenous metastasis. Invas. Metastasis, 8: 228-237; Weiss, L., Orr, F. W., Honn, K. V. 1988. Interactions of cancer cells with the microvasculature during metastasis. FSEB J., 2: 12-21; Weiss, L., Harlos, J. P., Elkin, G. 1989. Mechanism of mechanical trauma to Ehrlich ascites tumor cells in vitro and its relationship to rapid intravascular death during metastasis. Int. J. Cancer, 44: 143-148, incorporated herein by reference).
For example, the number of tumor cells in the lungs declined very rapidly after intravenous injection i.e., 90-99% had disappeared after 24 hours (Hewitt, H. B., Blake, A. 1975. Quantitative studies of translymphonodal passage of tumor cells naturally disseminating from a nonimmunogenic murine squamous carcinoma. Br. J. Cancer, 31: 25-35; Fidler, I. J. 1970. Metastasis: quantitative analysis of distribution and fate of tumor emboli labeled with 125I-5 iodo-2′-deoxyuridine. J. Natl. Cancer Inst., 45: 773-782; Proctor, J. W. 1976. Rat sarcoma model supports both soil seed and mechanical theories of metastatic spread. Br. J. Cancer, 34: 651-654; Proctor, J. W., Auclair, B. G., Rudenstam, C. M. 1976. The distribution and fate of blood-born 125IudR-labeled tumor cells in immune syngeneic rats. Int. J. Cancer, 18: 255-262; Weston, B. J., Carter, R. L., Eastry, G. C., Connell, D. I., Davies, A. J. C. 1974. The growth and metastasis of an allografted lymphoma in normal, deprived and reconstituted mice. Int. J. Cancer, 14: 176-185; Kodama, M., Kodama, T. 1975. Enhancing effect of hydrocortisone on hematogenous metastasis of Ehrlich ascites tumor in mice. Cancer Res., 35: 1015-1021, incorporated herein by reference) and after 3 days generally less than 1% remained (Fidler, I. J. 1970. Metastasis: quantitative analysis of distribution and fate of tumor emboli labeled with 125I-5 iodo-2′-deoxyuridine. J. Natl. Cancer Inst., 45: 773-782; Weston, B. J., Carter, R. L., Eastry, G. C., Connell, D. I., Davies, A. J. C. 1974. The growth and metastasis of an allografted lymphoma in normal, deprived and reconstituted mice. Int. J. Cancer, 14: 176-185; Kodama, M., Kodama, T. 1975. Enhancing effect of hydrocortisone on hematogenous metastasis of Ehrlich ascites tumor in mice. Cancer Res., 35: 1015-1021, incorporated herein by reference). This decline is due to a rapid degeneration of cancer cells (Fidler, I. J. 1970. Metastasis: quantitative analysis of distribution and fate of tumor emboli labeled with 125I-5 iodo-2′-deoxyuridine. J. Natl Cancer Inst., 45: 773-782; Roos, E., Dingemans, K. P. 1979. Mechanisms of metastasis. Biochim. Biophys. Acta, 560: 135-166, incorporated herein by reference). Therefore, the individual ‘average’ cancer cell survives only a short time in the circulation. The successful metastatic cancer cells are able to find a largely unknown survival and escape route. Patients at high risk for metastatic disease could be better managed if gene expression patterns correlated with a clinical metastatic phenotype are identified. The methods of the present invention identify such gene expression patterns. Patients' tumor samples can be tested to see whether the gene expression pattern is associated with an increased risk of metastasis, and if so, the patients can be treated with more aggressive therapies to lower the risk of metastasis. As explained in greater detail below, the present invention provides for methods that allow identification of such gene expression patterns, and sample classification based on those patterns.
Models of Human Cancer Metastasis of Graded Metastatic Potential
We have acquired several well-established and characterized model human cancer cell systems of graded metastatic potential (Table 1). The collection of these human cancer cell line panels provides different backgrounds upon which increased metastatic potential is superimposed. We have studied these cell line systems extensively for many years both in vitro and in vivo (Glinsky, G. V. 1998. Failure of Apoptosis and Cancer Metastasis. Berlin/Heidelberg: Springer-Verlag, pp. 178 et seq.; Glinsky, G. V., Mossine, V. V., Price, J. E., Bielenberg, D., Glinsky, V. V., Ananthaswamy, H. N., Feather, M. S. 1996. Inhibition of colony formation in agarose of metastatic human breast carcinoma and melanoma cells by synthetic glycoamines. Clin. Exp. Metastasis, 14: 253-267; Glinsky, G. V., Price, J. E., Glinsky, V. V., Mossine, V. V., Kiriakova, G., Metcalf, J. B. 1996. Inhibition of human breast cancer metastasis in nude mice by synthetic glycoamines. Cancer Res., 56: 5319-5324; Glinsky, G. V., Glinsky, V. V. 1996. Apoptosis and metastasis: a superior resistance of metastatic cancer cells to the programmed cell death. Cancer Lett., 101: 43-51; Glinsky, G. V., Glinsky, V. V., Ivanova, A. B., Hueser, C. N. 1997. Apoptosis and metastasis: increased apoptosis resistance of metastatic cancer cells is associated with the profound deficiency of apoptosis execution mechanisms. Cancer Lett., 115: 185-193, incorporated herein by reference) and, therefore, have considerable experience in the maintenance of cell lines preserving graded metastatic potentials. These models provide an excellent opportunity to test whether concordant changes in gene expression underlie the metastasis process and to test the efficacy of drugs designed to block one or more crucial targets.
Four important features of the selected models have been documented (Glinsky, G. V. 1997. Apoptosis in metastatic cancer cells. Crit. Rev. Onc/Hemat., 25:175-186; Glinsky, G. V. 1998. Anti-adhesion cancer therapy. Cancer and Metastasis Reviews, 17: 171-185. Glinsky, G. V. 1998. Failure of Apoptosis and Cancer Metastasis. Berlin/Heidelberg: Springer-Verlag, pp 178 et seq.; Glinsky, G. V., Mossine, V. V., Price, J. E., Bielenberg, D., Glinsky, V. V., Ananthaswamy, H. N., Feather, M. S. 1996. Inhibition of colony formation in agarose of metastatic human breast carcinoma and melanoma cells by synthetic glycoamines. Clin. Exp. Metastasis, 14: 253-267; Glinsky, G. V., Price, J. E., Glinsky, V. V., Mossine, V. V., Kiriakova, G., Metcalf, J. B. 1996. Inhibition of human breast cancer metastasis in nude mice by synthetic glycoamines. Cancer Res., 56: 5319-5324; Glinsky, G. V., Glinsky, V. V. 1996. Apoptosis and metastasis: a superior resistance of metastatic cancer cells to the programmed cell death. Cancer Lett., 101: 43-51; Glinsky, G. V., Glinsky, V. V., Ivanova, A. B., Hueser, C. N. 1997. Apoptosis and metastasis: increased apoptosis resistance of metastatic cancer cells is associated with the profound deficiency of apoptosis execution mechanisms. Cancer Lett., 115: 185-193, incorporated herein by reference): a) highly metastatic cell variants possess an increased survival ability, high clonogenic growth potential, and enhanced resistance to apoptosis compared to parental or poorly metastatic counterparts; b) treatment of highly metastatic cell variants with certain synthetic glycoamine analogues caused inhibition of clonogenic growth and survival and reversal of apoptosis resistance in vitro, as well as significant reduction of metastatic potential in vivo; c) these cell lines maintain their distinct in vivo metastatic potentials during in vitro passage for at least several months, indicating that metastatic ability is preserved in vitro; d) differential transcription profiles of four metastasis-associated genes between high and low metastatic cell variants was shown to be similar in vitro and in vivo (Greene, G. F., Kitadai, Y., Pettaway, C. A., von Eschenbach, A. C., Bucana, C. D., Fidler, I. J. 1997. Correlation of metastasis-related gene expression with metastatic potential in human prostate carcinoma cells implanted in nude mice using an in situ messenger RNA hybridization technique. American J. Pathology, 150: 1571-1582, incorporated herein by reference) indicating the potential relevance of in vitro gene expression patterns to the metastatic phenotype. Thus, in accordance with the methods of the present invention, these cellular systems can be used to identify relevant gene expression patterns associated with phenotypes of interest (such as, e.g., metastasis, invasiveness, etc.) by comparing patterns of differential gene expression in one or more independently selected cell line variants with those in different types of clinical human cancer samples.
Orthotopic Model of Human Cancer Metastasis in Nude Mice
When human tumor cells are injected into ectopic sites in nude mice most do not metastasize (Fidler, I. J. The nude mouse model for studies of human cancer metastasis. In: V. Schirrmacher and R. Schwartz-Abliez (eds.). pp. 11-17. Berlin: Springer-Verlag, 1989; Fidler, I. J. Critical factors in the biology of human cancer metastasis. 1990. Cancer Res., 50, 6130-6138, incorporated herein by reference). The normal host tissue environment influences metastatic ability of cancer cells in such a way that many human and animal tumors transplanted into nude mice metastasize only if placed in the orthotopic organ (Fidler, I. J. The nude mouse model for studies of human cancer metastasis. In: V. Schirrmacher and R. Schwartz-Abliez (eds.). pp. 11-17. Berlin: Springer-Verlag, 1989; Fidler, I. J. Critical factors in the biology of human cancer metastasis. 1990. Cancer Res., 50, 6130-6138; Fidler, I. J., Naito, S., Pathak, S. 1990. Orthotopic implantation is essential for the selection, growth and metastasis of human renal cell cancer in nude mice. Cancer Metastasis Rev., 9, 149-165; Giavazzi, R., Campbell, D. E., Jessup, J. M., Cleary, K., and Fidler, I. J. 1986. Metastatic behavior of tumor cells isolated from primary and metastatic human colorectal carcinomas implanted into different sites in nude mice. Cancer Res., 46: 1928-1948; Naito, S., von Eschenbach, A. C., Giavazzi, R., and Fidler, I. J. 1986. Growth and metastasis of tumor cells isolated from a renal cell carcinoma implanted into different organs of nude mice. Cancer Res., 46: 4109-4115; McLemore, T. L., et al. 1987. Novel intrapulmonary model for orthotopic propagation of human lung cancer in athymic nude mice. Cancer Res., 47: 5132-5140, incorporated herein by reference). These observations pointed out the unique opportunity to study gene expression changes associated with aggressive metastatic phenotype. A comparison of gene expression patterns using the same high metastatic variant implanted at orthotopic (metastasis promoting model) and ectopic (metastasis suppressing model) sites should provide unique information regarding differential gene expression profiles associated with metastatic behavior in vivo.
Several orthotopic models of human cancer metastasis have been developed (Fu, X., Herrera, H., and Hoffman, R. M. 1992. Orthotopic growth and metastasis of human prostate carcinoma in nude mice after transplantation of histologically intact tissue. Int. J. Cancer, 52: 987-990; Stephenson, R. A., Dinney, C. P. N., Gohji, K., Ordonez, N. G., Killion, J. J., and Fidler, I. J. 1992. Metastatic model for human prostate cancer using orthotopic implantation in nude mice. J. Natl. Cancer Inst., 84: 951-957; Pettaway, C. A., Stephenson, R. A., and Fidler, I. J. 1993. Development of orthotopic models of metastatic human prostate cancer. Cancer Bull. (Houst.), 45: 424-429; An, Z., Wang, X., Geller, J., Moossa, A. R., and Hoffman, R. M. 1998. Surgical orthotopic implantation allows high lung and lymph node metastasis expression of human prostate carcinoma cell line PC-3 in nude mice. The Prostate, 34: 169-174; Wang, X., An, Z., Geller, J., and Hoffman, R. M. 1999. High-malignancy orthotopic mouse model of human prostate cancer LNCaP. The Prostate, 39: 182-186; Yang, M., Jiang, P., Sun, F.-X., Hasegawa, S., Baranov, E., Chishima, T., Shimada, H., Moosa, A. R., and Hofman, R. M. 1999. A fluorescent orthotopic bone metastasis model of human prostate cancer. Cancer Res., 59: 781-786, incorporated herein by reference). The orthotopic model of human cancer metastasis in nude mice was used for in vivo selection of highly and poorly metastatic cell variants, employing either established panels of human cancer cell lines or cell variants derived from the same parental cell lines (Giavazzi, R., Campbell, D. E., Jessup, J. M., Cleary, K., and Fidler, I. J. 1986. Metastatic behavior of tumor cells isolated from primary and metastatic human colorectal carcinomas implanted into different sites in nude mice. Cancer Res., 46: 1928-1948; Morikawa, K., Walker, S. M., Jessup, J. M., Cleary, K., and Fidler, I. J. 1988. In vivo selection of highly metastatic cells from surgical specimens of different primary human colon carcinoma implanted in nude mice. Cancer Res., 48: 1943-1948; Dinney, C. P. N. et al. 1995. Isolation and characterization of metastatic variants from human transitional cell carcinoma passaged by orthotopic implantation in athymic nude mice. J. Urol., 154: 1532-1538, incorporated herein by reference).
This approach was successfully applied to develop a human breast cancer model of graded metastatic potential (see Glinsky, G. V., Mossine, V. V., Price, J. E., Bielenberg, D., Glinsky, V. V., Ananthaswamy, H. N., Feather, M. S. 1996. Inhibition of colony formation in agarose of metastatic human breast carcinoma and melanoma cells by synthetic glycoamines. Clin. Exp. Metastasis, 14: 253-267; Glinsky, G. V., Price, J. E., Glinsky, V. V., Mossine, V. V., Kiriakova, G., Metcalf, J. B. 1996. Inhibition of human breast cancer metastasis in nude mice by synthetic glycoamines. Cancer Res., 56: 5319-5324, incorporated herein by reference) as well as three independent panels of human prostate cancer cell lines with distinct metastatic potential (Pettaway, C. A., Pathak, S., Greene, G., Ramirez, E., Wilson, M. R., Killion, J. J., and Fidler, I. J. 1996. Selection of highly metastatic variants of different human prostatic carcinomas using orthotopic implantation in nude mice. Clinical Cancer Res., 2: 1627-1636; Bae, V. L., Jackson-Cook, C. K., Brothman, A. R., Maygarden, S. J., and Ware, J. Tumorugenicity of SV40 T antigen immortalized human prostate epithelial cells: association with decreased epidermal growth factor receptor (EGFR) expression. Int. J. Cancer 1994;58:721-29; Plymate, et al., The effect of the IGF system in human prostate epithelial cells of immortalization and transformation by SV-40 T antigen. J. Clin. Endocrinol. Met. 1996:81;3709-16; Jackson-Cook, C., Bae, V., Edelman W., Brothman, A., and Ware, J. Cytogenetic characterization of the human prostate cancer cell line P69SV40T and its novel tumorigenic sublines M2182 and M15. Cancer Genet. & Cytogenet 1996;87:14-23; Bae, V. L., Jackson-Cook, C. K., Maygarden, S. J., Plymate, S. R., Chen, J., and Ware, J. L. Metastatic subline of an SV40 large T antigen immortalized human prostate epithelial cell line. Prostate 1998;34:275-82, incorporated herein by reference). Recent experimental evidence indicates that enhancement of metastatic capability of human cancer cells transplanted orthotopically is associated with differential expression of several metastasis-associated genes that have been implicated earlier in certain key features of the metastatic phenotype (Greene, G. F., Kitadai, Y., Pettaway, C. A., von Eschenbach, A. C., Bucana, C. D., Fidler, I. J. 1997. Correlation of metastasis-related gene expression with metastatic potential in human prostate carcinoma cells implanted in nude mice using an in situ messenger RNA hybridization technique. American J. Pathology, 150: 1571-1582, incorporated herein by reference). These data support the rationale for the methods of the present invention to identify gene expression profiles associated with the phenotypes of clinical tumor samples based on a combination of in vitro gene expression analysis in one or more cell lines having a phenotype of interest (e.g., metastatic potential, invasiveness, etc.) and gene expression analysis of clinical samples.
A similar rationale supports the use of the methods of the present invention to identify gene expression patterns correlated with specific differentiation pathways associated with defined cell types (e.g., liver, skin, bone, muscle, blood, etc.), although in this instance, the preferred relevant comparisons are the gene expression profiles of one or more stem cell lines with that of the terminally differentiated cell type. (See, e.g., Table 2, supra.) In a related method of the present invention, expression analysis may be carried out on one or more different cell types using sets of genes (i.e., gene clusters) previously identified in, e.g., a biological sample analysis experiment such as the described tumor classification methods, to identify concordantly regulated genes that can be used as tissue-specific markers, or to screen for agents that may affect cellular differentiation or other aspects of cellular phenotype. Phenotype association indices can be calculated for normally differentiated tissue samples by calculating a correlation coefficient for a particular normally differentiated tissue sample against, e.g., −fold expression changes or expression differences for a minimum segregation set identified in a cancer analysis, as described above. The −fold expression changes or expression differences for the normally differentiated tissue sample can be calculated with reference to average values of gene x expression across a collection of different normal tissue samples. Expression data derived from the large collections of normal human and mouse tissue samples are available as supplemental data reported by Su, A. I. et al. Large-scale analysis of the human and mouse transcriptomes. PNAS 99: 4465-4470, 2002, incorporated herein by reference, and are available from the publicly accessible website http://expression.gnf.org, incorporated herein by reference.
Three possible outcomes are observed. In the first, no correlation is observed between the minimum segregation set and the normal tissue sample expression data implying that the regulatory pathway represented by the transcript abundance rank order within the minimum segregation set is not active. In the second, a positive correlation is seen between the −fold expression changes or differences in the minimum segregation set and the normal tissue sample implying that the regulatory pathway represented by the transcript abundance rank order within the minimum segregation set is active. In this outcome, the minimum segregation set represents a cluster of genes involved in a differentiation program and/or regulatory pathway that operates in the normal tissue sample and in the tumor cell lines. In the third outcome, a negative correlation is seen between the −fold expression changes or differences in the minimum segregation set and the normal tissue sample implying that the alternative regulatory pathway to one represented by the transcript abundance rank order within the minimum segregation set is active. In this outcome, the minimum segregation set represents a cluster of genes co-regulated in a differentiation program and/or regulatory pathway that operates in the normal tissue samples but that has failed in the tumor cell lines. Because the expression rank order of the genes within the minimum segregation class was derived from a comparison of the fold expression changes in tumor cell lines versus normal epithelial cells of the organ of cancer origin, this scenario may serve as an indicator of an active tumor suppression pathway. Gene expression profiles of human normal prostate epithelial cells and prostate cancer cell lines in culture
To identify genes expression of which is consistently altered in human prostate cancer cell lines, we searched for genes whose differential expression is retained as cells diverge through mutation, genomic instability, and possibly epigenetic mechanisms during repeated cycles of in vivo prostate cancer growth and progression in nude mice. To model this behavior, cell lines established from LNCap- and PC3-derived human prostate carcinoma xenografts were studied. Parental LNCap and PC3 cell lines represent divergent clinically relevant prostate cancer progression variants. LNCap is a relatively less aggressive, androgen-dependent cell line with wild-type p53, and PC3 is an aggressive, p53 mutated (21), and androgen independent cell line. The five cell lines, LNCapLN3, LNCapPro5, PC3M, PC3MLN4, PC3 MPro4 (Pettaway, C. A., Pathak, S., Greene, G., Ramirez, E., Wilson, M. R., Killion, J. J. and Fidler, I. J. Selection of highly metastatic variants of different human prostatic carcinomas using orthotopic implantation in nude mice. Clin Cancer Res. 1996;2:1627-36, incorporated herein by reference) represent lineages that have been derived from xenografts passaged repeatedly in the mouse to model prostate cancer growth and metastatic progression (see Table 1 and accompanying legend). The number of successive in vivo progression and in vitro expansion cycles varied from 1 to 5 in different lineages (Table 1).
The model design was based on the following considerations. Genes regulated similarly in five lineages would be expected to biased towards those genes that are relatively insensitive to the individual genetic differences in the cell's in vitro regulatory program. Furthermore, genes that are sensitive to environmental perturbations may be a source of changes that are stress-induced or are handling artifacts. This consideration also is relevant for changes associated with surgically-derived samples isolated from patients. We chose the early response to serum starvation (two hours) as a convenient method to identify and remove genes that are sensitive to environmental perturbations. Following these criteria, we identified 214 transcripts that are differentially expressed in the same direction in all five prostate cancer cell lines, relative to normal prostate epithelium (NPE), regardless of the presence or absence of serum (vs. 292 observed using data from high serum alone). 43 of these genes were consistently up-regulated and 171 were consistently down-regulated at least two-fold in all five cancer cell lines relative to NPE.
