METHOD FOR DETERMINING RCC SUBTYPES

FIELD OF THE INVENTION

The present invention relates to a method for determining in a subject's biological sample the relative proportions of papillary renal cell carcinoma (pRCC), clear cell renal cell carcinoma (ccRCC), and chromophobe renal cell carcinoma (chRCC), an array comprising capture molecules capable of specifically binding to RCC signature genes or coding sequences thereof or products encoded thereby, and to the use of RCC signature genes for classifying a subject into a renal cell carcinoma (RCC) risk group and/or for determining in a subject's biological sample the relative proportions of pRCC, ccRCC, and chRCC.

BACKGROUND OF THE INVENTION

Renal cell carcinoma (RCC) comprises several histologically defined tumors that differ in biology, clinical course and response to treatment. The major subtypes are clear cell RCC (ccRCC), papillary RCC (pRCC), and chromophobe RCC (chRCC), which account for 65-70%, 15-20%, and 5-7% of all RCCs, respectively (Inamura, Translocation Renal Cell Carcinoma: An Update on Clinicopathological and Molecular Features, Int. J. Mol. Sci. 9(9), p. 1-11 (2017)). In general, ccRCC has poor and chRCC has favorable prognosis. pRCC represents a heterogeneous group of RCC with intermediate prognosis compared to ccRCC and chRCC that has been subdivided in type 1 and type 2, a subset of tumors with mixed histology, and a small fraction of CpG island methylator phenotype (CIMP)-associated tumors (C. J. Ricketts et al., The Cancer Genome Atlas Comprehensive Molecular Characterization of Renal Cell Carcinoma, Cell Reports 23(1), p. 313-326 (2018)). Type 1 pRCC is associated with better prognosis than type 2 pRCC. CIMP tumors are characterized by poor survival.

Considering its significant prognostic as well as therapeutic implications, the correct determination of the subtype is of utmost importance. In clinical medicine, surgical specimen from RCC tumors are manually examined and classified by pathologists through histological and immunohistochemical analyses.

Pathological re-evaluation and bioinformatics analyses of molecular data have recently pointed to the shortcomings of pathological assessments of RCCs (Büttner et al., Survival Prediction of Clear Cell Renal Cell Carcinoma Based on Gene Expression Similarity to the Proximal Tubule of the Nephron, Eur. Urol. 68(6), p. 1016-1020 (2015); Chen et al., Multilevel Genomics-Based Taxonomy of Renal Cell Carcinoma, Cell Reports 14(10), p. 2476-2489 (2016); Schaeffeler et al., Metabolic and Lipidomic Reprogramming in Renal Cell Carcinoma Subtypes Reflects Regions of Tumor Origin, Eur. Urol. Focus (2018); C. J. Ricketts et al., loc. cit.). Manual classification is subjective and therefore bears potential for mislabeling or inconsistencies, especially in histologically ambiguous cases.

Rini et al., A 16-Gene Assay to Predict Recurrence After Surgery in Localised Renal Cell Carcinoma: Development and Validation Studies, Lancet Oncol. 16(6), p. 676-685 (2015), describe a prognostic multigene signature to improve prediction of recurrence risk in clear cell renal cell carcinoma. However, this method does not allow for an RCC subtype classification.

WO 2015/131095 discloses a method for distinguishing clear cell type A (ccA) renal cell carcinoma from clear cell type B (ccB) renal cell carcinoma in a subject. However, this method requires a statistically validated reference. Furthermore, it also does not allow for an RCC subtype classification beyond subtypes of ccRCC.

Wang et al., Identification and Validation of a 44-Gene Expression Signature for the Classification of Renal Cell Carcinomas, J. Exp. Clin. Cancer Res. 36:176, p. 1-11 (2017), disclose a 44-gene expression signature derived from microarray analysis which was associated with the histological differentiation of renal tumors and is proposed for tumor subtype classification. However, such gene expression signature has so far not proved successful in practice. Furthermore, the known method does not allow a direct subtype classification but only a clustering, i.e. individual biological samples cannot be classified.

Hence, there is a need for an objective subtype classification for RCC.

The present invention satisfies these and other needs.

SUMMARY OF THE INVENTION

The present invention provides a method for determining in a subject's biological sample the relative proportions of papillary renal cell carcinoma (pRCC), clear cell renal cell carcinoma (ccRCC), and chromophobe renal cell carcinoma (chRCC), the method comprising:

- (a) Providing a biological sample from a subject suspected of being affected by RCC,
- (b) Assaying said biological sample to determine expression level values of
  - at least one of the signature genes listed in Table 1,
  - at least one of the signature genes listed in Table 2, and
  - at least one of the signature genes listed in Table 3,
- (c) Subjecting the obtained expression level values to a signal separation method, thereby determining relative proportions of pRCC, ccRCC, and chRCC in said biological sample.

The inventors have developed an objective and reference-free RCC sub-type classification system based on gene expression data by which the disadvantages of the methods known in the art can be reduced or even avoided. The present invention can also be used to separate tumors that can be unambiguously assigned to the three major histological subtypes from those combining features from different subtypes. The method according to the invention also allows a clear statement about the probability of survival of the affected patients that is more accurate and less prone to errors than a common pathological evaluation.

The present invention is superior to currently performed manual histopathological classification because (1) it provides a precise and objective molecular-based procedure to classify RCC, (2) it quantifies the proportions of the major subtypes in histologically ambiguous RCC, (3) the predicted proportional subtype composition is directly associated to a prognostic estimate, and (4) it is the first molecular-based prognostic system that is applicable to ccRCC, pRCC, and chRCC.

The term “subject” as used herein refers to a member of any invertebrate or vertebrate species. Accordingly, the term “subject” is intended to encompass any member of the Kingdom Animalia including, but not limited to the phylum Chordata (i.e., members of classes Osteichythyes (bony fish), Amphibia (amphibians), Reptilia (reptiles), Ayes (birds), and Mammalia (mammals)), and all orders and families encompassed therein. In an embodiment, the subject is a human.

A “biological sample” as used herein refers to biological material originating from the subject and comprises nucleic acids, and/or proteins, and/or peptides and/or polypeptides and/or fragments thereof. In an embodiment of the invention the biological sample comprises cellular material, cells or tissues. Preferably, the biological material comprises cells suspected of including renal carcinoma cell(s) or cells being renal carcinoma cell(s). In the clinical routine the biological sample may be a biopsy sample taken from potentially tumorous or RCC tissue, blood plasma, urine etc.

The terms “nucleic acid molecule” and “nucleic acid” refer to deoxyribonucleotides, ribonucleotides, and polymers thereof, in single-stranded or double-stranded form. As used herein, the terms “peptide” and “polypeptide” refer to polymers of at least two amino acids linked by peptide bonds. Typically, “peptides” are shorter than “polypeptides” and the latter are typically shorter than proteins, but unless the context specifically requires, these terms are used interchangeably herein.

As used herein the term “gene” refers to a hereditary unit including a sequence of DNA that occupies a specific location on a chromosome and that contains the genetic instruction for a particular characteristic or trait in an organism. Similarly, the phrase “gene product” refers to biological molecules that are the transcription and/or translation products of genes. Exemplary gene products include, but are not limited to mRNAs and polypeptides that result from translation of mRNAs.

A “signature gene” as used herein refers to a gene listed in any of Table 1, 2, and 3 and being specifically expressed and indicative for pRCC (Table 1), ccRCC (Table 2), and chRCC (Table 3), respectively. The signature genes as referred to herein make up a so-called gene signature with a unique pattern of gene expression which is characteristic in cells of ccRCC, pRCC, and chRCC.

The signature genes can be clearly identified in Tables 1, 2, and 3 by means of their GeneID, i.e. the first column of the respective table. GeneID is a unique identifier that is assigned to a gene record in the meta search engine or database ‘Entrez Gene’ operated by the National Center for Biotechnology Information (NCBI). Synonyms for ‘GeneID’ are Gene Identifies (NCBI), NCBI gene ID, Entrez gene ID, NCBI geneid, or Gene identifier (Entrez). The ‘Symbol’ column lists the HUGO Gene symbols of the genes. The columns headlined ‘ccRCC’, ‘chRCC’, and ‘pRCC’ list the medians of relative expression values of the respective signature genes in the indicated RCC subtypes. The expression values are (non-log-transformed) processed signal intensities originally measured with the Affymetrix HTA2.0 array.