Of the 78 genes excluded by this experimental condition, only the Id3 protein and two alternatively spliced transcripts from the Id1 gene showed a common differential response to serum withdrawal within all five PC3— and LNCap-derived cell lines. Id1 and Id3 gene products are dominant negative regulators of the HLH transcription factors (Lyden, D., Young, A. Z., Zagzag, D., Yan, W., Gerald, W., O'Reilly, R., Bader, B. L., Hynes, R. O., Zhuang, Y., Manova, K., Benezra, R. Id1 and Id3 are required for neurogenesis, angiogenesis and vascularization of tumor xenografts. Nature 1999;401:670-77, incorporated herein by reference). The remaining 75 genes were differentially regulated with respect to serum withdrawal in ways that depended on the cell type. This is consistent with the view that the serum withdrawal criterion removes genes that are sensitive to both external environmental variables and internal cell line-specific context.
Gene Expression Profiles of PC3-Derived Orthotopic Tumors
To test whether the altered gene expression pattern of 214 genes identified in vitro is maintained in vivo, the common set of differentially expressed genes identified in the five cell lines relative to NPE were compared with genes that were differentially expressed in orthotopic tumors induced in nude mice using donor tumors for the PC3 lineage.
We identified a concordant gene expression profile for two tumors each independently derived from the three cell lines PC3 parental, PC3M, and PC3MLN4.79% (170 of 214 genes) of the transcripts differentially expressed in five prostate cancer cell lines in vitro were also differentially regulated in the same direction in vivo in all six orthotopic tumors. This gene set is exhaustively authenticated in thirty separate comparisons, which should, theoretically, put their regulation in these systems beyond doubt. Nevertheless, a sample of twelve up- and two down-regulated genes was tested using Q-PCR on an ABI7900 using the vendor's recommended protocols available at http://www.appliedbiosystems.con/support/tutorials/ (incorporated herein by reference). This PCR experiment used a further new batch of RNA from normal human prostate epithelial cell line and PC3M cells and human transcript-specific pairs of PCR primers. For several genes two separate sets of primers were designed and tested. Regulation was confirmed in the correct direction for these 14 genes, although the arrays tended to underestimate the magnitude of the change.
Therefore, the differential expression pattern of many of the prostate cancer-associated transcripts of PC3/LNCap consensus class identified in vitro using cell line concordance and media shift refractivity is retained in vivo in orthotopic human prostate tumors in mice. In the context of present invention, these data suggest that human prostate carcinoma xenografts may serve as a useful source of samples for identification of the reference standard data sets.
In Vivo Versus In Vitro Selection of Human Prostate Cancer-Associated Genes
To determine whether the consensus set of 214 differentially expressed genes identified here is retained in the parental cell lines, the PC3 and LNCaP cell lines that have not been serially passaged through mice were examined by microarray analysis, both in high and low serum. When concordance analysis was performed comparing the consensus list of 214 genes and genes that were differentially regulated relative to NPE in parental PC3 and LNCap cell lines, the majority of the down-regulated transcripts (133 genes; 78%) were similarly down-regulated in all 7 cell lines. However, only a small fraction (10 genes; 23%) of up-regulated transcripts was similarly differentially regulated in both parental cell lines. Thus, when compared with the five tumor-derived cell lines, PC3 and LNCaP parental cell lines have substantially smaller similarity with respect to the up-regulated transcripts, indicating that the transcripts with increased mRNA abundance levels in a set of 214 genes do not reflect in vitro selection. The significant degree of conservation of the consensus set of 214 genes in both xenograft-derived and plastic-maintained series of cancer cell lines supports the notion that plastic maintained cancer cell lines may serve as a useful source of samples for identification of the reference standard data sets.
Comparison with Clinical Human Prostate Tumors
While the genes described here are of undoubted interest as their expression is consistently altered in the multiple mouse model systems of human prostate cancer, it is not possible to say, as yet, whether they are of relevance to human disease. However, the expression levels of the genes in our stable set were analyzed published data from a group of clinical samples (Welsh, J. B., Sapinoso, L. M., Su, A. I., Kern, S. G., Wang-Rodriguez, J., Moskaluk, C. A., Frierson, H. F., Jr., Hampton, G. M. Analysis of gene expression identifies candidate markers and pharmacological targets in prostate cancer. Cancer Res., 61: 5974-5978, 2001, (supplemental data obtained from http://www.gnf.org/cancer/prostate), incorporated herein by reference).
These data must be treated with caution because the human clinical samples are highly heterogeneous, consisting of different amounts of cells of epithelial, stromal, and other origins. Nevertheless, of the genes that could be cross-referenced, 31 out of 41 up-regulated genes (76%) were more highly expressed in the majority of 24 human tumors than in a normal epithelial cell line. 32 of these genes were more highly expressed in the majority of tumors than the average expression found in nine adjacent normal prostate tissue samples. Similarly, 141 of 166 down-regulated genes (88%) were down regulated in tumors relative to normal epithelial cells, and 122 were down-regulated in tumors relative to adjacent normal prostate tissue. The similarity in the altered regulation of many of these genes in clinical tumors is an indication that these genes are relevant to the human disease.
Materials and Methods
Cell culture. Cell lines used in this study are described in Table 1. The PC3— and LNCap-derived cell lines were developed by consecutive serial orthotopic implantation, either from metastases to the lymph node (for the LN series), or reimplanted from the prostate (Pro series). This procedure generated cell variants with differing tumorigenicity, frequency and latency of regional lymph node metastasis (Pettaway, C. A., Pathak, S., Greene, G., Ramirez, E., Wilson, M. R., Killion, J. J. and Fidler, I. J. Selection of highly metastatic variants of different human prostatic carcinomas using orthotopic implantation in nude mice. Clin Cancer Res. 1996;2:1627-36, incorporated herein by reference). The LNCaP and PC-3 panels of human prostate carcinoma cell lines of graded metastatic potential were provided by Dr. C. Pettaway (M. D. Anderson Cancer Center, Houston, Tex.) and described earlier (Pettaway, C. A., Pathak, S., Greene, G., Ramirez, E., Wilson, M. R., Killion, J. J. and Fidler, I. J. Selection of highly metastatic variants of different human prostatic carcinomas using orthotopic implantation in nude mice. Clin Cancer Res. 1996;2:1627-36, incorporated herein by reference). A third progression model is represented by the P69 cell line, an SV40 large T-antigen-immortalized prostate epithelial line, and M12, a metastatic derivative of P69 (Bae, V. L., Jackson-Cook, C. K., Brothman, A. R., Maygarden, S. J., and Ware, J. Tumorugenicity of SV40 T antigen immortalized human prostate epithelial cells: association with decreased epidermal growth factor receptor (EGFR) expression. Int. J. Cancer 1994;58:721-29; Jackson-Cook, C., Bae, V., Edelman W., Brothman, A., and Ware, J. Cytogenetic characterization of the human prostate cancer cell line P69SV40T and its novel tumorigenic sublines M2182 and M15. Cancer Genet. & Cytogenet 1996;87:14-23; Bae, V. L., Jackson-Cook, C. K., Maygarden, S. J., Plymate, S. R., Chen, J., and Ware, J. L. Metastatic subline of an SV40 large T antigen immortalized human prostate epithelial cell line. Prostate 1998;34:275-82, incorporated herein by reference). The P69 cell line and M12 cell line were obtained from Dr. S. Plymate and Dr. J. Ware. Two primary human prostate epithelial and one primary human prostate stromal cell line were obtained from Clonetics/BioWhittaker (San Diego, Calif.) and grown in complete prostate epithelial and stromal growth medium provided by the supplier. Except where noted, other cell lines were grown in RPM11640 supplemented with 10% fetal bovine serum and gentamycin (Gibco BRL) to 70-80% confluence and subjected to serum starvation as described (14-16), or maintained in fresh complete media, supplemented with 10% FBS.
RNA extraction. For gene expression analysis, cells were harvested in lysis buffer 2 hrs after the last media change at 70-80% confluence and total RNA or mRNA was extracted using the RNeasy (Qiagen, Chatsworth, Calif.) or FastTract kits (Invitrogen, Carlsbad, Calif.). Cell lines were not split more than 5 times, except where noted.
Orthotopic xenografts. Orthotopic xenografts of human prostate PC3 cells and sublines (Table 1) were developed by surgical orthotopic implantation as previously described (An, Z., Wang, X., Geller, J., Moossa, A. R., Hoffman, R. M. Surgical orthotopic implantation allows high lung and lymph node metastatic expression of human prostate carcinoma cell line PC-3 in nude mice. Prostate 1998;34:169-74, incorporated herein by reference). Briefly, 2×106 cultured PC3 cells, PC3M cells, or PC3M sublines were injected subcutaneously into male athymic mice, and allowed to develop into firm palpable and visible tumors over the course of 2-4 weeks. Intact tissue was harvested from a single subcutaneous tumor and surgically implanted in the ventral lateral lobes of the prostate gland in a series of six athymic mice per cell line subtype. The mice were examined periodically for suprapubic masses, which appeared for all subline cell types, in the order PC3MLN4>PC3M>>PC3. Tumor-bearing mice were sacrificed by CO2 inhalation over dry ice and necropsy was carried out in a 2-4° C. cold room. Typically, bilaterally symmetric prostate gland tumors in the shape of greatly distended prostate glands were apparent. Prostate tumor tissue was excised and snap frozen in liquid nitrogen. The elapsed time from sacrifice to snap freezing was <20 min. A systematic gross and microscopic post mortem examination was carried out.
Tissue processing for mRNA isolation. Fresh frozen orthotopic tumor was examined by use of hematoxylin and eosin stained frozen sections. Orthotopic tumors of all sublines exhibited similar morphology consisting of sheets of monotonous closely packed tumor cells with little evidence of differentiation interrupted by only occasional zones of largely stromal components, vascular lakes, or lymphocytic infiltrates. Fragments of tumor judged free of these non-epithelial clusters were used for mRNA preparation. Frozen tissue (1-3 mm×1-3 mm) was submerged in liquid nitrogen in a ceramic mortar and ground to powder. The frozen tissue powder was dissolved and immediately processed for mRNA isolation using a Fast Tract kit for mRNA extraction (Invitrogen, Carlsbad, Calif., see above) according to the manufacturers instructions.
Affymetrix arrays. The protocol for mRNA quality control and gene expression analysis was that recommended by the array manufacturer, Affymetrix, Inc. (Santa Clara, Calif. http://www.affymetrix.com). In brief, approximately one microgram of mRNA was reverse transcribed with an oligo(dT) primer that has a T7 RNA polymerase promoter at the 5′ end. Second strand synthesis was followed by cRNA production incorporating a biotinylated base. Hybridization to Affymetrix Hu6800 arrays representing 7,129 transcripts or Affymetrix U95Av2 array representing 12,626 transcripts overnight for 16 h was followed by washing and labeling using a fluorescently labeled antibody. The arrays were read and data processed using Affymetrix equipment and software (Lockhart, D. J., Dong, H., Byrne, M. C., Follettie, M. T., Gallo, M. V., Chee, M. S., Mittmann, M., Wang, C., Kobayashi, M., Horton, H. and Brown, E. L. Expression monitoring by hybridization to high-density oligonucleotide arrays [see comments]. Nat. Biotechnol. 1996;14:1675-80, incorporated herein by reference). Detailed protocols for data analysis and documentation of the sensitivity, reproducibility and other aspects of the quantitative microarray analysis using Affymetrix technology have been reported (Lockhart, D. J., Dong, H., Byrne, M. C., Follettie, M. T., Gallo, M. V., Chee, M. S., Mittmann, M., Wang, C., Kobayashi, M., Horton, H. and Brown, E. L. Expression monitoring by hybridization to high-density oligonucleotide arrays [see comments]. Nat. Biotechnol. 1996;14: 1675-80, incorporated herein by reference).
To determine the quantitative difference in the mRNA abundance levels between two samples, in each individual sample for each gene the average expression differences were calculated from intensity measurements of perfect match (PM) probes minus corresponding control probes representing a single nucleotide mismatch (MM) oligonucleotides for each gene-specific set of 20 PM/MM pairs of oligonucleotides, after discarding the maximum, the minimum, and any outliers beyond 3 standard deviations (SD) from the average. The averages of pairwise comparisons for each individual gene were made between the samples, and the corresponding expression difference calls (see below) were made with Affymetrix software. Microsoft Access was used for other aspects of data management and storage. For each gene, a matrix-based decision concerning the difference in the mRNA abundance level between two samples was made by the software and reported as a “Difference call” (No change (NC), Increase (1), Decrease (D), Marginal increase (MI), and Marginal decrease (MD)) and the corresponding fold change ratio was calculated. 40-50% of the surveyed genes were called present by the Affymetrix software in these experiments. The concordance analysis of differential gene expression across the data set was performed using Microsoft Access and Affymetrix MicroDB software. For experiments involving study of prostate cancer, three of the normal prostate epithelial (NPE) microarrays are used as controls, and referred to as the NPE expression profile. Thus, when a gene is required to show a 2-fold or greater change relative to NPE, this must occur in all three microarrays, for either positive or negative changes. These stringent criteria exclude genes for which one of the three microarrays is in error. The strategy in this study is based on the idea that expression differences will not be called by chance in the same direction in multiple arrays (see below for statistical justification). Each gene in the final list of the 214 differentially expressed genes was required to be called exclusively as either concordantly up- or down-regulated in 30 separate comparisons (5 prostate cancer cell lines×2 experimental serum conditions×3 NPE controls) or 15 separate comparisons (5 prostate cancer cell lines×1 experimental serum condition×3 NPE controls).
Statistical analysis and quality performance criteria. We used a stringent analytical approach to test the hypothesis that there are common genes with altered mRNA abundance levels whish appear to be significantly associated with the studied phenotypes. The Affymetrix MicroDB and Affymetrix DMT software was used to identify in any given comparison of two chips only genes that are determined to be expressed at statistically significantly different (p<0.05) levels. These transcripts are called as differentially expressed. To be included in our final differentially regulated gene class the given transcript was required to be determined as differentially regulated in the same direction (up or down) at the statistically significant levels (p<0.05) e.g., in 30 independent comparisons (5 experimental cell lines×2 experimental conditions×3 control cell lines). To be recognized as differentially regulated in the orthotopic tumors any given gene of the PC3/LNCap consensus class was required to be determined differentially regulated in the same direction at the statistically significant level (p<0.05) in 18 additional independent comparisons (6 orthotopic tumors×3 controls). Despite that identified set of 214 genes is differentially expressed in described experimental systems with the extremely high level of confidence, we carried out Q-PCR confirmation analysis for a sub-set of identified genes and confirmed their differential expression in all instances using an additional independent normal human prostate epithelial cell line as a control.
Quality performance criteria adopted for the Affymetrix GeneChip system and applied in this study. 40-50% of the surveyed genes were called present by the Affymetrix software in these experiments. This is at the high end of the required standard adopted in many peer-reviewed publications using the same experimental system. Transcripts that are called present by the Affymetrix software in any given experiment were determined to have the signal intensities higher in the perfect match probe sets compared to single-nucleotide mismatch probe sets and background at the statistically significant level. This analysis was performed for each individual transcript using unique set of 20 perfect matches versus 20 single nucleotide mismatch probes. In our final list of 214 genes all transcripts were called present in at least one experimental setting. The inclusion error associated with two mRNA samples from identical cell lines was 2.7% for a difference called by the Affymetrix software. Thus, two independently obtained mRNA from the same cell lines will have 2.7% false positives. When a third independently derived epithelial cell line was included, only 4 genes (0.06%) out of 7,129 were called differentially expressed. The expression profiles of the normal prostate epithelial cell lines used in our experiments were determined to be indistinguishable. Therefore, controls are not likely source of errors in gene expression analysis performed in this study. This is particularly important, since the strategy adopted in this study is based on the idea that expression differences will not be called statistically significant by chance in the same direction in multiple arrays and during multiple independent comparisons of different phenotypes and variable experimental conditions. To impose additional stringent restrictions on possibility of a gene to be detected as concordantly differentially regulated by chance, we apply the use of multiple experimental models and vastly variable experimental settings such as in vitro and in vivo growth and varying growth conditions. Similar strategy for identification of consistent gene expression changes based on a concordant behavior of the differentially regulated genes using Affymetrix GeneChip system and software was applied and validated in several peer-reviewed published papers (see for example, Lee C K, Klopp, R G, Weindruch, R, Prolla, T A. Gene expression profile of aging and its retardation by caloric restriction. Science 1999; 285: 1390-1393; Ishida, S, Huang, E, Zuzan, H, Spang, R, Leone, G, West, M, Nevins, JR. Role for E2F in control of both DNA replication and mitotic function as revealed from DNA microarray analysis. Mol Cell Biol 2001; 21: 4684-4699, incorporated herein by reference). We applied more stringent criteria in our study requiring a concordance in at least 30 of 30 experiments compared to 6 of 6 comparisons in (Lee C K, Klopp, R G, Weindruch, R, Prolla, T A. Gene expression profile of aging and its retardation by caloric restriction. Science 1999; 285: 1390-1393, incorporated herein by reference); and 4 of 6 comparisons in (Ishida, S, Huang, E, Zuzan, H, Spang, R, Leone, G, West, M, Nevins, JR. Role for E2F in control of both DNA replication and mitotic function as revealed from DNA microarray analysis. Mol Cell Biol 2001; 21: 4684-4699, incorporated herein by reference). Ishida, et al. (Ishida, S, Huang, E, Zuzan, H, Spang, R, Leone, G, West, M, Nevins, JR. Role for E2F in control of both DNA replication and mitotic function as revealed from DNA microarray analysis. Mol Cell Biol 2001; 21: 4684-4699, incorporated herein by reference) provided a formal statistical justification that four or more concordant calls out of six comparisons cannot be explained by chance, with the probability in the range of 1 in 104.
Q-PCR confirmation analysis of the differentially regulated genes. To confirm differential regulation of the transcripts comprising a PC3/LNCap-consensus class using an independent method a sample of 14 genes (12 up-regulated and 2 down-regulated) was tested using Q-PCR on an ABI7900 according to the vendor's recommended protocols (available at http://www.appliedbiosystems.com/support/tutorials/). This PCR experiment used a further new batch of RNA from a third normal human prostate epithelial cell line and human transcript-specific pairs of PCR primers.
A. General
A first reference set for human prostate tumors was obtained by obtaining gene expression data from five prostate cancer cell lines (cell lines used were LNCapLN3; LNCapPro5; PC3M; PC3MLN4; PC3 Mpro4; see Table 1) and two different normal human prostate epithelial cell lines were obtained from Clonetics/BioWhittaker (San Diego, Calif.) and grown in complete prostate epithelial growth medium provided by the supplier. An original and a replicate data set was obtained for the first normal cell line, and the second cell line represented an independent data set from an independent epithelial cell line. Each of the tumor cell lines was derived from aggressively metastatic human prostate tumors. Consequently, we expected that these tumor cell lines should have an “invasive” phenotype because had they not been “invasive,” they would not have penetrated the prostate capsule, a step pre-requisite to metastasis.
The expression data were obtained using an Affymetrix Human Genome-U95Av2 (“HG-U95Av2”) expression array chip (Affymetrix, Santa Clara, Calif.). The HG-U95Av2 Array represents approximately 10,000 full-length genes. Data were obtained from the HG-U95Av2 according to the manufacturer's suggested protocols, as outlined in the Materials & Methods Section above
The original data set thus comprised a total of eight separate sets of gene expression data, five from the set of tumor cell lines and three from the set of epithelial cell lines. Fifteen separate pairwise comparisons were carried out to identify a first reference set of genes that were differentially expressed in the tumor cell lines and the epithelial cell lines. Differential expression was determined using Affymetrix's Microarray Suite software (versions 4.0 and 5.0). To be included in the first reference set, a candidate gene needed to meet two criteria: 1) the candidate gene was shown to be differentially expressed in each of the 15 pairwise comparisons; and 2) the direction of the differential (i.e. greater expression in the tumor cell lines cf. the epithelial cell lines or vice-versa) was consistent in each of the 15 pairwise comparisons. The first reference set comprised of 629 genes.
B. Recurrence Predictor Cluster and Sample Classification
The methods of the invention were used to identify gene clusters associated with increased likelihood of tumor recurrence. A second reference set was obtained using expression data obtained from clinical human prostate tumor samples. These data were the supplemental data reported in Singh, D., Febbo, P. G., et al., “Gene Expression Correlates of Clinical Prostate Cancer Behavior,” Cancer Cell March 20021:203-209, incorporated herein by reference. The clinical human prostate tumor samples were divided into two groups, recurrent and non-recurrent, as reported in Singh, et al. (2002). Data from twenty-one patients were evaluable with respect to recurrence following surgery. Recurrence was defined as two successive PSA values>0.2 ng/ml. Of the twenty-one patients, eight had recurrences, and thirteen patients remained relapse-free for at least four years.
Affymetrix MicroDB (version 3.0) and Affymetrix Data Mining Tools (DMT) (version 3.0) data analysis software were used to identify genes that were differentially regulated in recurrence group compared to relapse-free group of patients at the statistically significant level (p<0.05; Student T-test). Candidate genes were included in the second reference set if they were identified by the DMT software as having p values of 0.05 or less both for up-regulated and down-regulated genes. 316 genes were identified as being members of the second reference set.
A concordance set of genes was identified from the first and second reference sets. Genes were included in the concordance set if they met the following criteria: 1) the gene was identified as a member of both the first and the second reference sets; and 2) the direction of the differential was consistent in the first and the second reference sets (i.e., the gene transcript was more abundant in the tumor cell lines cf. the control cell lines and more abundant in the recurrent cf. the non-recurrent samples, or the gene transcript was less abundant in the tumor cell lines cf. the control cell lines and less abundant in the recurrent cf. the non-recurrent samples). The first criterion provides a way of minimizing the number of genes for which the pairwise comparisons are carried out for the sample data. Only those genes that are members of the first reference set need to be compared for generating the second reference set because the first criterion requires that the candidate gene be a member of both the first and second reference sets. The concordance set comprises of 19 genes.
The minimum segregation set was obtained as follows. For each gene in the concordance set, the −fold expression changes (as determined by the ratio of the relative transcript abundance levels) was determined. This was done for the cell line data by computing for each gene in the concordance set the ratio of the average expression in the tumor cell lines to the average expression in the control cell lines, and similarly the ratio of the average expression in the samples obtained from patients who relapsed (recurrent population) from those who did not relapse (non-recurrent population). Using the notation described above, this corresponds to calculating <expression>1/<expression>2 for the cell line and clinical samples data. For the cell line data, <expression>1 corresponds to the average expression value for gene x over all tumor cell lines and <expression>2 corresponds to the average expression value for gene x over all control cell lines. For the clinical sample data, <expression>1 corresponds to the average expression value for gene x over all samples from patients who relapsed and <expression>2 corresponds to the average expression value for gene x over all samples from patients who did not relapse.
The −fold expression change data were log10 transformed and the transformed data were entered as two arrays in a Microsoft Excel spreadsheet. The Excel CORREL function was used to generate a correlation coefficient that characterizes the degree to which the concordance set −fold expression changes were correlated between the cell line and clinical sample data. Typically, we observe correlation coefficients at this stage of the analysis in the range of about 0.7 to about 0.9. A scatter plot showing the relationship between the log-transformed −fold expression changes in the cell line and clinical sample data is shown in
A minimum segregation set was selected from the concordance set. This set was chosen by looking at the scatter plot (
Using this method, a total of nine genes was selected for the recurrence predictor minimum segregation set. This recurrence predictor minimum segregation set had a correlation coefficient of 0.995 for the cell line and sample −fold expression change differences. See
1LocusLink provides a single query interface to curated sequence and descriptive information about genetic loci. It presents information on official nomenclature, aliases, sequence accessions, phenotypes, EC numbers, MIM numbers, UniGene clusters, homology, map locations, and related web sites. It may be accessed through the National Center for Biotechnology Information (NCBI) website at http://www.ncbi.nlm.nih.gov/LocusLink/.
2The first entry in each cell of this column corresponds to the HUGO Gene Nomenclature Committee (“HGNC”) Approved Symbol for the gene corresponding to the Affymetrix Probe Set and LocusLink Identifiers within the same row. Information for the subject gene, associated cDNA, mRNA, and protein sequences may be obtained using the LocusLink identifier or the HGNC Approved Symbol by querying the search page at http://www.ncbi.nlm.nih.gov/LocusLink.
Note, the footnotes associated with Table 5 apply to every table in this specification that follows the same or similar format as Table 3 (i.e., column 1 contains information on the Affymetrix Probe Set ID, column 2 contains the LocusLink Identifier, and column 3 contains the gene description.
The recurrence predictor minimum segregation set was used to calculate a phenotype association indices for each of the twenty-one tumors removed from the patients described in Singh, et al. (2002) that were evaluated for recurrence. The phenotype association index was obtained by calculating for each individual tumor sample, the −fold expression change for each of the nine genes in the recurrence predictor minimum segregation set. The −fold expression change was calculated as:
expression/<expression1+expression2>
where “expression” is the observed expression level for gene x for the individual tumor, and “<expression1+expression2>” is the average gene expression level for gene x across the set of 21 tumors used to generate the recurrence predictor minimum segregation set. The −fold expression changes for these nine genes were log10 transformed, the transformed data entered as an array in a Microsoft Excel spreadsheet, and the Excel CORREL function was used to generate a correlation coefficient between the individual tumor data array and the corresponding log10 transformed data for the average −fold expression changes in the cell lines for the same nine genes (i.e., log10(<expression>1/<expression>2). This second correlation coefficient is the phenotype association index. The phenotype association index has the surprising and unexpected property of allowing the samples to be classified according to the sign of the index.
B-1. Prostate Cancer Predictor Clusters and Sample Classification
The methods of the invention were used to identify gene clusters associated with the presence of prostate carcinoma cells in a tissue sample compared to the adjacent normal tissue samples that were determined to be cancer cell free. The first reference data set was derived as described above in A. A second reference set was obtained using expression data obtained from clinical human prostate tumor samples. These data were two independent sets of the supplemental data reported in Welsh, J. B., et al., “Analysis of gene expression identifies candidate markers and pharmacological targets in prostate cancer,” Cancer Research, 2001, 61: 5974-5978; and Singh, D., Febbo, P. G., et al., “Gene Expression Correlates of Clinical Prostate Cancer Behavior,” Cancer Cell March 2002 1:203-209, incorporated herein by reference. The clinical human prostate tumor samples were divided into two groups, cancer samples and adjacent normal tissue samples, as reported in Welsh, et al. (2001). Data from twenty-five cancer samples (analysis of one tumor samples was carried out in duplicate) and nine adjacent normal tissue samples were used to identify the concordance gene set with high correlation coefficient and significant sample segregation power thus comprising genes with the properties of the minimum segregation class.
Genes were included in the concordance set if the direction of the differential was consistent in the first reference set and in the clinical samples (i.e., the gene transcript was more abundant in the tumor cell lines cf. the control cell lines and more abundant in the cancer samples cf. the adjacent normal tissue (ANT) samples, or the gene transcript was less abundant in the tumor cell lines cf. the control cell lines and less abundant in the cancer samples cf. the ANT samples. The concordance set comprising 54 genes was identified with correlation coefficient 0.823. Members of this concordance set are shown in Table 6. When applied to individual clinical samples, this gene set yielded sample segregation power of 91%. 30 of 33 clinical samples were classified correctly; 9 of 9 ANT samples displayed negative phenotype association indices while 21 of 24 cancer samples had positive phenotype association indices (
The minimum segregation set was obtained as follows. For each gene in the concordance set, the −fold expression changes (as determined by the ratio of the relative transcript abundance levels) was determined. This was done for the cell line data by computing for each gene in the concordance set the ratio of the average expression in the tumor cell lines to the average expression in the control cell lines, and similarly the ratio of the average expression values in the samples obtained from cancer samples (malignant population) from those from ANT samples (non-malignant population). Using the notation described above, this corresponds to calculating <expression>1/<expression>2 for the cell line and clinical samples data. For the cell line data, <expression>1 corresponds to the average expression value for gene x over all tumor cell lines and <expression>2 corresponds to the average expression value for gene x over all control cell lines. For the clinical sample data, <expression>1 corresponds to the average expression value for gene x over all cancer samples and <expression>2 corresponds to the average expression value for gene x over all ANT samples.
The −fold expression change data were log10 transformed and the transformed data were entered as two arrays in a Microsoft Excel spreadsheet. The Excel CORREL function was used to generate a correlation coefficient that characterizes the degree to which the concordance set −fold expression changes were correlated between the cell line and clinical sample data. Typically, we observe correlation coefficients at this stage of the analysis in the range of about 0.7 to about 0.9. A scatter plot showing the relationship between the log-transformed −fold expression changes in the cell line and clinical samples data for the 54 genes of a concordance set is shown in
A minimum segregation set was selected from the concordance set. This set was chosen by looking at the scatter plot (
Using this method, a total of ten genes were selected for the prostate cancer/normal tissue predictor minimum segregation set 1 (i.e. cluster 1) and a total of five genes was selected for the prostate cancer/normal tissue minimum segregation set 2 (i.e., cluster 2). These prostate cancer predictor minimum segregation clusters had a correlation coefficient of 0.995 (cluster 1) and 0.997 (cluster 2) for the cell line and sample −fold expression change differences. Members of these two prostate cancer minimum segregation clusters are shown in Table 7.
The prostate cancer/normal tissue minimum segregation clusters were used to calculate phenotype association indices for each of the thirty-three samples from the patients described in Welsh, et al. (2001). The phenotype association index was obtained by calculating for each individual clinical sample, the −fold expression change for each of the ten and five genes in the prostate cancer predictor minimum segregation set 1 and 2. The −fold expression change was calculated as:
expression/<expression1+expression2>
To test the performance of prostate cancer/normal tissue minimum segregation sets or clusters on independent data sets, we applied the method to classify 94 ANT and cancer samples described in Singh, D., Febbo, P. G., et al., “Gene Expression Correlates of Clinical Prostate Cancer Behavior,” Cancer Cell March 2002 1:203-209, incorporated herein by reference. This set of samples comprises of 47 cancer samples and 47 adjacent normal tissue samples obtained in each instances from the same patients. The phenotype association index was obtained by calculating for each individual clinical sample, the −fold expression change for each of the ten and five genes in the prostate cancer predictor minimum segregation set 1 and 2. The −fold expression change was calculated as:
expression/<expression1+expression2>
C. Invasion Clusters and Sample Classification
The methods of the invention were used along with the data reported by Singh, et al. (2002) to identify gene clusters associated with an invasive phenotype. Invasive phenotype was assessed by determining the presence or absence of positive surgical margins. The same first reference set described above in part A was used to generate the concordance and minimum segregation sets for invasiveness. The second reference set was obtained following the procedures described above in part B, using the supplemental data reported in Singh, et al. (2002) for fourteen invasive and 38 non-invasive human prostate tumors. Thus, the second reference set was obtained by using the Affymetrix MicroDB (version 3.0) and Affymetrix Data Mining Tools (DMT) (version 3.0) data analysis software to identify genes that were differentially regulated in invasion group compared to non-invasive group of patients at the statistically significant level (p<0.05; Student T-test). Candidate genes were included in the second reference set if they were identified by the DMT software as having p values of 0.05 or less both for up-regulated and down-regulated genes. 3869 genes were identified as being members of the second reference set.
The concordance set was obtained by selecting only those genes having a consistent direction of the differential in both the first and the second reference sets (i.e., greater gene expression in the tumor lines cf. the control lines and greater gene expression in the invasive tumor samples cf. the non-invasive tumor samples or vice-versa). The concordance set comprised 104 genes with an overall correlation coefficient of 0.755 (
A minimum segregation set was selected following the procedures described above in section B. A scatter plot was generated of the log10 transformed average −fold expression change in the cell line and average −fold expression change in the sample data. For the clinical sample data, <expression>1 corresponds to the average expression value for gene x over all samples from patients who had invasive tumors and <expression>2 corresponds to the average expression value for gene x over all samples from patients who had non-invasive tumors. The overall correlation coefficient for the invasiveness concordance set was 0.755. The invasiveness concordance set is shown in
A minimum segregation set was identified by selecting a subset of the highly correlated genes from the invasiveness concordance set. This minimum segregation set (invasion minimum segregation set 1 or invasion cluster 1) included 20 genes listed below in Table 8. The overall correlation coefficient between the cell lines and clinical samples for invasion cluster 1 was 0.980.
Note that three entries in the table correspond to the same genes, i.e., 34853_at, 209_at, and 115_at. They most likely represent the splice variants of the same gene (Hs.31989). According to Affymetrix annotation, the 34853_at is an alternative splice 3 variant of the FGFR2.
Individual phenotype association indices were calculated for each of the 14 invasive and each of the 38 non-invasive human prostate tumors according to the methods described in section B, above, using data for the 20 genes that make up invasion cluster 1. The phenotype association index for each tumor sample was calculated using the average −fold expression change data for the tumor cell line data and the individual −fold expression change data for the tumor sample. The data were log10 transformed and a correlation coefficient (phenotype association index) was calculated. The results are shown in
The greatest percentage of misclassifications obtained using invasion cluster 1 involved false positives, i.e., 17/38=44% of the non-invasive tumors were mis-classified as having an expression profile associated with the invasive phenotype. To improve the overall accuracy of the method, the sample set was re-structured so as to include data only from the twelve invasive tumors correctly classified using invasion cluster 1, and from the seventeen tumors mis-classified as false positives. (The false positives were considered to be non-invasive tumors (as, in fact they were) in carrying out the method steps to generate the second reference set, the concordance set, and the minimum segregation set.) Using this set of twenty-nine samples, another second reference set was generated by using the Affymetrix MicroDB (version 3.0) and Affymetrix Data Mining Tools (DMT) (version 3.0) data analysis software to identify genes that were differentially regulated in invasion group compared to non-invasive group of patients at the statistically significant level (p<0.05; Student T-test). Candidate genes were included in the second reference set if they were identified by the DMT software as having p values of 0.05 or less both for up-regulated and down-regulated genes. 458 genes were identified as being members of the second reference set.
Once the second reference set was generated, it was used to generate a concordance set by applying the criterion that the direction of the differential was consistent in the cell line and the clinical sample data. That is, the concordance set included only those genes present in the first and second reference sets whose expression was always greater in the tumor cell line cf. the control cell line and always greater in the invasive tumor sample cf. the non-invasive tumor sample, or vice-versa. The concordance set comprised 23 genes (r=0.809).
Once the concordance set was obtained using the data from the 29-member set of clinical samples, average expression values for genes within the concordance set were generated for the tumor cell lines, the control cell lines, the invasive tumors, and the non-invasive tumors. Average −fold expression changes were obtained, log10 transformed, and used to generate scatter plots and first correlation coefficients, as described above. A second minimum segregation set (invasion cluster 2) was identified by selecting a subset of genes from the concordance set whose −fold expression changes were highly correlated in the cell line and clinical samples. Invasion cluster 2 included 12 genes, and had an overall correlation coefficient of 0.983. See
Individual phenotype association indices were calculated for each of the 12 invasive and each of the 17 non-invasive human prostate tumors used to generate invasion cluster 2 according to the methods described in section B, above, using data for the 12 genes that make up invasion cluster 2. The phenotype association index for each tumor sample was calculated using the average −fold expression change data for the tumor cell line data and the individual −fold expression change data for the tumor sample. The data were log10 transformed and a correlation coefficient (phenotype association index) was calculated. The results are shown in
The method was iterated using the 11 properly classified invasive tumors and the 7 non-invasive tumors mis-classified as false positives using invasion cluster 2. Using the expression data from these 18 tumors (11 invasive and 7 non-invasive) and following the identical procedures as outlined above, a new second reference set of 449 genes, concordance set of 16 genes (r=0.908), and minimum segregation set (minimum segregation set 3 or invasion cluster 3) were generated. Invasion cluster 3 includes the 10 genes listed in Table 10, and had an overall correlation coefficient of 0.998, as shown in
As was done with the previous invasion clusters, individual phenotype association indices were calculated for each of the 11 invasive and each of the 7 non-invasive human prostate tumors used to generate invasion cluster 3 according to the methods described in section B, above, using data for the 10 genes that make up invasion cluster 3. The results are shown in
Of the fourteen invasive tumors comprising the original data set, 10/14=71% scored positive phenotype association indices in all three invasion clusters, 3/14=21% scored positive phenotype association indices in two of the three invasion clusters, and 1/14=7% scored a positive phenotype association index in only a single of the three invasion clusters. These data are summarized in Table 11.
Note:
1 = Positive phenotype association index;
0 = negative phenotype association index.
A similar analysis can be carried out for the 38 non-invasive tumors that comprised the original sample set. Of these thirty eight non-invasive tumors, 17/38=45% scored a positive phenotype association index in one of the three invasion clusters (one non-invasive tumor (T5) scored negatively in all three invasion clusters and included in this group), and 21/38=55% scored a positive phenotype association index in two of the three invasion clusters. These data are summarized in Table 12.
Note:
1 = Positive phenotype association index;
0 = negative phenotype association index.
Three of the invasive tumors scored positively in two of the three invasion clusters, and twenty-one of the non-invasive tumors also scored positively in two of the three invasion clusters. We iterated the method, as described above, using this group of three invasive and twenty-one non-invasive tumors to generate another second reference set, concordance set and minimum segregation set (minimum segregation set 4 or invasion cluster 4). The purpose of this experiment was to determine how well invasion cluster 4 could differentiate this set of three invasive and twenty-one non-invasive prostate tumors.
Invasion cluster 4 includes the 13 genes listed in Table 13, and had an overall correlation coefficient of 0.986, as shown in
As shown in
D. Gleason Score Clusters and Sample Classifications
The methods of the invention were used along with the data reported by Singh, et al. (2002) to identify gene clusters capable of distinguishing tumor samples having a Gleason score of 6 or 7 (low grade tumors) from those having a Gleason score of 8 or 9 (high grade tumors). The same first reference set described above in part A was used to generate concordance and minimum segregation sets for Gleason score stratification. The second reference set was obtained following the procedures described above in part B, using the supplemental data reported in Singh, et al. (2002) for 46 low grade tumors and six high-grade tumors. Thus, the second reference set was generated by using the Affymetrix MicroDB (version 3.0) and Affymetrix Data Mining Tools (DMT) (version 3.0) data analysis software to identify genes that were differentially regulated in high grade group compared to low grade group of patients at the statistically significant level (p<0.05; Student T-test). Candidate genes were included in the second reference set if they were identified by the DMT software as having p values of 0.05 or less both for up-regulated and down-regulated genes. 2144 genes were identified as being members of the second reference set.
The concordance set was obtained by selecting only those genes having a consistent direction of the differential in both the first and the second reference sets (i.e., greater gene expression in the tumor lines cf. the control lines and greater gene expression in the high grade cf. the low-grade tumor samples or vice-versa). The concordance set comprised 58 genes with an overall correlation coefficient equal to 0.823 (see
A minimum segregation set was selected following the procedures described above in section B. A scatter plot was generated of the log10 transformed average −fold expression change in the cell line and average −fold expression change in the sample data. For the clinical sample data, <expression>1 corresponds to the average expression value for gene x over all samples from patients who had tumors with Gleason scores of 8 or 9 (high grade) and <expression>2 corresponds to the average expression value for gene x over all samples from patients who had tumors with Gleason scores of 6 or 7 (low grade). The overall correlation coefficient for the high grade concordance set was 0.823. The high grade concordance set is shown in
A minimum segregation set was identified by selecting a subset of the highly correlated genes from the high grade concordance set. This minimum segregation set (Gleason Score 8/9 minimum segregation set 1 or high grade cluster 1) included 17 genes listed below in Table 14. The overall correlation coefficient between the cell lines and clinical samples for high grade cluster 1 was 0.986.
Individual phenotype association indices were calculated for each of the six high grade and each of the 46 low grade human prostate tumors used to generate high grade cluster 1 according to the methods described in section B, above, using data for the 17 genes that make up high grade cluster 1 (data not shown). Application of the classification method using high grade cluster 1 resulted in 6/6 high grade tumors having positively signed association indices, and so were correctly classified, while 26/46 of the low grade tumors had negative association indices and so were correctly classified. There thus were 20 false positives (i.e., low grade tumors improperly classified as high grade tumors) identified using high grade cluster 1. Overall, high grade cluster 1 accurately classified 32/52=62% of the tumors in this sample set.
To improve the accuracy of the method, we selected from the concordance set of 58 genes additional minimum segregation sets and tested their ability to classify tumor samples. A second minimum segregation set was identified by selecting a smaller subset of the highly correlated genes from the high grade minimum segregation cluster 1. This minimum segregation set (Gleason Score 8/9 minimum segregation set 2 or high grade cluster 2) included 12 genes listed below in Table 15. The overall correlation coefficient between the cell lines and clinical samples for high grade cluster 2 was 0.994.