As used herein “at least one” signature gene refers to the minimum of one signature gene of each Table or group that needs to be analyzed. In embodiments of the invention 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, or all 58 signature genes of each Table or group are analyzed in respect of their expression levels. Further, in embodiments of the invention the numbers of signatures genes can be the same or different from each of the Tables or groups, i.e. x genes out of Table 1, y genes out of Table 2, and z genes out of Table 3 can be analyzed, while x, y, and z stand for the same or different integers.

While, according to the findings of the inventors, the analysis of one signature gene per Table is sufficient for a determination of the relative proportions of pRCC, ccRCC, and chRCC in the biological sample, the accuracy of determination and the reliability of the method is increasingly enhanced by the inclusion of more than one signature gene up to all 58 signature genes of each Table or group, respectively.

As is known to one of ordinary skill in the art, gene expression levels can be assayed at the level of RNA and/or at the level of protein. As such, in some embodiments RNA is extracted from the biological sample and analyzed by techniques that include, but are not limited to, PCR analysis (in some embodiments, quantitative reverse transcription PCR), nucleotide sequencing and/or array analysis. Alternatively or in addition, gene expression levels can be assayed by determining the levels at which proteins or polypeptides are present in the biological sample. This can also be done using arrays, and exemplary methods for producing peptide and/or polypeptide arrays attached to a suitable carrier are well known to the skilled person. In each case, one of ordinary skill in the art would be aware of techniques that can be employed to determine the expression level of a gene in the biological sample.

A “signal separation method” as used herein refers to a process for the analysis of mixtures of signals with the objective to recover the original component signals from the mixture. In particular, it is referred to a method for determining the relative proportions of ccRCC, pRCC, and chRCC in said biological sample. It includes, but is not limited to methods like blind signal separation (BSS) such as deconvolution, principal component analysis (PCA), independent component analysis (ICA), machine learning (supervised learning/classification/regression) and data mining (unsupervised learning/clustering); see Vandesompele et al., Computational deconvolution of transcriptomics data from mixed cell populations, Bioinformatics 34(11): 1969-1979 (2018).

The object underlying the invention is herewith completely solved.

The inventors have realized that determining the expression level values of only at least one of the signature genes listed in each of Tables 1, 2, and 3, i.e. of only at least three different genes, and subjecting the obtained expression level values to a signal separation method allows the determination of relative proportions of pRCC, ccRCC, and chRCC in the biological sample.

The method according to the invention allows an objective determination of the RCC subtype, thereby avoiding an incorrect subjective classification made by a pathologist. Another advantage of the method according to the invention over the methods in the art is that no reference is required to allow a correct subtype classification.

Knowledge of a pRCC versus ccRCC versus chRCC class assignment allows for an assessment of risk for recurrence or cancer specific death, and can be used to augment clinical information to make more accurate risk assessments. Knowledge of risk allows clinicians to tailor the post-operative evaluations, and to consider adjuvant therapy options. Particular changes to care that might arise when a subject's RCC is classified as comprising significant proportions of any of pRCC, ccRCC and chRCC might include, but not be limited to more intensive monitoring, consideration of surgical intervention, drug/radiation therapy, and/or finding an adjuvant therapy trial for the subject to reduce risk for recurrence.

In an embodiment of the method according to the invention in step (b) said biological sample is assayed to determine expression level values of at least two of the signature genes listed in Table 1, at least two of the signature genes listed in Table 2, and at least two of the signature genes listed in Table 3.

This measure has the advantage that the accuracy of the determination of the relative proportions of pRCC, ccRCC, and chRCC in said biological sample is further increased.

In another embodiment of the present invention the signal separation method is a blind signal separation method (BSS).

Blind signal separation (BSS), also known as blind source separation, refers to a method for the separation of a set of source signals from a set of mixed signals, without the aid of information (or with very little information) about the source signals or the mixing process. The inventors have realized that BSS, if used in the method of the invention, allows a high degree of signal separation and ensures the achievement of reliable results.

In a further embodiment of the method according to the invention the blind separation method is deconvolution, preferably computational deconvolution.