Individual phenotype association indices were calculated for each of the six high grade and each of the 46 low grade human prostate tumors according to the methods described in section B, above, using data for the 12 genes that make up high grade cluster 2 (data not shown). Application of the classification method using high grade cluster 2 resulted in 6/6 high grade tumors having positively signed association indices, and so were correctly classified, while 30/46 of the low grade tumors had negative association indices and so were correctly classified. There thus were 16 false positives (i.e., low grade tumors improperly classified as high grade tumors) identified using high grade cluster 2. Overall, high grade cluster 2 accurately classified 36/52=69% of the tumors in this sample set.
A third minimum segregation set was identified by selecting a smaller subset of the highly correlated genes from the high grade minimum segregation cluster 2. This minimum segregation set (Gleason Score 8/9 minimum segregation set 3 or high grade cluster 3) included the 7 genes listed below in Table 16. The overall correlation coefficient between the cell lines and clinical samples for high grade cluster 3 was 0.970 (
Individual phenotype association indices were calculated for each of the six high grade and each of the 46 low grade human prostate tumors according to the methods described in section B, above, using data for the 7 genes that make up high grade cluster 3 (data not shown). Application of the classification method using high grade cluster 3 again resulted in 6/6 high grade tumors having positively signed association indices, and so were correctly classified, while 17/46 of the low grade tumors had negative association indices and so were correctly classified. There thus were 29 false positives (i.e., low grade tumors improperly classified as high grade tumors) identified using high grade cluster 3. Overall, high grade cluster 3 accurately classified 23/52=44% of the tumors in this sample set.
A summary of the accuracy with which the first three high grade clusters distinguished high grade (Gleason score 8 or 9) from low grade (Gleason score 6 or 7) tumors is provided in Table 17.
Note:
1 = Positive phenotype association index;
0 = negative phenotype association index.
Since the overall classification accuracy of high grade cluster 3 was lower than that of high grade cluster 1 and 2, additional high grade clusters were generated from a high grade concordance set of 58 genes. The resulting alternative minimum segregation set (ALT high grade cluster) included a total of 38 genes listed below in Table 18. The overall correlation coefficient between the cell line and clinical samples for this high grade cluster (Gleason Score 8/9 ALT high grade cluster) was 0.929 (
To further improve the overall classification accuracy, additional high grade clusters were generated by culling a subset of sample data made up of all the true positives (ie., the 6 high grade tumors correctly classified using each of the first three high grade clusters) and the set of 12 low grade tumors that scored as false positives in 3/3 of the first 3 high grade clusters (i.e., all the Gleason score 6&7 tumors that had a “0” in the “No. of Correct Classifications” column in Table 15). This subset was used to generate another second reference set, and concordance set using the same procedures outlined above. From this concordance set of 33 genes (r=0.731), a fourth minimum segregation set was identified by selecting a subset of the highly correlated genes from the new high grade concordance set. This minimum segregation set (Gleason Score 8/9 minimum segregation set 4 or high grade cluster 4) included 5 genes listed below in Table 19. The overall correlation coefficient between the cell lines and clinical samples for high grade cluster 4 was 0.995.
Phenotype association indices were calculated using the average cell line and individual sample −fold change expression data for the genes in high grade cluster 4. The sample included the 6 high grade tumors and the set of 17 low grade tumors that scored as false positives in 2/3 or 3/3 of the first three high grade clusters (i.e., all the Gleason score 6&7 tumors that had a “0” or “1” in the “No. of Correct Classifications” column in Table 17).
High grade cluster 4 correctly classified 6/6 high grade tumors, and 12/17 low grade tumors. Overall, high grade cluster 4 accurately characterized 18/23=78% of the tumors in this set.
To improve the accuracy of the classification, several additional minimum segregation sets of highly correlated genes were selected. Gleason Score 8/9 minimum segregation set 5, or high grade cluster 5, was used to generate phenotype association indices for the 6 high grade tumors (true positives) and the set of 17 low grade tumors that scored as false positives in 2/3 or 3/3 of the first three high grade clusters (i.e., all the Gleason score 6&7 tumors that had a “0” or “1” in the “No. of Correct Classifications” column in Table 17). High grade cluster 5 correctly classified 6/6 high grade tumors and 9/17 low grade tumors. Overall, high grade cluster 5 correctly classified 15/23=65% of the samples in this set.
High grade cluster 5 included 4 genes listed below in Table 20. The overall correlation coefficient between the cell lines and clinical samples for high grade cluster 5 was 0.998.
High grade cluster 6 included 7 genes and had an overall correlation coefficient of 0.995 (
Note:
1 = Positive phenotype association index;
0 = negative phenotype association index.
Application of the methods of present invention to classification of human prostate tumors according to Gleason grade revealed that high grade tumors can be readily distinguished from the majority of low grade prostate cancers based on gene expression analysis of small discrete clusters of genes. However, there is a significant fraction of low grade tumors that closely resemble transcriptional profiles of more advanced and aggressive high grade tumors suggesting that these low grade tumors may represent a precursor of aggressive metastatic disease.
D. Benign Prostatic Hyperplasia (BPH) Sample Classification
Applying method of present invention we identified a BPH vs. prostate cancer discrimination cluster comprising 14 genes listed in Table 22. In this example we utilized human prostate carcinoma cell line gene expression data to develop a first reference set and clinical sample data set presented in Stamey T A, Warrington J A, Caldwell M C, Chen Z, Fan Z, Mahadevappa M, McNeal J E, Nolley R, Zhang Z. Molecular genetic profiling of Gleason grade 4/5 prostate cancers compared to benign prostatic hyperplasia. J Urol 2001 166(6):2171-2177, 2001; incorporate herein by reference. The clinical data set consists of 17 samples obtained from 8 patients with BPH and 9 patients with prostate cancer (Stamey, T. A., et al., 2001).
We identified a concordance set of 54 genes (r=0.842) exhibiting concordant gene expression changes between prostate cancer cell lines vs. normal prostate epithelial cells and clinical samples of prostate cancer vs. BPH. As shown in
Applying the methods of the present invention, we next identified a minimum segregation set of genes (BPH minimum segregation set 1 or BPH cluster 1 (MAGE-1 cluster)—Table 24) that is able accurately discriminates between BPH and prostate cancer in clinical tissue samples derived from human prostate. This BPH vs. prostate cancer discrimination cluster comprises 14 genes displaying a high correlation coefficient of −fold expression changes in prostate cancer cell lines vs. normal prostate epithelial cells and clinical samples of prostate cancer vs. BPH (r=0.990) and high accuracy of sample classification. As shown in
E. Metastatic Prostate Cancer Sample Classification
Applying method of present invention we identified two gene clusters comprising 17 and 19 genes useful for classifying prostate cancer metastases. In this example we utilized human prostate carcinoma cell line gene expression data and clinical sample data set presented in Dhanasekaran, S. M., Barrette, T. R., Ghosh, D., Shah, R., Varambally, S., Kurachi, K., Pienta, K. J., Rubin, M. A., Chinnalyan, A. M. Delineation of prognostic biomarkers in prostate cancer. Nature, 412:822-826, 2001, incorporated herein by reference. As a starting gene set we utilized a set of 242 genes that was identified using a combination of statistical and clustering analyses approach in Dhanasekaran, S. M., et al., 2001 and was found to be useful in classification of various clinical samples using hierarchical clustering algorithm. Our initial analysis applying the methods of the present invention was performed on a small training data set comprising three human prostate cancer cell lines (LNCap; PC3; DU145), three samples of adjacent to cancer normal prostate, one sample of prostatitis, five samples of BPH, ten samples of hormone dependent localized prostate cancer, and seven samples of hormone refractory metastatic prostate cancer.
The original gene expression data were presented as log transformed −fold expression changes of a gene in a sample compared to normal human prostate. For the set of 242 genes we calculated average gene expression values for three prostate cancer cell lines (first reference set) and average expression values for group of metastatic prostate tumors vs. localized prostate tumors (second reference set). The initial set of 242 genes displayed only a weak correlation coefficient of the −fold expression changes in prostate cancer cell lines and clinical samples of metastatic prostate cancer vs. localized prostate cancer (r=0.323).
Applying the methods of the present invention, we identified a concordance set of 72 genes (r=0.866) exhibiting concordant gene expression changes between prostate cancer cell lines and clinical samples of metastatic prostate cancer vs. localized prostate cancer. When we utilized genes of this concordance set to calculate the phenotype association indices in individual clinical samples, 3 of 3 samples from ANP group, 5 of 5 samples from the BPH group, one sample of prostatitis, and five of ten samples of localized prostate cancer had negative phenotype association indices, whereas 7 of 7 samples from the metastatic prostate cancer group had positive phenotype association indices yielding overall accuracy of 84% in sample classification.
Applying the methods of the present invention, we next identified two minimum segregation sets of genes capable of accurately discriminating between metastatic prostate cancer and localized prostate cancer in clinical tissue samples derived from human prostate. The first metastatic prostate cancer (MPC) vs. localized prostate cancer (LPC) minimum segregation set or cluster (metastasis minimum segregation set 1) comprises 17 genes displaying a high correlation coefficient of fold expression changes in prostate cancer cell lines and clinical samples of metastatic prostate cancer prostate cancer vs. localized prostate cancer (r=0.988) and is highly accurate in discriminating among these different types of samples. As shown in
The second metastatic prostate cancer vs. localized prostate cancer discrimination cluster (metastasis minimum segregation set 2) comprises 19 genes displaying a high correlation coefficient of −fold expression changes in prostate cancer cell lines and clinical samples of metastatic prostate cancer prostate cancer vs. localized prostate cancer (r=0.988) and also is highly accurate in discriminating among these different types of samples. As shown in
To further validate the sample classification accuracy using an independent data set, we tested the performance of the two metastatic prostate cancer discrimination clusters on a larger set of clinical samples consisting of four samples of adjacent to cancer normal prostate (ANP), one sample of prostatitis, fourteen samples of BPH, fourteen samples of hormone dependent localized prostate cancer (LPC), and twenty samples of hormone refractory metastatic prostate cancer. As shown in
As shown in
Homo sapiens mRNA; cDNA DKFZp586G0623 (from clone DKFZp586G0623)
A recent study on gene expression profiling of breast cancer identifies 70 genes whose expression pattern is strongly predictive of a short post-diagnosis and treatment interval to distant metastases (van't Veer, L. J., et al., “Gene expression profiling predicts clinical outcome of breast cancer,” Nature, 415: 530-536, 2002, incorporated herein by reference). The expression pattern of these 70 genes discriminates with 81% (optimized sensitivity threshold) or 83% (optimal accuracy threshold) accuracy the patient's prognosis in the group of 78 young women diagnosed with sporadic lymph-node-negative breast cancer. This group comprises 34 patients who developed distant metastases within 5 years and 44 patients who continued to be disease-free after a period of at least 5 years; they constitute a poor prognosis and good prognosis group, correspondingly.
We applied the methods of the present invention to further reduce the number of genes whose expression patterns represent genetic signatures of breast cancers with “poor prognosis” or “good prognosis.” Measurements of mRNA expression levels of 70 genes in established human breast carcinoma cell lines (MCF7; MDA-MB-435; MDA-MB-468; MDA-MB-23 1; MDA-MB-435Br1; MDA-MB-435BL3) and primary cultures of normal human breast epithelial cells were performed utilizing Q-PCR method, which generally is accepted as the current most reliable method of gene expression analysis and unambiguous confirmation of gene identity. Applying the methods of the present invention, for each breast cancer cell line, concordant sets of genes were identified exhibiting both positive and negative correlation between −fold expression changes in cancer cell lines versus control cell line and the poor prognosis group versus the good prognosis group. Minimum segregation sets were selected from corresponding concordance sets and individual phenotype association indices were calculated. Three top-performing breast cancer metastasis predictor gene clusters are listed in Tables 27-29, and corresponding phenotype association indices are presented in
A breast cancer poor prognosis predictor cluster comprising 6 genes was identified (r=0.981) using MDA-MB-468 cell line gene expression profile as a reference standard (
A breast cancer good prognosis predictor cluster comprising 14 genes was identified (r=−0.952) using MDA-MB-435Br1 cell line gene expression profile as a reference standard (
Another breast cancer good prognosis minimum segregation set 2 comprising 13 genes (r=−0.992) was identified using MCF7 cell line gene expression profile as a reference standard (
To validate the classification accuracy using an independent data set, we tested performance of the 13 genes good prognosis predictor cluster (good prognosis minimum segregation set 2) on a set of 19 samples obtained from 11 breast cancer patients who developed distant metastases within five years after diagnosis and treatment and 8 patients who remained disease free for at least five years (van't Veer et al., 2002). As shown in
Lack of effective diagnostic and prognostic markers is generally considered a major problem in the clinical management of ovarian cancer—an epithelial neoplasm that has one of the worst prognoses among epithelial malignancies in women and is the leading cause of death from gynecologic cancer. The clinical utility of the most widely used biomarker of ovarian cancer, CA125, is largely limited to follow-up the response to therapy and progression of the disease and considered to be less efficient in diagnostic and prognostic applications (Meyer, T., Rustin, G. J. Br. J. Cancer, 82: 1535-1538, 2000, incorporated herein by reference).
We applied the methods of the present invention to identify gene expression profiles distinguishing poorly differentiated ovarian epithelial tumors, often exhibiting invasive, highly malignant phenotype, from less aggressive, well and moderately differentiated ovarian epithelial malignancies. Both clinical and cell line data sets utilized in this example were published in Welsh, J. B., et al., “Analysis of gene expression profiles in normal and neoplastic ovarian tissue samples identifies candidate molecular markers of epithelial ovarian cancer,” PNAS, 98: 1176-1181, 2001, incorporated herein by reference. As a starting point for identification of the concordant set of genes for established ovarian cancer cell lines and ovarian tumor tissue samples we utilized a set of the top 501 genes selected by a multidimensional statistical metric that was devised to identify genes with an expression pattern considered ideal for the molecular detection of epithelial ovarian cancer (Welsh et al., 2001). There determined that there was no significant correlation between the −fold changes in the expression levels of these 501 genes in the three cancer cell lines (SKOV8; MDA2774; CAOV3) compared to a control sample (HuOVR) and three poorly differentiated tumors (OVR—1; OVR—12; OVR—27) compared to eleven moderately and well differentiated tumors (OVR—1; —2; —5; —8; —10; —13; —16; —19; —22; —26; —28), (r=0.101).
According to the methods of present invention, we selected from the set of 501 genes two concordant sets of genes: concordant set 1 comprising 251 genes and exhibiting positive correlation (r=0.504) between cell lines and tissue samples data sets and concordant set 2, comprising 248 genes and exhibiting negative correlation (r=−0.296) between cell lines and clinical samples. We selected from concordance set 1 a set of 11 genes (ovarian cancer poor prognosis minimum segregation set 1) (ovarian cancer poor prognosis cluster—see Table 30) displaying a high positive correlation (r=0.988) between the cell lines and tissue samples data sets and exhibiting a 93% success rate in clinical sample classification based on individual phenotype association indices. As shown in
Applying the methods of the present invention, we selected from concordance set 2 a set of 10 genes (ovarian cancer good prognosis minimum segregation set 1) (ovarian cancer good prognosis cluster—see Table 31) displaying a high negative correlation (r=−0.964) between the tumor cell lines and clinical samples data sets and exhibiting a 93% success rate in clinical sample classification based on individual phenotype association indices. As shown in
Lung cancer accounts for more than 150,000 cancer-related deaths every year in the United States, thus exceeding the combined mortality caused by breast, prostate, and colorectal cancers (Greenlee, R. T., Hill-Harmon, M. B., Murray, T., Thun, M. CA Cancer J. Clin. 51: 15-36, 2001, incorporated herein by reference). Late stage of cancer at diagnosis and lack of efficient diagnostic and prognostic biomarkers are significant factors that adversely affect the clinical management of lung cancer (Mountain, C. F. Revisions in the international system for staging lung cancer. Chest, 111:1710-1717, 1997; Ihde, D. C. Chemotherapy of lung cancer. N. Engl. J. Med., 327:1434-1441, 1992; Sugita, M., Geraci, M., Gao, B., Powell, R. L., Hirsch, F. R., Johnson, G., Lapadat, R., Gabrielson, E., Bremnes, R., Bunn, P. A., Franklin, W. A. Combined use of oligonucleotide and tissue microarrays identifies cancer/testis antigens as biomarkers in lung cancer. Cancer Res., 62:3971-3979, 2002). Non-small-cell lung carcinoma (NSCLC) is a clinically and histopathologically distinct major form of lung cancer and is further classified as adenocarcinoma (most common form of NSCLC), squamous cell carcinoma, and large-cell carcinoma (Travis, W. D., Travis, L. B., Devesa, S. S. Cancer, 75:191-202, 1995).
We applied the methods of the present invention to identify gene expression profiles distinguishing lung adenoracinoma samples from normal lung specimens as well as a highly malignant phenotype of lung adenocarcinoma, associated with short survival after diagnosis and therapy, from less aggressive lung cancers, associated with longer patient survival. Both clinical and cell line data sets utilized in this example were published (Clinical data: Bhattacharjee, A., Richards, W. G., Staunton, J., Li, C., Monti, S., Vasa, P., Ladd, C., Beheshti, J., Bueno, R., Gillette, M., Loda, M., Weber, G., Mark, E. J., Lander, E. S., Wong, W., Johnson, B. E., Golub, T. R., Sugarbaker, D. J., Meyerson, M. Classification of human lung carcinomas by mRNA expression profiling reveals distinct adenocarcinoma subclasses. PNAS, 98: 13790-13795, 2001; incorporated herein by reference; Cell line data: Sugita, M., Geraci, M., Gao, B., Powell, R. L., Hirsch, F. R., Johnson, G., Lapadat, R., Gabrielson, E., Bremnes, R., Bunn, P. A., Franklin, W. A. Combined use of oligonucleotide and tissue microarrays identifies cancer/testis antigens as biomarkers in lung cancer. Cancer Res., 62:3971-3979, 2002; incorporated herein by reference. As a starting point for identification of the concordant set of genes for established lung cancer cell lines and lung cancer tissue samples we utilized a set of the 675 transcripts selected based on a statistical analysis of the quality of the dataset and variability of gene expression across dataset (Bhattacharje et al., 2001). Initial analysis showed that there was no significant correlation between the −fold changes in the expression levels of these 675 genes in the two NSCLC cancer cell lines (H647 and A549 cell lines) compared to a control sample (normal bronchial epithelial cell cultures obtained from a healthy 48-year-old donor) and 139 samples of lung adenoracinomas compared to the 17 normal lung specimens (r=0.163).
According to the methods of present invention, we selected from the set of 675 genes a concordant set of transcripts comprising 355 genes and exhibiting positive correlation (r=0.523) between cell lines and tissue samples data sets. Next we selected from the concordant set of 355 genes two minimum segregation sets of genes: a set of 13 genes (lung adenoracinoma minimum segregation set 1, also referred to as lung adenocarcinoma cluster 1—see Table 32) and a set of 26 genes (lung adenoracinoma minimum segregation set 2, also referred to as lung adenocarcinoma cluster 2—see Table 33) both displaying high positive correlation (r=0.979 and r=0.966, respectively) between the cell lines and tissue samples data sets (
Next we applied the methods of the present invention to identify gene expression profiles distinguishing highly malignant phenotype of lung adenocarcinoma, associated with short patient survival after diagnosis and therapy, from less aggressive lung cancers, associated with longer patient survival. Using the clinical data set and associated clinical history published in Bhattacharje et al., 2001, we selected two groups of adenocarcinoma patients having markedly distinct survival after diagnosis and therapy: poor prognosis group 1 comprising 34 patients with median survival of 8.5 months (range 0.1-17.3 months) and good prognosis group 2 comprising 16 patients with median survival of 84 months (range 75.4-106.1 months).
Applying the methods of the present invention, we selected from the set of 675 genes a concordant set of transcripts comprising 302 genes and exhibiting positive correlation (r=0.444) between cell lines data (NSCLC cell lines versus normal bronchial epithelial cells) and tissue samples data sets (poor prognosis samples versus good prognosis samples). We selected from the concordant set of 302 genes a set of 38 genes (lung adenocarcinoma poor prognosis predictor cluster 1—see Table 34) displaying high positive correlation (r=0.881) between the cell lines and tissue samples data sets (FIG.
44). This gene cluster exhibited a 64% success rate in clinical sample classification based on individual phenotype association indices (
Using the sample iteration and cluster reduction strategies described in the previous examples, we selected four additional sets of genes displaying high positive correlation between the cell lines (NSCLC cell lines versus normal bronchial epithelial cells) and tissue samples data sets (poor prognosis samples versus good prognosis samples) (see Tables 35-38) and thus having potential discriminating power in classification of lung adenocarcinoma samples.
The scoring summary of the individual phenotype association indices calculated for each of the five poor prognosis predictor clusters are presented in Table 39 for the good prognosis patients and in Table 40 for the poor prognosis patients. Only a single patient in the good prognosis group had one positive association index. All the remaining 15 good prognosis patients had negative phenotype association indices for each of the five poor prognosis gene clusters (Table 39). In contrast, 30 of 34 poor prognosis patients had at least one positive association index and 27 of 34 poor prognosis patients scored at least two positive phenotype association indices (Table 40). Thus, applying the methods of the present invention and applying a criterion requiring at least 1 positive phenotype association index for poor prognosis classification, 45 of 50 (90%) adenocarcinoma patients in this data set could be correctly classified as having a good or a poor prognosis.
When human cancer cells derived from the metastatic tumors are injected into ectopic sites in nude mice most do not metastasize (1, 2). The host tissue environment influences metastatic ability of cancer cells in such a way that many human and animal tumors transplanted into nude mice metastasize only if placed in the orthotopic organ (3-8). Several orthotopic models of human cancer metastasis have been developed (9-15). The orthotopic model of human cancer metastasis in nude mice was utilized for in vivo selection of highly and poorly metastatic cell variants (6, 13-15). This approach was successfully applied for development of human prostate cancer cell variants with distinct metastatic potential (15). Experimental evidence indicates that enhancement of metastatic capability of human cancer cells transplanted orthotopically is associated with differential expression of several metastasis-associated genes that have been implicated earlier in certain key features of the metastatic phenotype (16). It is well established that even highly metastatic cells, when implanted ectopically, are not able to consistently produce metastasis.
Here we identified metastasis-associated gene expression signatures based on expression profiling human prostate carcinoma xenografts derived from the same highly metastatic variant implanted at orthotopic (metastasis promoting setting) and ectopic (metastasis suppressing setting) sites, demonstrating that distinct malignant behavior of highly metastatic cells associated with the site of inoculation in a nude mouse is dependent upon differential gene expression in prostate cancer cells implanted either orthotopically or ectopically. We utilized the Affymetrix GeneChip system to compare the expression profiles of 12,625 transcripts in highly metastatic variant PC-3MLN4 implanted at orthotopic (metastasis promoting setting) (“PC3MLN40R”) and ectopic (metastasis suppressing setting) (“PC3MLN4SC”) sites. PC-3MLN4 tumors growing in orthotopic metastasis-promoting setting appear to dramatically over-express a set of genes with well-established invasion-activation functions (
Expression profiling experiments imply that human prostate carcinoma cells growing in orthotopic metastasis-promoting setting display many clinically relevant gene expression features. Highly aggressive clinically relevant biological behavior of human prostate cancer cells growing in the prostate of nude mice is particularly evident in a fluorescent orthotopic bone metastasis model recapitulating to a significant degree the clinical pattern of metastatic spread of advanced prostate cancer in men (12). Recent gene expression analysis experiments showed that molecular signatures of metastasis could be identified in primary solid tumors (43). We sought to determine whether human prostate carcinoma xenografts growing in the prostate of nude mice would carry the clinically relevant gene-expression signatures of metastasis. We compared the gene expression profiles of 9 metastatic and 23 primary human prostate tumors (the original clinical data were published in LaTulippe, E., Satagopan, J., Smith, A., Scher, H., Scardino, P., Reuter, V., Gerald, W. L. Comprehensive gene expression analysis of prostate cancer reveals distinct transcriptional programs associated with metastatic disease. Cancer Res., 62: 4499-4506, 2002) to identify a broad spectrum of transcripts differentially regulated at the statistically significant level (p<0.05) in metastatic human prostate cancer. Next, we compared a set of transcripts differentially regulated in clinical metastatic human prostate tumors with transcripts differentially regulated in orthotopic human prostate carcinoma xenografts versus subcutaneous (“s.c.”)(i.e., ectopic) tumors of the same lineage. This comparison identified a set of 131 genes that exhibited highly concordant behavior in clinical metastatic samples and orthotopic metastasis-promoting tumors (Pearson correlation coefficient, r=0.799;
Remarkably, when we compared the expression profiles of these 131 transcripts in orthotopic xenografts and individual clinical samples, we found that all metastatic prostate carcinomas have expression patterns highly similar to orthotopic xenografts as reflected in positive correlation of expression profiles, whereas all primary tumors displayed a negative correlation of expression profiles (
Homo sapiens DNA for H4 histone, complete cds
Interestingly, the 9-gene molecular signature cluster (
To further test the potential clinical relevance of the models, we attempted to utilize expression profiling of highly metastatic orthotopic human prostate carcinoma xenografts for identification of gene expression correlates of clinically significant phenotypes such as invasive behavior and recurrence propensity of human prostate tumors (the original clinical data utilized in these examples were recently published in Singh, D., Febbo, P. G., Ross, K., Jackson, D. G., Manola, C. L., Tamayo, P., Renshaw, A. A., D'Amico, A. V., Richie, J. P., Lander, E. S., Loda, M., Kantoff, P. W., Golub, T. R., Sellers, W. R. Gene expression correlates of clinical prostate cancer behavior. Cancer Cell, 1: 203-209, 2002). Using gene expression profiles of metastasis-promoting orthotopic xenografts as a predictive reference of expected transcript abundance behavior in clinical samples, we identified a five-gene cluster (Table 43) of co-regulated transcripts discriminating with 75% accuracy invasive versus non-invasive human prostate tumors (
20 of 26 samples (77%) obtained from the patients with invasive prostate cancer defined by histology as having positive surgical margins (“PSM”) and/or extra-capsular penetration (“PCP”) exhibited a positive correlation coefficient of expression of the five-gene cluster (Table 43) compared to orthotopic xenografts. In contrast, 19 of 26 samples (73%) from the patients with organ-confined disease showed a negative correlation coefficient of expression of the five-gene cluster (Table 43) compared to orthotopic xenografts (
In this example we compared a set of transcripts differentially regulated in recurrent versus non-recurrent human prostate tumors with transcripts differentially regulated in orthotopic human prostate carcinoma xenografts derived from highly metastatic PC3MLN4 cell variant versus orthotopic tumors of the less metastatic parental lineages, PC3 and PC3M.
Next we compared a set of transcripts differentially regulated in recurrent versus non-recurrent human prostate tumors with transcripts differentially regulated in orthotopic human prostate carcinoma xenografts derived from highly metastatic PC3MLN4 cell variant versus subcutaneous (“s.c.”) ectopic tumors of the same lineage. This comparison identified a set of 25 genes (
When we compared the expression profiles of these 25 transcripts in orthotopic xenografts and individual clinical samples, we found that all recurrent prostate carcinomas have expression patterns highly similar to orthotopic xenografts as reflected in positive correlation of expression profiles, whereas 12 of 13 non-recurrent tumors displayed a negative correlation of expression profiles (
In conclusion, using gene expression profiles of metastasis-promoting orthotopic xenografts as a predictive reference of expected transcript abundance behavior in clinical samples, we identified clusters of co-regulated genes discriminating with 75-100% accuracy among metastatic versus primary, invasive versus non-invasive, and recurrent versus non-recurrent human prostate tumors. Our data indicate that human prostate cancer cells derived from metastatic lesions have stable “genetic memory” of metastatic behavior and that genetic signatures associated with metastatic phenotype could be revived by growth in a metastasis-promoting orthotopic environment. The genetic signatures of metastatic prostate cancer have the ability to be used as nucleic acid-based and/or protein-based clinical prognostic and diagnostic tests useful in clinical management of prostate cancer patients, and as a source of targets for novel therapeutic approaches for disease management.
Application of the present invention for identification of gene clusters with useful clinical properties was not limited by the availability of the suitable reference standard such as the appropriate cell lines and/or in vivo model systems. When a suitable reference standard was not readily available an algorithm utilizing the expression profile(s) of the best-fit sample(s) as a reference standard was applied for selection of the minimum segregation set of genes. As the first step of such analysis we compared the gene expression profiles of two distinct sets of samples that are subjects of classification (for example, metastatic and non-metastatic human breast tumors) to identify a broad spectrum of transcripts differentially regulated at a statistically significant level (p<0.05) in metastatic human breast cancer. If desirable, further criteria such as a particular cut-off based on fold expression changes (e.g., 2-fold, 3-fold, etc.) can be applied for selecting differentially expressed genes. Next, we calculated the average expression values for each transcript of the differentially expressed genes in the metastatic and non-metastatic tumors and determined the average fold expression change in metastatic versus non-metastatic tumors (“average” metastatic expression profile). We then determined the individual expression profiles for each sample within the two classification groups by calculating fold expression change for each transcript of the differentially expressed class of genes in a given sample by dividing an individual expression value of a gene by the average expression value for a particular gene across the entire data set. At the next step, we determined the individual phenotype association indices across the entire data set by calculating the Pearson correlation coefficient between the “average” metastatic expression profile and individual expression profiles. Next, the selection of the best-fit sample(s) was performed based on a highest positive and/or negative value(s) of the individual phenotype association index. The expression profile(s) of the best-fit sample(s) was utilized to refine the gene-expression signature associated with a particular phenotype to a small set of transcripts that would exhibit high discrimination accuracy between metastatic and non-metastatic tumors. To achieve this we used the increase in correlation coefficient of gene expression profiles between the “average” metastatic expression profile and an expression profile(s) of the best-fit sample(s) as a guide for reducing the number of members within a cluster.
The methods of the invention were used along with the data reported by Singh, et al. (2002) to identify gene clusters associated with an invasive phenotype. These data were the supplemental data reported in Singh, D., Febbo, P. G., et al., “Gene Expression Correlates of Clinical Prostate Cancer Behavior,” Cancer Cell March 2002 1:203-209, incorporated herein by reference. The clinical human prostate tumor samples were divided into two groups, invasive and non-invasive, as reported in Singh, et al. (2002). Invasive phenotype was assessed by determining the presence or absence of positive surgical margins (“PSM”) and positive or negative capsular penetration (“PCP”). The reference set was obtained following the procedures described above in part B, using the supplemental data reported in Singh, et al. (2002) for 26 invasive (identified as having positive surgical margins and/or positive capsular penetration) and 26 non-invasive (identified as having no evidence of positive surgical margins and/or positive capsular penetration) human prostate tumors. Thus, the first reference set was obtained by using the Affymetrix MicroDB (version 3.0) and Affymetrix Data Mining Tools (DMT) (version 3.0) data analysis software to identify genes that were differentially regulated in invasive group compared to non-invasive group of patients at the statistically significant level (p<0.05; Student T-test). Candidate genes were included in the first reference set if they were identified by the DMT software as having p values of 0.05 or less both for up-regulated and down-regulated genes. 114 genes were identified as being members of the reference set (Table 47).
Next, we calculated phenotype association indices for all 52 samples and determined that this gene cluster exhibited a 77% success rate in clinical sample classification based on individual phenotype association indices (Table 48). As shown in Table 48, 22/26 (or 85%) of the invasive prostate cancer samples had positive phenotype association indices, whereas 18/26 (or 69%) of non-invasive prostate cancer samples displayed negative phenotype association indices. Overall, 40 of 52 samples (or 77%) were correctly classified.
Next, we identified a single best-fit invasive prostate cancer sample displaying the correlation coefficient of 0.704 to the average expression profile of the 26 invasive prostate cancer samples. The expression profile of this single best-fit invasive prostate cancer sample was utilized as a second reference set.
The concordance set was obtained by selecting only those genes having a consistent direction of the differential expression in both the first and the second reference sets (i.e., greater gene expression difference in the invasive cf. the non-invasive samples and greater gene expression in the best-fit tumor sample cf. the average expression value across the entire data set or vice-versa). The concordance set comprised of 107 genes (r=0.721). A minimum segregation set was selected following the procedures described in above. Scatter plots were generated of the log10 transformed average −fold expression change in the first reference set and average −fold expression change in the second reference set (in case of a single best-fit tumor it was the log10 transformed ratio of the expression value for a gene to the average expression value across the entire data set). For the samples of the first reference set, <expression>1 corresponds to the average expression value for gene x over all samples from patients who had invasive tumors and <expression>2 corresponds to the average expression value for gene x over all samples from patients who had non-invasive tumors. A minimum segregation set was identified by selecting a subset of the highly correlated genes between two reference sets from the invasiveness concordance set. Using this approach we identified five gene clusters discriminating with high accuracy between invasive and non-invasive human prostate tumors. The members of these invasion predictors or invasion minimum segregation sets (invasion minimum segregation gene clusters) are listed in Tables 49-54. The classification performance for each of these gene clusters is presented in the Table 48.
In this example we utilized gene expression data and associated clinical information published in the recent study on gene expression profiling of breast cancer (van't Veer, L. J., et al., “Gene expression profiling predicts clinical outcome of breast cancer,” Nature, 415: 530-536, 2002, incorporated herein by reference). This study identifies 70 genes whose expression pattern is strongly predictive of a short post-diagnosis and treatment interval to distant metastases (van't Veer, L. J., et al., 2002). The expression pattern of these 70 genes discriminate with 81% (optimized sensitivity threshold) or 83% (optimal accuracy threshold) accuracy the patient's prognosis in the group of 78 young women diagnosed with sporadic lymph-node-negative breast cancer (this group comprises of 34 patients who developed distant metastases within 5 years and 44 patients who continued to be disease-free after a period of at least 5 years; they constitute a poor prognosis and good prognosis group, correspondingly). The authors described in this paper the second independent groups of breast cancer patients comprising 11 patients who developed distant metastases within 5 years and 8 patients who continued to be disease-free after a period of at least 5 years. We applied the method of the present invention to further reduce the number of genes whose expression patterns represent genetic signatures of breast cancer with “poor prognosis” or “good prognosis.” In our example we utilized the data derived from a group of 19 patients as a training set of samples, and the data derived from a group of 78 patients as a test set of samples.
Using the methods of present invention, we calculated the phenotype association indices for 19 samples of the training set and determined that this gene cluster exhibited a 84% success rate in clinical sample classification based on individual phenotype association indices (Table 54). As shown in Table 54, 7/8 (or 88%) of the good prognosis breast cancer samples had negative phenotype association indices, whereas 9/11 (or 82%) of poor prognosis breast cancer samples displayed negative phenotype association indices. Overall, 16 of 19 samples (or 84%) were correctly classified.
Next, we identified two best-fit poor prognosis breast cancer samples displaying the correlation coefficient of 0.751 and 0.832 to the average expression profile of the 11 poor prognosis breast cancer samples. The average expression profile of the 11 poor prognosis breast cancer samples was utilized as a first reference set. The average expression profile of these two best-fit poor prognosis breast cancer samples was utilized as a second reference set.
The concordance set was obtained by selecting only those genes having a consistent direction of the differential expression in both the first and the second reference sets (i.e., greater gene expression difference in the poor prognosis cf. the good prognosis samples and greater gene expression in the best-fit tumor sample cf. the average expression value across the entire data set or vice-versa). The concordance set comprised of 44 genes (r=0.950). A minimum segregation set was selected following the procedures described above. Scatter plots were generated of the log10 transformed average −fold expression change in the first reference set and average −fold expression change in the second reference set (in case of a single best-fit tumor it was the log10 transformed ratio of the expression value for a gene to the average expression value across the entire data set). For the samples of the first reference set, <expression>1 corresponds to the average expression value for gene x over all samples from patients who had invasive tumors and <expression>2 corresponds to the average expression value for gene x over all samples from patients who had non-invasive tumors. A minimum segregation set was identified by selecting a subset of the highly correlated genes between two reference sets from the concordance set. Using this approach we identified two gene clusters (19-gene cluster and 9-gene cluster) discriminating with high accuracy between poor prognosis and good prognosis human breast tumors in both training and test sets of clinical samples. These two breast cancer metastasis predictors or poor prognosis minimum segregation sets are listed in Tables 55 & 56. The classification performance for each of these gene clusters is presented in the Table 54.
In the next example, the average expression profile of all 19 breast cancer samples obtained from 11 patients with poor prognosis and 8 patients with good prognosis was utilized as a first reference set. Next, we calculated the individual phenotype association indices and identified a single best-fit poor prognosis breast cancer sample displaying the correlation coefficient of 0.677 to the average expression profile of the 19 breast cancer samples. The average expression profile of this single best-fit poor prognosis breast cancer sample was utilized as a second reference set.
The concordance set was obtained by selecting only those genes having a consistent direction of the differential expression in both the first and the second reference sets (i.e., greater gene expression difference in the poor prognosis cf. the good prognosis samples and greater gene expression in the best-fit tumor sample cf. the average expression value across the entire data set or vice-versa). The concordance set comprised of 47 genes (r=0.822). A minimum segregation set was selected following the procedures described in the introduction to the Detailed Description of the Preferred Embodiments and the Materials & Methods sections. Scatter plots were generated of the log10 transformed average −fold expression change in the first reference set and average −fold expression change in the second reference set (in case of a single best-fit tumor it was the log10 transformed ratio of the expression value for a gene to the average expression value across the entire data set). For the samples of the first reference set, <expression>1 corresponds to the average expression value for gene x over all samples from patients who had invasive tumors and <expression>2 corresponds to the average expression value for gene x over all samples from patients who had non-invasive tumors. A minimum segregation set was identified by selecting a subset of the highly correlated genes between two reference sets from the concordance set. Using this approach we identified two gene clusters (22-gene cluster and 12-gene cluster) discriminating with high accuracy between poor prognosis and good prognosis human breast tumors in both training and test sets of clinical samples. These two breast cancer metastasis predictors or poor prognosis minimum segregation sets are listed in Tables 57 & 58. The classification performance for each of these gene clusters is presented in the Table 54.
We applied the methods of the present invention to identify gene expression profiles distinguishing lung adenocarcinoma samples from normal lung specimens as well as highly malignant phenotype of lung adenocarcinoma, associated with short survival after diagnosis and therapy, from less aggressive lung cancers, associated with longer patient's survival. Clinical data set utilized in this example was published (Bhattacharjee, A., Richards, W. G., Staunton, J., Li, C., Monti, S., Vasa, P., Ladd, C., Beheshti, J., Bueno, R., Gillette, M., Loda, M., Weber, G., Mark, E. J., Lander, E. S., Wong, W., Johnson, B. E., Golub, T. R., Sugarbaker, D. J., Meyerson, M. Classification of human lung carcinomas by mRNA expression profiling reveals distinct adenocarcinoma subclasses. PNAS, 98: 13790-13795, 2001; incorporated herein by reference).
Using the clinical data set and associated clinical history (Bhattacharje et al., 2001), we selected two groups of adenocarcinoma patients having markedly distinct survival after diagnosis and therapy: poor prognosis group 1 comprising 34 patients with the median survival of 8.5 month (range 0.1-17.3 month) and good prognosis group 2 comprising 16 patients with the median survival of 84 month (range 75.4-106.1 month). As a starting point, we utilized a set of the 675 transcripts selected based on a statistical analysis of the quality of the dataset and variability of gene expression across the dataset (Bhattacharje et al., 2001). Applying methods of the present invention, we identified a set of 38 genes displaying at least a 2-fold difference in the average values of the mRNA expression levels between 34 poor prognosis samples versus 16 good prognosis samples (Table 59).
Next, we calculated the phenotype association indices for all 50 samples and determined that this gene cluster exhibited a 72% success rate in clinical sample classification based on individual phenotype association indices (Table 60). As shown in Table 60, 12/16 (or 75%) of the lung adenocarcinoma samples of the good prognosis group had negative phenotype association indices, whereas 24/34 (or 71%) of lung adenocarcinoma specimens of the poor prognosis group displayed positive phenotype association indices. Overall, 36 of 50 samples (or 72%) were correctly classified.
Next, we identified 8 best-fit poor prognosis samples displaying the correlation coefficient of 0.3 or higher to the average expression profile of the 34 poor prognosis samples. We calculated the average expression profile for these 8 best-fit poor prognosis samples by dividing the average expression value for each gene in the 8 samples of the best-fir set by the average expression value across the entire data set.
Next, we selected from an initial set of 38 genes a set of 26 genes (lung adenocarcinoma poor prognosis predictor cluster 1—see Table 61) displaying high positive correlation (r=0.938) between the best-fit tumors and poor prognosis samples data sets. This gene cluster exhibited a 56% success rate in clinical sample classification based on individual phenotype association indices (Table 60). As shown in Table 60, 15/16 (or 94%) of the lung adenocarcinoma samples of the good prognosis group had negative phenotype association indices, whereas 13/34 of lung adenocarcinoma specimens of the poor prognosis group displayed positive phenotype association indices. Overall, 28 of 50 samples (or 56%) were correctly classified.
To improve the classification accuracy, we selected from an initial set of 38 genes a set of 15 genes (lung adenocarcinoma poor prognosis predictor cluster 2—see Table 62) displaying high positive correlation (r=0.942) between the best-fit tumors and poor prognosis samples data sets.