Deconvolution is an algorithm-based process used to reverse the effects of convolution on recorded data. Initially, deconvolution has been mainly used in the techniques of signal processing and image processing. Computational deconvolution refers to a computer-assisted deconvolution method which has been used to address specific questions of biology or bioinformatics, as e.g. described in S. S. Shen-Orr and R. Gaujoux, Computational Deconvolution: Extracting Cell Type-Specific Information from Heterogeneous Samples, Current Opinion in Immunology 25, p. 571-578 (2013); F. Avlia Cobos et al., Computational Deconvolution of Transcriptomics Data from Mixed Cell Populations, Bioinformatics 34, p. 1969-1979 (2018); A. R. Abbas et al., Deconvolution of Blood Microarray Data Identifies Cellular Activation Patterns in Systemic Lupus Erythematosus., PloS one 4, e6098 (2009); R. Gaujoux and C. Seoighe, CellMix: A Comprehensive Toolbox for Gene ExpressionDeconvolution., Bioinformatics (Oxford, England), p. 1-2 (2013). The inventors, however, realized for the very first time that the deconvolution method can be used in an advantageous manner to determine in a heterogeneous biological or RCC sample the relative proportions of the respective RCC subtypes.

In an embodiment of the present invention after step (c) the following step is carried out: Classifying the subject into a risk group on the basis of the relative proportions of at least one of ccRCC, pRCC, and chRCC in said biological sample, preferably of ccRCC in said biological sample.

The inventors have realized that by the developed inventive concept not only the relative proportions of the respective RCC subtypes can be determined in a biological or RCC sample but also a prediction of the risk for cancer-specific death of the patient. This measure implements the invention into the clinic in an advantageous maner.

In another embodiment of the present invention the risk group is selected from “low risk”, “intermediate risk”, and “high risk” according to the prognosis for the subject.

This measure allows a rapid allocation of a prognosis for the affected subject in daily hospital routines. “Low risk” refers to a high likelihood of the subject to survive for more than 5 years, “high risk” refers to a low likelihood of the subject to survive for more than 5 years, and “intermediate” refers to a medium likelihood of the subject to survive for more than 5 years, each 5 years period beginning to run at the date of the initial diagnosis on the basis of the biological sample from the subject obtained by surgery. In an embodiment of the invention in the “low risk” group the likelihood is about 87-96% or higher, preferably 91%, in the “high risk” group the likelihood is about 34-69%, preferably 48%, and in the “intermediate risk” group the likelihood is about 72-81%, preferably 76%.

In another embodiment of the invention the “low risk” group is determined by a relative ccRCC proportion in the range of about 0 to 12%, further preferably of about 0 to 5%, further preferably of about 0 to 3%, and highly preferably of about 0%.

In yet another embodiment of the invention the “intermediate risk” group is determined by a relative ccRCC proportion in the range of about ≥7.5 to ≤25%, further preferably of about ≥10 to ≤20%, and highly preferably of about ≥13 to ≤17.

In another embodiment of the invention the “intermediate risk” group is determined by a relative ccRCC proportion in the range of about ≥62.5%, further preferably of about ≥70%, %, further preferably of about ≥77.5%, further preferably of about 90%, and highly preferably of about 100%.

In still another embodiment of the invention the “high risk” group is determined by a relative ccRCC proportion in the range of about ≥16 to ≤77.5%, preferably of about ≥20 to ≤70%, further preferably of about ≥25 to ≤62.5%, and highly preferably of about 40%.

The inventors have realized that the indicated thresholds of the relative proportion of the respective ccRCC subtype allow an allocation of the subject to a risk group “low risk”, “high risk”, and/or “intermediate risk”. It is accepted that by using the rough or less-specific thresholds each mentioned for the less preferred embodiments a subject may fall into more than one risk group. However, it is clear that the more specific thresholds each mentioned for the more or further preferred embodiments allow an increasingly distinctive allocation of a subject to a specific risk group.

In another embodiment of the invention in step (b) the assaying involves the use of RNA sequencing, a PCR-based method, a microarray-based method, a hybridization-based method and/or an antibody-based method.

This measure takes advantage of such methods for assaying the biological sample which have been proven their suitability for determining expression level values of genes or gene products.

Another subject-matter of the present invention is an array comprising capture molecules capable of specifically binding to

- biomolecules encoding or encoded by at least one, preferably at least two of the signature genes listed in Table 1 or segments thereof,
- biomolecules encoding or encoded by at least one, preferably at least two of the signature genes listed in Table 2 or segments thereof, and
- biomolecules encoding or encoded by at least one, preferably at least two of the signature genes listed in Table 3.