This gene cluster exhibited a 78% success rate in clinical sample classification based on individual phenotype association indices (Table 60). As shown in Table 60, 11/16 (or 69%) of the lung adenocarcinoma samples of the good prognosis group had negative phenotype association indices, whereas 28/34 (or 82%) of lung adenocarcinoma specimens of the poor prognosis group displayed positive phenotype association indices. Overall, 39 of 50 samples (or 78%) were correctly classified.
The methods of the present invention were used along with the data reported by Ramaswamy et al. (2003) to identify gene clusters distinguishing between the human primary adenocarcinomas of diverse origin and metastatic adenocarcinoma lesions. These data were the supplemental data reported in Ramaswamy, S., Ross, K. N., Lander, E. S., Golub, T. R. “A molecular signature of metastasis in primary solid tumors,” Nature Genetics, January 2003, 33: 49-54, incorporated herein by reference. Ramaswamy et al. (2003) identified the 17-gene cluster expression profile of which distinguishes 12 metastatic adenocarcinoma nodules of diverse origin and 64 human primary adenocarcinomas of diverse origin (lung, breast, prostate, colorectal, uterus, ovary). Both metastatic lesions and primary adenocarcinomas were representing the same diverse spectrum of tumor types obtained from different individuals (Ramaswamy et al., 2003).
The expression profile of the 17-gene cluster in metastatic versus primary tumors was utilized as a first reference set.
Next, we calculated the phenotype association indices for all 76 samples and determined that this gene cluster exhibited a 45% success rate in clinical sample classification based on individual phenotype association indices (Table 63). As shown in Table 63, 12/12 (or 100%) of the metastatic samples had positive phenotype association indices, whereas 22/64 (or 34%) of primary tumor samples displayed negative phenotype association indices. Overall, 34 of 76 samples (or 45%) were correctly classified.
The classification accuracy of the 17-gene cluster was much improved when the discrimination threshold was set at the level of 0.400 of a correlation coefficient. As shown in Table 64, 12/12 (or 100%) of the metastatic samples had phenotype association indices higher than 0.400, whereas 48/64 (or 75%) of primary tumor samples displayed phenotype association indices lower than 0.400. Overall, 60 of 76 samples (or 79%) were correctly classified.
Next, we identified three best-fit metastatic samples displaying the correlation coefficient of 0.870, 0.923, and 0.874 to the average expression profile of the 12 metastatic samples. The average expression profile of these three best-fit metastatic samples was utilized as a second reference set.
The expression profile of the best-fit samples was utilized to refine the gene-expression signature associated with a metastatic phenotype to a small set of transcripts that would exhibit high discrimination accuracy between metastatic lesions and primary tumors. Thus, selecting a subset of the highly correlated genes between two reference sets identified a minimum segregation set suitable for clinical samples classification. Using this approach we identified four gene clusters discriminating with high accuracy between metastatic lesions and primary tumors. The members of these metastases minimum segregation sets (metastases minimum segregation gene clusters) are listed in Tables 65-68. The classification performance for each of these gene clusters is presented in the Tables 63 and 64.
Introduction
Critical clinical need in development of reliable prognostic markers suitable for stratification of prostate cancer patients is clearly demonstrated by the results of a recent randomized study of the therapeutic efficacy of surgery versus watch and wait strategy demonstrating only modest 6.6% absolute reduction in mortality after prostatectomy compared to observation, despite the association of surgery with a 50% reduction in hazard ration of death from prostate cancer (1). It appears that a measurable clinical benefit of surgery is limited to poorly defined sub-population of prostate cancer patients. Therefore, an improved ability to identify a sub-group of prostate cancer patients who would benefit from therapy should have a significant immediate positive clinical and socioeconomic impact.
Widely used biochemical, histopathological, and clinical criteria such as PSA level, Gleason score, the clinical tumor stage and molecular genetic approaches assaying loss of tumor suppressors or gain of oncogenes (2) had only limited success with respect to prostate cancer patients stratification and demonstrated a significant variability in predictive value among different clinical laboratories and hospitals. Furthermore, best existing markers cannot reliably identify at the time of diagnosis a poor prognosis group of prostate cancer patients that ultimately would fail therapy (3). Classification nomograms that incorporate measurements of several individual pre- and postoperative parameters are generally recognized as most efficient clinically useful models currently available for prediction of the probability of relapse-free survival after therapy of individual prostate cancer patients (4-7). However, one of the significant deficiencies of these classification systems is that they have only limited utility in predicting the differences in outcomes readily observed between patients diagnosed with prostate cancers exhibiting similar clinical, histopathological, and biochemical features. Therefore, a critical clinical need exists to improve the classification accuracy of prostate cancer patients with respect to clinical outcome after therapy.
Expression profiling of prostate tumor samples using oligonucleotide or cDNA microarray technology revealed gene expression signatures associated with human prostate cancer (8-19), including potential prostate cancer prognosis markers (9, 14, 16, 17). However, one of the major limitations of these studies was that the same clinical data set was utilized for both signature discovery and validation. Furthermore, usually only a single or few hits were validated using independent methods and independent clinical data sets, thus diminishing the potential advantage of the use of a panel of markers over a single marker in diagnostic and/or prognostic applications.
Here we applied a microarray-based gene expression profiling approach to identify molecular signatures distinguishing sub-groups of patients with differing outcome and develop a stratification algorithm demonstrating high discrimination accuracy between sub-groups of prostate cancer patients with distinct clinical outcome after therapy in a training set of 21 prostate cancer patients. To validate a potential clinical utility of discovered genetic signatures, we confirmed the discrimination power of proposed prostate cancer prognosis stratification algorithm using an independent set of 79 clinical tumor samples.
Our data indicate that identified molecular signatures provide the bases for developing clinical prognostic tests suitable for stratification of prostate cancer patients at the time of diagnosis with respect to likelihood of negative or positive clinical outcome after therapy. Our results provide experimental evidence of a transcriptional resemblance between metastatic human prostate carcinoma xenografts in nude mice and primary prostate tumors from patients subsequently developing relapse after therapy. These data suggest that genetically defined metastasis-promoting features of primary tumors are one of the major contributing factors of aggressive clinical behavior and unfavorable prognosis in prostate cancer patients.
Materials and Methods
Clinical Samples. We utilized in our experiments two independent sets of clinical samples for signature discovery (training outcome set of 21 samples) and validation (validation outcome set of 79 samples). Original gene expression profiles of the training set of 21 clinical samples analyzed in this study were recently reported (14). Primary gene expression data files of clinical samples as well as associated clinical information were provided by Dr. W. Sellers and can be found at http://www-genome.wi.mit.edu/cancer/.
Prostate tumor tissues comprising validation data set were obtained from 79 prostate cancer patients undergoing therapeutic or diagnostic procedures performed as part routine clinical management at MSKCC. Clinical and pathological features of 79 prostate cancer cases comprising validation outcome set are presented in the Table 70. Median follow-up after therapy in this cohort of patients was 70 months. Samples were snap-frozen in liquid nitrogen and stored at −80° C. Each sample was examined histologically using H&E-stained cryostat sections. Care was taken to remove normeoplastic tissues from tumor samples. Cells of interest were manually dissected from the frozen block, trimming away other tissues. All of the studies were conducted under MSKCC Institutional Review Board-approved protocols.
Cell Culture. Cell lines used in this study were previously described (19). The LNCap- and PC-3-derived cell lines were developed by consecutive serial orthotopic implantation, either from metastases to the lymph node (for the LN series), or reimplanted from the prostate (Pro series). This procedure generated cell variants with differing tumorigenicity, frequency and latency of regional lymph node metastasis (19). Except where noted, cell lines were grown in RPMI1640 supplemented with 10% FBS and gentamycin (Gibco BRL) to 70-80% confluence and subjected to serum starvation as described (19), or maintained in fresh complete media, supplemented with 10% FBS.
Orthotopic Xenografts. Orthotopic xenografts of human prostate PC-3 cells and sublines used in this study were developed by surgical orthotopic implantation as previously described (19). Briefly, 2×106 cultured PC3 cells, PC3M or PC3MLN4 sublines were injected subcutaneously into male athymic mice, and allowed to develop into firm palpable and visible tumors over the course of 2-4 weeks. Intact tissue was harvested from a single subcutaneous tumor and surgically implanted in the ventral lateral lobes of the prostate gland in a series of six athymic mice per cell line subtype. The mice were examined periodically for suprapubic masses, which appeared for all subline cell types, in the order PC3MLN4>PC3M>>PC3. Tumor-bearing mice were sacrificed by CO2 inhalation over dry ice and necropsy was carried out in a 2-4° C. cold room. Typically, bilaterally symmetric prostate gland tumors in the shape of greatly distended prostate glands were apparent. Prostate tumor tissue was excised and snap frozen in liquid nitrogen. The elapsed time from sacrifice to snap freezing was <5 min. A systematic gross and microscopic post mortem examination was carried out.
Tissue Processing for mRNA and RNA Isolation. Fresh frozen orthotopic tumor was examined by use of hematoxylin and eosin stained frozen sections. Orthotopic tumors of all sublines exhibited similar morphology consisting of sheets of monotonous closely packed tumor cells with little evidence of differentiation interrupted by only occasional zones of largely stromal components, vascular lakes, or lymphocytic infiltrates. Fragments of tumor judged free of these non-epithelial clusters were used for mRNA preparation. Frozen tissue (1-3 mm×1-3 mm) was submerged in liquid nitrogen in a ceramic mortar and ground to powder. The frozen tissue powder was dissolved and immediately processed for mRNA isolation using a Fast Tract kit for mRNA extraction (Invitrogen, Carlsbad, Calif., see above) according to the manufacturers instructions.
RNA and mRNA Extraction. For gene expression analysis, cells were harvested in lysis buffer 2 hrs after the last media change at 70-80% confluence and total RNA or mRNA was extracted using the RNeasy (Qiagen, Chatsworth, Calif.) or FastTract kits (Invitrogen, Carlsbad, Calif.). Cell lines were not split more than 5 times prior to RNA extraction, except where noted.
Affymetrix Arrays. The protocol for mRNA quality control and gene expression analysis was that recommended by Affymetrix (http://www.aff metrix.com). In brief, approximately one microgram of mRNA was reverse transcribed with an oligo(dT) primer that has a T7 RNA polymerase promoter at the 5′ end. Second strand synthesis was followed by cRNA production incorporating a biotinylated base. Hybridization to Affymetrix U95Av2 arrays representing 12,625 transcripts overnight for 16 h was followed by washing and labeling using a fluorescently labeled antibody. The arrays were read and data processed using Affymetrix equipment and software as reported previously (18, 19).
Data Analysis. Detailed protocols for data analysis and documentation of the sensitivity, reproducibility and other aspects of the quantitative statistical microarray analysis using Affymetrix technology have been reported (18, 19). 40-50% of the surveyed genes were called present by the Affymetrix Microarray Suite 5.0 software in these experiments. The concordance analysis of differential gene expression across the data sets was performed using Affymetrix MicroDB v. 3.0 and DMT v.3.0 software as described earlier (18, 19). We processed the microarray data using the Affymetrix Microarray Suite v.5.0 software and performed statistical analysis of expression data set using the Affymetrix MicroDB and Affymetrix DMT software. This analysis identified a set of 218 genes (91 up-regulated and 127 down-regulated transcripts) differentially regulated in tumors from patients with recurrent versus non-recurrent prostate cancer at the statistically significant level (p<0.05) defined by both T-test and Mann-Whitney test (Table 69). The concordance analysis of differential gene expression across the clinical and experimental data sets was performed using Affymetrix MicroDB v. 3.0 and DMT v.3.0 software as described earlier (19). The Pearson correlation coefficient for individual test samples and appropriate reference standard was determined using the Microsoft Excel software as described in the signature discovery protocol.
Survival Analysis. The Kaplan-Meier survival analysis was carried out using the Prism 4.0 software. Statistical significance of the difference between the survival curves for different groups of patients was assessed using Chi square and Logrank tests.
Discovery and validation of the prostate cancer recurrence predictor algorithm. According to the present invention, clinically relevant genetic signatures can be found by searching for clusters of co-regulated genes that display highly concordant transcript abundance behavior across multiple experimental models and clinical settings that model or represent malignant phenotypes of interest (Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003; Example 5, supra; Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B. Malignancy-associated regions of transcriptional activation: gene expression profiling identifies common chromosomal regions of a recurrent transcriptional activation in human prostate, breast, ovarian, and colon cancers. Neoplasia, 5: 21-228; Glinsky, G. V., Ivanova, Y. A., Glinskii, A. B. Common malignancy-associated regions of transcriptional activation (MARTA) in human prostate, breast, ovarian, and colon cancers are targets for DNA amplification. Cancer Letters, in press, 2003). Thus, a primary criterion in selecting genes for inclusion within the cluster is the concordance of changes in expression rather than a magnitude of changes (e.g., fold change). Accordingly, transcripts of interest are expected to have a tightly controlled “rank order” of expression within a cluster of co-regulated genes reflecting a balance of up- and down-regulation as a desired regulatory end-point in a cell. A degree of resemblance of the transcript abundance rank order within a gene cluster between a test sample and reference standard is measured by a Pearson correlation coefficient and designated as a phenotype association index (PAI), as described fully in the introduction of the Detailed Description of Preferred Embodiments section. To identify genes with consistently concordant expression patterns across multiple data sets and various experimental conditions, we compared the expression profile of 218 genes (test samples) to the expression profiles of transcripts differentially regulated in multiple experimental models (reference standard) of human prostate cancer (Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003).
The transcripts comprising each signature were selected based on Pearson correlation coefficients (r>0.95) reflecting a degree of similarity of expression profiles in clinical tumor samples (recurrent versus non-recurrent tumors) and experimental samples using the following protocol.
Step 1. Sets of differentially regulated transcripts were independently identified for each experimental conditions (see below) and clinical samples using the Affymetrix microarray processing and statistical analysis software package as described in this examples's Materials and Methods section.
Step 2. Sub-sets of transcripts exhibiting concordant expression changes in clinical and experimental samples were identified using the Affymetrix MicroDB and DMT software. Sub-sets of transcripts were identified with concordant changes of transcript abundance behavior in recurrent versus non-recurrent clinical tumor samples (218 transcripts) and experimental conditions independently defined for each signature (Signature 1: PC-3MLN4 orthotopic versus s.c. xenografts; Signature 2: PC-3MLN4 versus PC-3M & PC-3 orthotopic xenografts; Signature 3: PC-3/LNCap consensus class, Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003). Thus, from a set of 218 transcripts three concordant sub-sets of transcripts were identified corresponding to each binary comparison of clinical and experimental samples.
Step 3. Small gene clusters were selected as sub-sets of genes exhibiting concordant changes of transcript abundance behavior in recurrent versus non-recurrent clinical tumor samples (218 transcripts) and experimental conditions defined for each signature (Signature 1: PC-3MLN4 orthotopic versus s.c. xenografts; Signature 2: PC-3MLN4 versus PC-3M & PC-3 orthotopic xenografts; Signature 3: PC-3/LNCap consensus class, Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003). Expression profiles were presented as log10 average fold changes for each transcript and processed for visualization and Pearson correlation analysis using Microsoft Excel software. The cut-off criterion for cluster formation was set to exceed a Pearson correlation coefficient 0.95 among the log10 transformed average expression values in the compared groups.
Step 4. Small gene clusters exhibiting highly concordant pattern of expression (Pearson correlation coefficient, r>0.95) in clinical and experimental samples (identified in step 3) were evaluated for their ability to discriminate clinical samples with distinct outcomes after the therapy. To assess a potential prognostic relevance of individual gene clusters, we calculated a Pearson correlation coefficient for each of 21 tumor samples (training data set) by comparing the expression profiles of individual samples to the reference expression profiles of relevant experimental samples defined for each signature and an “average” expression profile of recurrent versus non-recurrent tumors. As explained above, we named the corresponding correlation coefficients calculated for individual samples the phenotype association indices (PAIs). We evaluated the prognostic power of identified clusters of co-regulated transcripts based on their ability to segregate the patients with recurrent and non-recurrent prostate tumors into distinct sub-groups and selected a single best performing cluster for each binary condition (
Step 5. We used Kaplan-Meier survival analysis to assess the prognostic power of each best-performing cluster in predicting the probability that patients would remain disease-free after therapy (
Step 6. We developed a prostate cancer recurrence predictor algorithm taking into account calls from all three individual signatures. We selected the common prognosis discrimination cut-off value for all three signatures based on highest level of statistical significance in patient's stratification into poor and good prognosis groups as determined by Kaplan-Meier survival analysis (lowest P value and highest hazard ratio defined by the log-rank test; Table 70 &
Step 7. We validated the prognostic power of prostate cancer recurrence predictor algorithm alone and in combination with the established markers of outcome using an independent clinical set of 79 prostate cancer patients (
Results
Identification of molecular signatures distinguishing sub-groups of prostate cancer patients with distinct clinical outcomes after therapy. To identify the outcome predictor signatures, we utilized as a training data set the expression analysis of 12,625 transcripts in 21 prostate tumor samples obtained from prostate cancer patients with distinct clinical outcomes after therapy. Using biochemical evidence of relapse after therapy as a criterion of treatment failure, 21 patients were divided into two sub-groups, representing prostate cancer with recurrent (8 patients) and non-recurrent (13 patients) clinical behavior (14). We processed the original U95Av2 GeneChip CEL files using the Affymetrix Microarray Suite 5.0 software and performed statistical analysis of expression data set using the Affymetrix MicroDB and Affymetrix DMT software. This analysis identified a set of 218 genes (91 up-regulated and 127 down-regulated transcripts) differentially regulated in tumors from patients with recurrent versus non-recurrent prostate cancer at the statistically significant level (p<0.05) defined by both T-test and Mann-Whitney test (Table 70).
To reduce the number of hits in potential outcome predictor clusters and identify transcripts of potential biological relevance, we compared the expression profile of 218 genes to the expression profiles of transcripts differentially regulated in multiple experimental models of human prostate cancer (Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003, and Example 5, supra) in search for genes with consistently concordant expression patterns across multiple data sets and various experimental conditions. We identified several small gene clusters exhibiting highly concordant pattern of expression (Pearson correlation coefficient, r>0.95) in clinical and experimental samples. We evaluated the prognostic power of each identified cluster of co-regulated transcripts based on ability to segregate the patients with recurrent and non-recurrent prostate tumors into distinct sub-groups. To assess a potential prognostic relevance of individual gene clusters, we calculated a Pearson correlation coefficient for each of 21 tumor samples by comparing the expression profiles of individual samples to the “average” expression profile of recurrent versus non-recurrent tumors and expression profiles of relevant experimental samples (Table 69 and
Using this strategy we identified several gene clusters (Tables 69 & 70) discriminating with 86-95% accuracy human prostate tumors exhibiting recurrent or non-recurrent clinical behavior (
Table 70 illustrates data from 21 prostate cancer patients who provided tumor samples comprising a signature discovery (training) data set that were classified according to whether they had a good-prognosis signature or poor-prognosis signature based on PAI values defined by either individual recurrence predictor signatures or a recurrence predictor algorithm that takes into account calls from all three signatures. The number of correct predictions in the poor-prognosis and good-prognosis groups is shown as a fraction of patients with the observed clinical outcome after therapy (8 patients developed relapse and 13 patients remained disease-free). Correlation coefficients reflect a degree of similarity of expression profiles in clinical tumor samples (recurrent versus non-recurrent tumors) and experimental samples (Signature 1: PC-3MLN4 orthotopic versus s.c. xenografts; Signature 2: PC-3MLN4 versus PC-3M & PC-3 orthotopic xenografts; Signature 3: PC-3/LNCap consensus class, Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003; and Example 5, supra). P values were calculated with use of the log-rank test and reflect the statistically significant difference in the probability that patients would remain disease-free between poor-prognosis and good-prognosis sub-groups.
The figure shows the phenotype association indices for eight samples from patients who later had recurrence as bars 1 through 8, while the association indices for thirteen samples from patients whose tumors did not recur is shown as bars 11 through 23. Eight of the eight samples (or 100%) from patients who later experienced recurrence had positive phenotype association indices and so were properly classified. Twelve of the thirteen samples (or 92.3%) from patients whose tumors did not recur had negative phenotype association indices and so were properly classified as non-recurrent tumors. Thus, overall, twenty of the twenty-one samples (or 95.2%) were properly classified using a five-gene recurrence predictor signature. Two alternative clusters identified using this strategy showed similar sample classification performance (Tables 69 & 70).
To further evaluate the prognostic power of the identified gene expression signatures, we performed Kaplan-Meier survival analysis using as a clinical end-point disease-free interval (“DFI”) after therapy in prostate cancer patients with positive and negative PAIs. The Kaplan-Meier survival curves showed a highly significant difference in the probability that prostate cancer patients would remain disease-free after therapy between the groups with positive and negative PAIs defined by the signatures (FIGS. 58A-C), suggesting that patients with positive PAIs exhibit a poor outcome signature whereas patients with negative PAIs manifest a good outcome signature. The estimated hazard ration for disease recurrence after therapy in the group of patients with positive PAIs as compared with the group of patients with negative PAIs defined by the recurrence predictor signature 3 (Table 69) was 9.046 (
Theoretically, the recurrence predictor algorithm based on a combination of signatures should be more robust than a single predictor signature, particularly during the validation analysis using an independent test cohort of patients. Next we analyzed whether a combination of the three signatures would perform in the patient's classification test with similar accuracy as the individual signatures. We found that the cut-off value of PAIs>0.2 scored in two of three individual clusters allowed to achieve the 90% recurrence prediction accuracy (Table 70). This recurrence predictor algorithm correctly identified 88% of patients with recurrent and 92% of patients with non-recurrent disease (Table 70). The Kaplan-Meier survival analysis (
Validation of the outcome predictor signatures using independent clinical data set. To validate the potential clinical utility of identified molecular signatures, we evaluated the prognostic power of signatures applied to an independent set of 79 clinical samples obtained from 37 prostate cancer patients who developed recurrence after the therapy and 42 patients who remained disease-free. The Kaplan-Meier survival analysis demonstrated that all three recurrence predictor signatures (Table 69) segregate prostate cancer patients into sub-groups with statistically significant differences in the probability of remaining relapse-free after therapy (Table 71). Interestingly, application of the recurrence predictor algorithm (requiring a cut-off value of PAIs>0.2 scored in two of three individual clusters) appears to perform better than individual signatures in patient's stratification test using an independent data set (Table 71).
Table 71 summarizes classification of 79 prostate cancer patients who provided tumor samples. These samples comprise a signature validation (test) data set and were classified according to whether they had a good-prognosis signature or poor-prognosis signature based on PAI values defined by either individual recurrence predictor signatures or recurrence predictor algorithm that takes into account calls from all three signatures. Kaplan-Meier analysis was performed to evaluate the probability that patients would remain disease free according to whether they had a poor-prognosis or a good-prognosis signature and determine the proportion of patients who would remain disease-free at least 5 years after therapy in a poor-prognosis and a good-prognosis sub-groups. Hazard ratios, 95% confidence intervals, and P values were calculated with use of the log-rank test.
Kaplan-Meier survival analysis (
Recurrence predictor signatures provide additional predictive value over conventional markers of outcome. Next we determined that application of the recurrence predictor signatures provides additional predictive value when combined with conventional markers of outcome such as preoperative PSA level and Gleason score. Both preoperative PSA level and RP Gleason sum were significant predictors of prostate cancer recurrence after therapy in the validation cohort of 79 patients (
Kaplan-Meier survival analysis (
Table 72 shows the number of correct predictions in poor-prognosis and good-prognosis groups as a fraction of patients with the observed clinical outcome after therapy (37 patients developed relapse and 42 patients remained disease-free). PSA and Gleason sum cut-off values for segregation of poor-prognosis and good-prognosis sub-groups were defined to achieve the most accurate and statistically significant recurrence prediction in this cohort of patients. Multiparameter nomogram-based prognosis predictor was defined as described in this example's Materials & Methods using 50% relapse-free survival probability as a cut-off for patient's stratification into poor and good prognosis subgroups.
We next determined that application of the recurrence predictor algorithm identifies sub-groups of patients with distinct clinical outcome after therapy in both high and low PSA-expressing groups, thus adding additional predictive value to the therapy outcome classification based on preoperative PSA level alone.
In the group of patients with high preoperative PSA level (
In the group of patients with low preoperative PSA level (
Radical prostatectomy (“RP”) Gleason sum is a significant predictor of relapse-free survival in the validation cohort of 79 prostate cancer patients (
In the group of patients with RP Gleason sum 6 & 7 (
In the group of patients with RP Gleason sum 8 & 9 (
Recurrence predictor signatures provide additional predictive value over outcome prediction based on multiparameter nomogram. Classification nomograms are generally recognized most efficient clinically useful models currently available for prediction of the probability of relapse-free survival after therapy of individual prostate cancer patients (Kattan M. W., Eastham J. A., Stapleton A. M., Wheeler T. M., Scardino P. T. A preoperative nomogram for disease recurrence following radical prostatectomy for prostate cancer. J. Natl. Cancer Inst., 90: 766-771, 1998; D'Amico A. V., Whittington R., Malkowicz S. B., Fondurulia J., Chen M-H, Kaplan I., Beard C. J., Tomaszewski J. E., Renshaw A. A., Wein A., Coleman C. N. Pretreatment nomogram for prostate-specific antigen recurrence after radical prostatectomy or external-beam radiation therapy for clinically localised prostate cancer. J. Clin. Oncol., 17: 168-172, 1999; Graefen M., Noldus J., Pichlmeier A., Haese P., Hammerer S., Fernandez S., Conrad R., Henke E., Huland E., Huland H. Early prostate-specific antigen relapse after radical retropubic prostatectomy: prediction on the basis of preoperative and postoperative tumor characteristics. Eur. Urol., 36: 21-30, 1999; Kattan M. W., Wheeler T. M., Scardino P. T. Postoperative nomogram for disease recurrence after radical prostatectomy for prostate cancer. J. Clin. Oncol., 17: 1499-1507, 1999.). We applied the Kattan nomogram utilizing multiple postoperative parameters (Kattan, et al. (1999)) for prognosis prediction classification in the test group of 79 prostate cancer patients.
Kaplan-Meier survival analysis (
Application of the recurrence predictor algorithm identified sub-groups of patients with distinct clinical outcome after therapy in both poor and good prognosis groups defined by the Kattan nomogram, thus adding additional predictive value to the therapy outcome classification based on nomogram alone.
In the poor prognosis group of patients defined by the Kattan nomogram the application of the recurrence predictor algorithm appears to identify two sub-groups of patients with statistically significant difference in the probability to remain relapse-free after therapy (
Similarly, in the good prognosis group of patients identified based on application of the Kattan nomogram, the recurrence predictor algorithm seems to define two sub-groups of patients with statistically significant difference in the probability to remain relapse-free after therapy (
Recurrence predictor algorithm defines poor and good prognosis sub-groups of patients diagnosed with the early stage prostate cancer. Identification of sub-groups of patients with distinct clinical outcome after therapy would be particularly desirable in a cohort of patients diagnosed with the early stage prostate cancer. Next we determined that recurrence predictor signatures are useful in defining sub-groups of patients diagnosed with early stage prostate cancer and having a statistically significant difference in the likelihood of disease relapse after therapy.
In the group of patients diagnosed with the stage 1C prostate cancer (
In the group of patients diagnosed with the stage 2A prostate cancer (
2. Discussion
As a result of the broad application of measurements of PSA level in the blood for early detection of prostate cancer in the United States, an increasing proportion of prostate cancer patients are diagnosed with early-stage tumors that apparently confined to the prostate gland and many patients have seemingly indolent disease not affecting individual's survival (Potosky, A., Feuer, E., Levin, D. Impact of screening on incidence and mortality of prostate cancer in the United States. Epidemiol. Rev., 23: 181-186, 2001). The considerable clinical heterogeneity of the early stage prostate cancer represents a highly significant health care and socioeconomic challenge because prostate cancer is expected to be diagnosed in ˜200,000 individuals every year (Greenlee, R. T., Hill-Hamon, M. B., Murray, T., Thun, M. Cancer statistics, 2001. CA Cancer J. Clin., 51: 15-36, 2001). Consequently, it can be argued that, unlike other types of cancer, development of efficient prognostic tests rather than early detection is critical for improvement of clinical decision-making and management of prostate cancer.
We hypothesized that clinically relevant genetic signatures can be found by searching for clusters of co-regulated genes that display highly concordant transcript abundance behavior across multiple experimental models and clinical settings that model or represent malignant phenotypes of interest (Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B., Gebauer, G. Microarray analysis of xenograft-derived cancer cell lines representing multiple experimental models of human prostate cancer. Molecular Carcinogenesis, 37: 209-221, 2003; Glinsky, G. V., Krones-Herzig, A., Glinskii, A. B. Malignancy-associated regions of transcriptional activation: gene expression profiling identifies common chromosomal regions of a recurrent transcriptional activation in human prostate, breast, ovarian, and colon cancers. Neoplasia, 5: 21-228; Glinsky, G. V., Ivanova, Y. A., Glinskii, A. B. Common malignancy-associated regions of transcriptional activation (MARTA) in human prostate, breast, ovarian, and colon cancers are targets for DNA amplification. Cancer Letters, in press, 2003). Thus, according to this model the primary criterion in a transcript selection process should be the concordance of changes in expression rather the magnitude of changes (e.g., fold change). One of the predictions of this model is that transcripts of interest are expected to have a tightly controlled “rank order” of expression within a cluster of co-regulated genes reflecting a balance of up- and down-regulated mRNAs as a desired regulatory end-point in a cell. A degree of resemblance of the transcript abundance rank order within a gene cluster between a test sample and reference standard is measured by a Pearson correlation coefficient and designated a phenotype association index (“PAI”).
Using this strategy we discovered and validated a prostate cancer recurrence predictor algorithm that is suitable for stratifying patients at the time of diagnosis into poor and good prognosis sub-groups with statistically significant differences in the disease-free survival after therapy. The algorithm is based on application of gene expression signatures associated with biochemical recurrence of prostate cancer. The signatures (Table 69) were defined using clusters of co-regulated genes exhibiting highly concordant expression profiles (r>0.95) in metastatic nude mouse models of human prostate carcinoma and tumor samples from patients with recurrent prostate cancer (see Example 5).
A few previous studies have applied oligonucleotide or cDNA microarrays for identification of gene expression signatures associated with biochemical recurrence of human prostate cancer (Dhanasekaran, S. M., Barrette, T. R., Ghosh, D., Shah, R., Varambally, S., Kurachi, K., Pienta, K. J., Rubin, M. A., Chinnalyan, A. M. Delineation of prognostic biomarkers in prostate cancer. Nature, 412:822-826, 2001; Singh, D., Febbo, P. G., Ross, K., Jackson, D. G., Manola, C. L., Tamayo, P., Renshaw, A. A., D'Amico, A. V., Richie, J. P., Lander, E. S., Loda, M., Kantoff, P. W., Golub, T. R., Sellers, W. R. Gene expression correlates of clinical prostate cancer behavior. Cancer Cell, 1: 203-209, 2002; Varambally, S., Dhanasekaran, S. M., Zhou, M., Barrette, T. R., Kumar-Sinha, C., Sanda, M. G., Ghosh, D., Pineta, K. J., Sewalt, R. G., Otte, A. P., Rubin, M. A., Chinnalyan, A. M. The polycomb group protein EZH2 is involved in progression of prostate cancer. Nature, 419: 624-629, 2002; Henshall, S. M., Afar, D. E., Hiller, J., Horvath, L. G., Quinn, D. I., Rasiah, K. K., Gish, K., Willhite, D., Kench, J. G., Gardiner-Garden, M., Stricker, P. D., Scher, H. I., Grygiel, J. J., Agus, D. B., Mack, D. H., Sutherland, R. L. Survival analysis of genome-wide gene expression profiles of prostate cancers identifies new prognostic targets of disease relapse. Cancer Res., 63: 4196-4203, 2003). One of the major deficiencies of these studies that somewhat limited their significance was that a single clinical data set was utilized for both signature discovery and validation. To our knowledge, the work reported here is the first genome-wide expression profiling study of human prostate cancer that utilizes one clinical data set for signature discovery and algorithm development, and a second independent data set for validation of the prostate cancer recurrence predictor algorithm.
One of the interesting features of described here prostate cancer recurrence predictor algorithm is that it provides additional predictive value over conventional markers of outcome such as pre-operative PSA level and Gleason sum. Another important feature of identified recurrence predictor algorithm is its ability to stratify patients diagnosed with the early stage prostate cancer into sub-groups with statistically-distinct likelihoods of biochemical relapse after therapy. Importantly, the recurrence predictor algorithm segregates into poor prognosis group 88% of patients who subsequently developed disease recurrence within one year after prostatectomy. Based on this analysis we concluded that identified in this study genetic signatures (as well as others that can be determined using the methods of the invention) have a significant potential for developing highly accurate clinical prognostic tests suitable for stratifying prostate cancer patients at the time of diagnosis with respect to likelihood of negative or positive clinical outcome after therapy.
The causal genetic, molecular, and biological distinctions between prostate tumors with recurrent and indolent clinical behavior remain largely unknown. The results reported in this example and in Example 5 provide the first experimental evidence of a clinically relevant transcriptional resemblance between metastatic human prostate carcinoma xenografts growing orthotopically in nude mice and primary prostate tumors from patients that subsequently developed a biochemical relapse after therapy. This work provides a model for investigation of the potential functional relevance of identified transcriptional aberrations and suggests that genetically defined metastasis-promoting features of primary tumors seem to be one of the major contributing factors of aggressive clinical behavior and unfavorable prognosis in prostate cancer patients. This conclusion is consistent with results of the several recent studies aimed at definition of metastasis predictor signatures in the primary human tumors representing multiple types of epithelial cancers (van 't Veer, L. J., Dai, H., van de Vijver, M. J., et al. Gene expression profiling predicts clinical outcome of breast cancer. Nature, 415: 530-536, 2002; van de Vijver, M. J., He, Y. D., van 't Veer, L. J., et al. A gene expression signature as a predictor of survival in breast cancer. N. Engl. J. Med., 347: 1999-2009, 2002; Ramaswamy, S., Ross, K. N., Lander, E. S., Golub, T. R. A molecular signature of metastasis in primary solid tumors. Nature Genetics, 33: 49-54, 2003). Our results indicate that sub-groups of prostate cancer patients with poor and good prognosis gene expression signatures reflect the presence of two genetically defined sub-types of human prostate carcinoma manifesting dramatic statistically significant differences in response to therapy and clinically distinct courses of disease progression.
One of the dominant views on prostate cancer pathogenesis is the concept of progression from hormone-dependent early stage prostate cancer to hormone-refractory metastatic late stage disease with the apparent implication of increased proportion of patients with poor prognosis at the advanced stage of progression. However, in our validation data set of 79 samples the actual frequency of recurrence remains relatively constant among the patients with different stages of prostate cancer: 47% (16 of 34) in stage 1C; 56% (9 of 16) in stage 2A; and 41% (12 of 29) in stages 2B/2C/3A. These data suggest that progression of the disease occurs only in a sub-group of patients. Interestingly, in a sub-group of patients with good prognosis signatures the frequency of recurrence appears to increase in the patients with the late-stage prostate cancer: 24% (5 of 21) in stage 1C; 22% (2 of 9) in stage 2A; 33% (3 of 9) in stage 2B; 40% (2 of 5) in stage 2C/3A. These results seem to imply that patients with the good prognosis signatures may represent a sub-group undergoing a classical prostate cancer progression with a gradual increase in malignant potential. The patients with poor prognosis signatures may represent a genetically and biologically distinct sub-type of prostate cancer exhibiting highly malignant behavior at the early stage of disease with the frequency of recurrence 85% (11 of 13) in stage 1C and 100% (7 of 7) in stage 2A patients.
In summary, using expression profiles of highly metastatic models of human prostate cancer in nude mice as a predictive reference of expected transcript abundance behavior in recurrent prostate tumors, we identified and validated recurrence predictor signatures of human prostate cancer. Prostate cancer recurrence predictor signatures provide additional predictive value to the conventional markers of outcome and will be clinically useful in stratifying prostate cancer patients into sub-groups with distinct clinical manifestation of disease and different response to therapy.
Introduction
Highly accurate prognostic tests are essential for individualized decision-making process during clinical management of cancer patients leading to rational and more efficient selection of appropriate therapeutic interventions and improved outcome after therapy. In breast cancer, patients are classified into broad subgroups with poor and good prognosis reflecting a different probability of disease recurrence and survival after therapy. Distinct prognostic subgroups are identified using a combination of clinical and pathological criteria: age, primary tumor size, status of axillary lymph nodes, histologic type and pathologic grade of tumor, and hormone receptor status (Goldhirsch, A., Glick, J. H., Gelber, R. D., Coates, A. S., Seen, H. J. Meeting highlights: International Consensus Panel on the Treatment of Primary Breast Cancer: Seventh International Conference on Adjuvant Therapy of Primary Breast Cancer. J. Clin. Oncol., 19: 3817-3827, 2001; Eifel, P., Axelson, J. A., Costa, J., et al. National Institute of Health Consensus Development Conference Summary: adjuvant therapy for breast cancer, Nov. 1-3, 2000. J. Natl. Cancer Inst., 93: 979-989, 2001.)
One of the most critical treatment decisions during the clinical management of breast cancer patients is the use of adjuvant systemic therapy. Adjuvant systemic therapy significantly improves disease-free and overall survival in breast cancer patients with both lymph-node negative and lymph-node positive disease (Early Breast Cancer Trialists' Collaborative Group. Polychemotherapy for early breast cancer: an overview of the randomized trials. Lancet, 352: 930-942, 1998; Early Breast Cancer Trialists' Collaborative Group. Tamoxifen for early breast cancer: an overview of the randomized trials. Lancet, 351: 1451-1467, 1998). It is generally accepted that breast cancer patients with poor prognosis would gain the most benefits from the adjuvant systemic therapy (Goldhirsch, et al., 2001; Eifel et al., 2001).
Diagnosis of lymph-node status is important in therapeutic decision-making, prediction of disease outcome, and probability of breast cancer recurrence. Invasion into axillary lymph nodes is recognized as one of the most important prognostic factors (Krag, D., Weaver, D., Ashikaga, T., et al. The sentinel node in breast cancer—a multicenter validation study. N. Engl. J. Med., 339: 941-946, 1998; Singletary, S. E., Allred, C., Ashley, P., et al. Revision of the American Joint Committee on cancer staging system for breast cancer. J. Clin. Oncol., 20: 3628-3636, 2002; Jatoli, I., Hilsenbeck, S. G., Clark, G. M., Osborne, C. K. Significance of axillary lymph node metastasis in primary breast cancer. J. Clin. Oncol., 17: 2334-2340, 1999). Most patients diagnosed with lymph-node negative breast cancer can be effectively treated with surgery and local radiation therapy. However, results of several studies show that 22-33% of breast cancer patients with no detectable lymph-node involvement and classified into a good prognosis subgroup develop recurrence of disease after a 10-year follow-up (Early Breast Cancer Trialists' Collaborative Group. Tamoxifen for early breast cancer: an overview of the randomized trials. Lancet, 351: 1451-1467, 1998). Therefore, accurate identification of breast cancer patients with lymph-node negative tumors who are at high risk of recurrence is critically important for rational treatment decision and improved clinical outcome in the individual patient.
Microarray-based gene expression profiling of human cancers rapidly emerged as a new powerful screening technique generating hundreds of novel diagnostic, prognostic, and therapeutic targets (Golub, T. R., Slonim, D. K., Tamayo, P., et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science, 286: 531-537, 1999; Alizadeh, A. A., Eisen, M. B., Davis, R. E., et al. Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature, 403: 503-511, 2000; Alizadeh, A. A., Ross, D. T., Perou, C. M., van de Rijn, M. Towards a novel classification of human malignancies based on gene expression patterns. J. Pathol., 195: 41-52, 2001; Battacharjee, A., Richards, W. G., Staunton, J., et al. Classification of human lung carcinomas by mRNA expression profiling reveals distinct adenocarcinoma subclasses. Proc. Natl. Acad. Sci. USA, 98: 13790-13795, 2001; Yeoh, E.-J., Ross, M. E., Shurtleff, S. A., et al. Classification, subtype discovery, and prediction of outcome in pediatric acute lymphoblastic leukemia by gene expression profiling. Cancer Cell, 1: 133-143, 2002; Dyrskot, L., Thykjaer, T., Kruhoffer, M., Jensen, J. L., Marcussen, N., Hamilton-Dutoit, S., Wolf, H., Orntoft, T. Identifying distinct classes of bladder carcinoma using microarrays. Nature Genetics, 33: 90-96, 2003). Recently breast cancer gene expression signatures have been identified that are associated with the estrogen receptor and lymph node status of patients and can aid in classification of breast caner patients into subgroups with different clinical outcome after therapy (Perou, C. M., Sorlie, T., Eisen, M. B., et al. Molecular portrait of human breast tumors. Nature, 406: 747-752, 2000; Gruvberger, S., Ringner, M., Chen, Y., et al. Estrogen receptor status in breast cancer is associated with remarkably distinct gene expression patterns. Cancer Res., 61: 5979-5984, 2001; West, M., Blanchette, C., Dressman, H., et al. Predicting the clinical status of human breast cancer by using gene expression profiles. Proc. Natl. Acad. Sci. USA, 98: 11462-11467, 2001; Ahr, A., Karn, T., Sollbach, C., et al. Identification of high risk breast cancer patients by gene expression profiling. Lancet, 359: 131-132, 2002; van 't Veer, L. J., dai, H., van de Vijver, M. J., et al. Gene expression profiling predicts clinical outcome of breast cancer. Nature, 415: 530-536, 2002; Sorlie, T., Perou, C. M., Tibshirani, R., et al. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc. Natl. Acad. Sci. USA, 98: 10869-10874, 2001; Heedenfalk, I., Duggan, D., Chen, Y., et al. Gene-expression profiles in hereditary breast cancer. N. Engl. J. Med., 344: 539-548, 2001; van de Vijver, M. J., He, Y. D., van 't Veer, L. J., et al. A gene expression signature as a predictor of survival in breast cancer. N. Engl. J. Med., 347: 1999-2009, 2002; Huang, E., Cheng, S. H., Dressman, H., Pittman, J., Tsou, M. H., Horng, C. F., Bild, A., Iversen, E. S., Liao, M., Chen, C. M., West, M., Nevins, J. R., Huang, A. T. Gene expression predictors of breast cancer outcome. Lancet, 361: 1590-1596, 2003).
One of the significant limitations of these array-based studies is that they generated vast data sets comprising many attractive targets with diagnostic and prognostic potential. Design and performance of meaningful follow-up experiments such as translation of the array-generated hits into quantitative RT-PCR-based analytical assays would require a significant data reduction. Furthermore, clinical implementation of novel prognostic tests would require integration of genomic data and best-established conventional markers of the outcome.
Here, we translate a large microarray-based breast cancer outcome predictor signature into quantitative RT-PCR-based assays of mRNA abundance levels of small gene clusters performing with similar classification accuracy. We demonstrate that identified molecular signatures provide additional predictive values over well-established conventional prognostic markers for breast cancer such as hormone receptor status and lymph node involvement. These data indicate that quantitative laboratory tests measuring expression profiles of identified small gene clusters are useful for stratifying breast cancer patients into sub-groups with distinct likelihood of positive outcome after therapy and assisting in selection of optimal treatment strategies.
Materials and Methods
The same general methods as described in Example 11 were used to carry out the work reported in this example.
Results and Discussion
The 70-gene breast cancer metastasis and survival predictor signature represents a heterogeneous set of small gene clusters independently performing with high therapy outcome prediction accuracy. Recent study on gene expression profiling of breast cancer identifies 70 genes whose expression pattern is strongly predictive of a short post-diagnosis and treatment interval to distant metastases (van 't Veer, et al., 2002). The expression pattern of these 70 genes discriminates with 81% (optimized sensitivity threshold) or 83% (optimal accuracy threshold) accuracy the patient's prognosis in the group of 78 young women diagnosed with sporadic lymph-node-negative breast cancer (this group comprises of 34 patients who developed distant metastases within 5 years and 44 patients who continued to be disease-free at least 5 years after therapy; they constitute clinically defined poor prognosis and good prognosis groups, correspondingly). We reduced the number of genes whose expression patterns represent genetic signatures of breast cancer with “poor prognosis” or “good prognosis.” Measurements of mRNA expression levels of 70 genes in established human breast carcinoma cell lines (MCF7; MDA-MB-435; MDA-MB-468; MDA-MB-231; MDA-MB-435Br1; MDA-MB-435BL3) and primary cultures of normal human breast epithelial cells were performed utilizing Q-RT-PCR method, which is generally accepted as the most reliable method of gene expression analysis and unambiguous confirmation of a gene identity. For each breast cancer cell line concordant sets of genes were identified exhibiting both positive and negative correlation between fold expression changes in cancer cell lines versus control cell line and poor prognosis group versus good prognosis group patient samples. Minimum segregation sets were selected from corresponding concordance sets and individual phenotype association indices were calculated. The four top-performing breast cancer metastasis predictor gene clusters are listed in Table 73.
A breast cancer poor prognosis predictor cluster comprising 6 genes was identified (r=0.981) using MDA-MB-468 cell line gene expression profile as a reference standard. 32 of 34 samples from the poor prognosis group had positive phenotype association indices, whereas 29 of 44 samples from the good prognosis group had negative phenotype association indices yielding 78% overall accuracy in sample classification. Another breast cancer poor prognosis predictor cluster comprising 4 genes was identified (r=0.944) using MDA-MB-435BL3 cell line gene expression profile as a reference standard. Using this 4-gene cluster, 28 of 34 samples from the poor prognosis group had positive phenotype association indices, whereas 28 of 44 samples from the good prognosis group had negative phenotype association indices overall yielding 72% accuracy in sample classification.
A breast cancer good prognosis predictor cluster comprising 14 genes was identified (r=−0.952) using MDA-MB-435Br1 cell line gene expression profile as a reference standard. 30 of 34 samples from the poor prognosis group had negative phenotype association indices, whereas 34 of 44 samples from the good prognosis group had positive phenotype association indices yielding 82% overall accuracy in sample classification. Another breast cancer good prognosis predictor cluster comprising 13 genes (r=−0.992) was identified using MCF7 cell line gene expression profile as a reference standard. 30 of 34 samples from the poor prognosis group had negative phenotype association indices, whereas 32 of 44 samples from the good prognosis group had positive phenotype association indices yielding 79% overall accuracy in sample classification.
The transcripts comprising each signature listed in Table 73 were selected based on Pearson correlation coefficients (r>0.95) reflecting a degree of similarity of expression profiles in clinical tumor samples (34 recurrent versus 44 non-recurrent tumors) and experimental cell line samples. Selection of transcripts was performed from sets of genes exhibiting concordant changes of transcript abundance behavior in recurrent versus non-recurrent clinical tumor samples (70 transcripts) and experimental conditions independently defined for each signature (6-gene signature: MDA-MB468 cells versus control; 4-gene signature: MDA-MB-435BL3 cells versus control; 13-gene signature: MCF7 cells versus control; 14-gene signature: MDA-MB-435Br1 cells versus control)(see also Example 2). mRNA expression levels of 70 genes comprising parent microarray-defined signature (van't Veer, L. J., et al., 2002; van de Vijver, M. J., et al., 2002) were measured by standard quantitative RT-PCR method in multiple established human breast cancer cell lines using GAPDH expression for normalization and compared to the expression in a control cell line. Control cells were primary cultures of normal human breast epithelial cells. Expression profiles were presented as log10 average fold changes for each transcript.
To demonstrate the ability to reduce the number of genes in the cluster, while maintaining predictive power, we selected subsets of genes within a minimum segregation set so as to raise the correlation coefficient, and tested the performance of the cluster as the set was reduced from 9 to 2 genes. Specifically, classification was performed in a cohort of 78 breast cancer patients. The outcome predictor clusters were identified using MDA-MB-435BL3 human breast carcinoma cell line as a reference standard. These results are shown in Tables 73.1 and 73.2.
As described in Example 2, we validated the classification accuracy using an independent data set, and tested performance of the 13 genes good prognosis predictor cluster on a set of 19 samples obtained from 11 breast cancer patients who developed distant metastases within five years after diagnosis and treatment and 8 patients who remained disease free for at least five years (van 't Veer, L. J., et al., 2002). 9 of 11 samples from the poor prognosis group had negative phenotype association indices, whereas 6 of 8 samples from the good prognosis group had positive phenotype association indices yielding 79% overall accuracy in sample classification.
Kaplan-Meier analysis showed that metastasis-free survival after therapy was significantly different in breast cancer patients segregated into good and poor prognosis groups based on relative values of expression signatures defined by all four small gene clusters (
Small gene clusters and a large parent signature perform with similar therapy outcome prediction accuracy in an independent cohort of 295 breast cancer patients. Recently the breast cancer prognosis prediction accuracy of the 70-gene signature was validated in a large cohort of 295 patients with either lymph node-negative or lymph node-positive breast cancer (van de Vijver, M. J., et al., 2002). The expression profile of the 70-gene breast cancer outcome predictor signature was highly informative in forecasting the probability of remaining free of distant metastasis and predicting the overall survival after therapy (id.). We compared the classification accuracy of small gene clusters and a large 70-gene parent signature applied to a cohort of 295 patients.
As shown in the Table 74, identified small gene clusters and a large parent signature perform similarly in identifying sub-groups of breast cancer patients with poor and good prognosis defined by differences in the probability of the overall survival after therapy. At the several classification threshold levels small gene clusters fully recapitulate or even outperform the 70-gene parent signature in classification accuracy of the 295 breast cancer patients (Table 74). Taken together these data are consistent with the idea that the 70-gene breast cancer prognosis signature represents a heterogeneous set of small gene clusters with high therapy outcome prediction potential. Consistent with this idea, the application of the 14-gene survival predictor signature was highly informative in classification of breast cancer patients into sub-groups with statistically significant difference in the probability of survival after therapy (
To generate the data in Table 74, 295 breast cancer patients were classified according to whether they had a good-prognosis signature or poor-prognosis signature defined by individual therapy outcome predictor signatures. Kaplan-Meier analysis was performed to evaluate the probability that patients would survive according to whether they had a poor-prognosis or a good-prognosis signature and determine the proportion of patients who would survive at least 5 or 10 years after therapy in poor-prognosis and good-prognosis sub-groups. Hazard ratios, 95% confidence intervals, and P values were calculated with use of the log-rank test. The number of correct predictions in poor-prognosis and good-prognosis groups is shown as a fraction of patients with the observed clinical outcome after therapy (79 patients died and 216 patients remained alive). The classification performance of different signatures were evaluated using one common threshold level (0.00) and optimized threshold levels adjusted for each gene cluster to achieve the most statistically significant (highest hazard ratio and lowest P value) discrimination in survival probability between patients assigned to poor and good prognosis groups.
The 70-gene signature, in contrast to small gene clusters, is not suitable for breast cancer outcome prediction in patients with estrogen receptor negative tumors. Consistent with well-established prognostic value of the estrogen receptor status of breast tumors (see Introduction), 97 percent of patients in the good prognosis group defined by the 70-gene signature had estrogen receptor positive (ER+) tumors (van de Vijver, M. J., et al., 2002). Conversely, ninety six percent of breast cancer patients with the estrogen receptor negative (ER−) tumors (66 of 69 patients at the cut off level <0.45) had expression profile of the 70 genes predictive of a poor outcome after therapy. Two important conclusions can be drawn from this association. First, breast cancer patients with ER+tumors and poor prognosis expression profile of the 70 genes may have yet unidentified functional defect within an ER-response pathway. Second, a 70-gene signature appears to assign rather uniformly a vast majority of the patients with ER-tumors into poor prognosis category and, therefore, is not suitable for prognosis prediction in this group of breast cancer patients.
In agreement with many previous observations, patients with ER− tumors had significantly worst survival after therapy compared to the patients with ER+tumors in the cohort of 295 breast cancer patients (
Next we determined that application of a survival predictor algorithm would identify sub-groups of patients with distinct clinical outcome after therapy in breast cancer patients with ER-negative tumors, thus providing additional predictive value to the therapy outcome classification based on ER status alone. We were unable to generate statistically meaningful prognostic stratification of ER-negative breast cancer patients using a parent 70-gene signature (data not shown). However, we were able to identify two small gene clusters comprising 5 and 3 genes (Table 75) that appear highly informative in classifying breast cancer patients with ER-negative tumors into good and poor prognosis sub-groups with statistically distinct probability of survival after therapy (
In the group of 69 breast cancer patients with ER-negative tumors (
Outcome classification of breast cancer patients with ER-positive tumors using a 14-gene survival predictor signature. To further validate the clinical utility of identified signatures, we determined that application of a 14-gene survival predictor cluster is informative in classifying breast cancer patients with ER-positive tumors. Kaplan-Meier analysis showed that application of the 14-gene survival predictor signature identified three sub-groups of patients with statistically distinct probabilities of survival after therapy in the cohort of 226 breast cancer patients with ER-positive tumors (FIGS. 67A&B). The median survival after therapy of patients in the poor prognosis sub-group defined by the 14-gene survival predictor signature was 7.2 years (
Therapy outcome prediction in breast cancer patients with lymph node-negative disease using survival predictor signatures. Invasion into axillary lymph nodes is considered as one of the most important negative prognostic factors in breast cancer and patients with no detectable lymph node involvement are classified as having good prognosis (Krag, et al., 1998; Singletary, et al., 2002; Jatoli, et al., 1999). Breast cancer patients with lymph node negative disease typically would not be selected for adjuvant systemic therapy and usually treated with surgery and radiation. Recent data demonstrated that up to 33% of these patients would fail therapy and develop recurrence of the disease after a 10-year follow-up (Early Breast Cancer Trialists' Collaborative Group. Tamoxifen for early breast cancer: an overview of the randomized trials. Lancet, 351: 1451-1467, 1998). Therefore, we tested whether application of the 14-gene survival predictor signature would aid in identifying breast cancer patients with lymph-node negative tumors that are at high risk of treatment failure.
Kaplan-Meier analysis showed that the 14-gene survival predictor signature (Tables 29 and 73) identified two sub-groups of patients with statistically distinct probability of survival after therapy in the cohort of 151 breast cancer patients with lymph node negative disease (
Kaplan-Meier analysis also demonstrated that the 14-gene survival predictor signature identified two sub-groups of patients with statistically distinct probability of survival after therapy in the cohort of 109 breast cancer patients with ER-positive tumors and lymph node negative disease (
Next we determined that application of small gene clusters comprising 5 and 3 genes (Table 75) that appear highly informative in classification of breast cancer patients with ER-negative tumors into good and poor prognosis sub-groups with statistically distinct probability of survival after therapy (
Therapy outcome prediction in breast cancer patients with lymph node-positive disease using survival predictor signatures. Breast cancer patients with invasion into axillary lymph node are considered as having a poor prognosis and usually treated with adjuvant systemic therapy. The patients with poor prognosis are thought to benefit most from adjuvant systemic therapy (see Introduction). In the cohort of 295 breast cancer patients, ten of 151 (6.6%) patients who had lymph node-negative disease and 120 of the 144 (83.3%) patients who had lymph node-positive disease had received adjuvant systemic therapy (van de Vijver, et al. 2002). This treatment strategy was clearly beneficial for patients with lymph node-positive disease, because sub-groups of patients with distinct lymph node status in the cohort of 295 patients had statistically indistinguishable survival after therapy (data not shown). Next we determined therapy outcome prediction using survival predictor signatures identified in accordance with the present invention to be informative in breast cancer patients with lymph node-positive disease.
Kaplan-Meier analysis show that application of the 14-gene survival predictor signature identify three sub-groups of patients with statistically distinct probability of survival after therapy in the cohort of 144 breast cancer patients with lymph node positive disease (
Using the 14-gene survival predictor signature we identified two sub-groups of patients with statistically distinct probabilities of survival after therapy in the cohort of 117 breast cancer patients with ER-positive tumors and lymph node positive disease (
Next we determined that the small gene clusters comprising 5 and 3 genes (Table 75) also are informative in classifying sub-groups of ER-negative patients with lymph node-positive disease. In the group of 27 breast cancer patients with ER-negative tumors and lymph node-positive disease (
Estimated long-term survival benefits of using gene expression profiling as a component of multiparameter therapy outcome classification of breast cancer patients. Next we estimated the potential clinical benefits of applying gene expression survival predictor signatures identified in accordance with the methods of the present invention for classifying breast cancer patients at the time of diagnosis into sub-groups with distinct probabilities of survival after therapy. In our estimate we used the assignment of the patient into a poor outcome classification sub-group as a criterion of treatment failure and reason for prescription of additional cycle(s) of adjuvant systemic therapy. We have made the estimate of potential therapeutic benefits in the cohort of 295 breast cancer patients (van de Vijver, et al. 2002) and based our estimate on the assumption that the use of additional cycle(s) of adjuvant systemic therapy would be prescribed to patients classified within a poor prognosis sub-group. In the cohort of 295 breast cancer patients, ten of 151 (6.6%) patients who had lymph node-negative disease and 120 of the 144 (83.3%) patients who had lymph node-positive disease had received adjuvant systemic therapy (id.), indicating that a major difference in treatment protocols between LN+ and LN− sub-groups was the application of adjuvant systemic therapy in patients with lymph node positive disease. We accepted the actual 5- and 10-year survival in the corresponding classification categories as the expected therapy outcome for a given sub-group. We assumed that each additional cycle of adjuvant systemic therapy would result in the same therapy outcome as was actually documented in the most relevant sub-groups of the 295 patients. Therapy outcome for patients classified into poor prognosis sub-groups and treated with additional cycle(s) of adjuvant systemic therapy is expected to be in 37% of patients in good therapy outcome category for ER+LN+ and ER+LN-poor signature sub-groups and in 41% of patients in good therapy outcome category for ER-LN+ and ER-LN− poor signature sub-groups (Table 76). Finally, we assumed that patients classified into good prognosis sub-groups would receive the same treatment and would have the same outcome as in the original cohort of 295 patients (van de Vijver, et al., 2002). Based on these assumptions we calculated the number of patients that would be expected to have good and poor survival outcome after therapy and estimated the expected 10-year survival in each classification sub-groups (Table 76).
The estimate of potential therapeutic benefits provided in Table 76 is based on the cohort of 295 breast cancer patients (van de Vijver, et al. 2002) and premised on the assumption that additional cycle(s) of adjuvant systemic therapy would be prescribed to patients classified into poor prognosis sub-groups. In the cohort of 295 breast cancer patients, ten of 151 (6.6%) patients who had lymph node-negative disease and 120 of the 144 (83.3%) patients who had lymph node-positive disease had received adjuvant systemic therapy (id.). We accepted the actual 5- and 10-year survival in the corresponding classification categories as the expected therapy outcome for a given sub-group. We assumed that each additional cycle of adjuvant systemic therapy would result in the same therapy outcome as was actually documented in the most relevant sub-groups of the 295 patients. Therapy outcome for patients classified into poor prognosis sub-groups and treated with additional cycle(s) of adjuvant systemic therapy is expected to be in 37% of patients in good therapy outcome category for ER+LN+ and ER+LN− poor signature sub-groups and in 41% of patients in good therapy outcome category for ER-LN+ and ER-LN− poor signature sub-groups.
One of the most interesting end-points of this analysis is the prediction that patients with ER-LN− and ER-LN+breast cancer classified into poor prognosis sub-groups would be expected to show a most dramatic increase in 10-year survival after therapy (Table 76). This prediction is consistent with the generally accepted notion that breast cancer patients with poor prognosis would benefit most from adjuvant systemic therapy (see Introduction). The estimated modest increase in the overall 10-year survival (Table 76) may translate every year into ˜7,000-9,000 more breast cancer survivors after 10-year follow-up. Our ability to accurately segregate at the time of diagnosis breast cancer patients with low probability of survival after therapy should lead to more rapid development of novel efficient therapeutic modalities specifically targeting most aggressive therapy-resistant breast cancers.
While the invention has been described with reference to specific methods and embodiments, it will be appreciated that various modifications may be made without departing from the invention, the scope of which is limited only by the appended claims. All references cited, including scientific publications, patent applications, and issued patents, are herein incorporated by reference in their entirety for all purposes.
This application is a continuation of co-pending U.S. application Ser. No. 10/660,434, filed Sep. 10, 2003, which claims the benefit of U.S. Provisional Application 60/410,018 filed Sep. 10, 2002; U.S. Provisional Application 60/411,155, filed Sep. 16, 2002; U.S. Provisional Application 60/429,168, filed Nov. 25, 2002; U.S. Provisional Application 60/444,348, filed Jan. 31, 2003; and U.S. Provisional Application 60/460,826, filed Apr. 3, 2003, each of which is incorporated by reference in its entirety.
This invention was made using federal funds awarded by the National Institutes of Health, National Cancer Institute under contract number 1RO1CA89827-01. The government has certain rights to this invention.
| Number | Date | Country | |
|---|---|---|---|
| 60410018 | Sep 2002 | US | |
| 60411155 | Sep 2002 | US | |
| 60429168 | Nov 2002 | US | |
| 60444348 | Jan 2003 | US | |
| 60460826 | Apr 2003 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10660434 | Sep 2003 | US |
| Child | 10861003 | Jun 2004 | US |