The “biomolecules” include, but are not limited hereto, nucleic acid molecules encoding the signature genes, proteins, peptides or polypeptides encoded by the signature genes. The “capture molecules” include, but are not limited hereto, nucleic acid molecules (e.g. hybridization probes, aptamers etc.), antibodies and fragments thereof.

The embodiments, features, characteristics, and advantages disclosed for the method according to the invention apply likewise to the array according to the invention.

The term “array” is to be understood in its broadest sense and refers to any kind of test format suitably adapted to comprise the capture molecules and to carry out a binding reaction of the signature genes or gene products or equivalents to the capture molecules. Preferably, the array is a microarray.

Another subject-matter of the present invention is the use of

- at least one, preferably at least two of the signature genes listed in Table 1,
- at least one, preferably at least two of the signature gene listed in Table 2, and
- at least one, preferably at least two of the signature genes listed in Table 3,
  
  for classifying a subject into a renal cell carcinoma (RCC) risk group.

Another subject-matter of the present invention is the use of

- at least one, preferably at least two of the signature genes listed in Table 1,
- at least one, preferably at least two of the signature genes listed in Table 2, and
- at least one, preferably at least two of the signature genes listed in Table 3,
  
  for determining in a subject's biological sample the relative proportions of papillary renal cell carcinoma (pRCC), clear cell renal cell carcinoma (ccRCC), and chromophobe renal cell carcinoma (chRCC), further preferably for classifying the subject into a RCC risk group.

The embodiments, features, characteristics, and advantages disclosed for the method according to the invention apply likewise to the uses according to the invention.

It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination or as suitable in any other described embodiment of the invention.

Certain features described in the context of various embodiments are not to be considered essential features of those embodiments, unless the embodiment is inoperative without those elements.

It is to be understood that the before-mentioned features and those to be mentioned in the following cannot only be used in the combination indicated in the respective case, but also in other combinations or in an isolated manner without departing from the scope of the invention.

The invention is now described and explained in further detail by referring to the following non-limiting examples and drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1: Overview of the data analysis workflow including different cohorts and RNA quantification technologies (microarray and RNA-Seq) used in the development of the present invention.

FIG. 2: Selection of candidate genes for the signature matrix using cohort C1. (A) Hierarchical clustering of cohort C1 (n=52) using Ward's method. (B) Tumor purity as determined by the ESTIMATE method varies between RCC subtypes. (C) The scatter plot shows P-values obtained from model comparison for each gene. The goal of the analysis was to identify genes for which expression variability is better explained by differences in RCC subtypes than by tumor purity. 28464 genes were stronger associated to RCC subtype than to tumor purity. (D) 11195 genes remained after four filtering steps. “TCGA”: genes covered in TCGA RNA-Seq data; “HG U133 Plus 2.0”: genes covered in this microarray; “exp. level”: genes with median expression above the global median in C1 in minimum one subtype; “purity ind.”: genes independently expressed of tumor purity. (E) Log2 fold changes of 3686 subtype-specific genes obtained by analysis of variance and subsequent post-hoc testing using Tukey's method. For each gene and subtype the minimum log fold change compared to the two respective other subtypes was calculated.

FIG. 3: Hierarchical clustering of cohort C2. Hierarchical clustering of cohort C2 (n=143) using Ward's method. Cohort C2 was a combined cohort containing RCC samples from five different studies (Table S1).

FIG. 4: Signature matrices with increasing number of genes were tested. The initial matrix included the top two genes per subtype exhibiting the highest log fold change compared to the respective other subtypes (FIG. 2E). The median gene expression per subtype based on cohort C1 was used. Matrix sizes ranged from 6, i.e. the top two genes per subtype, to 1500 genes. Each matrix was used to deconvolve the 143 transcriptomes from cohort C2. The maximum absolute difference (MAD) in PSA was computed between consecutive matrices for each sample. (A) The 0.95-quantile MAD between two consecutive matrices is shown. (B) Subsets including 50% of the samples were randomly drawn 10000 times from cohort C2 and for each tested matrix and subset the percentage of samples experiencing a MAD>5% compared to the predecessor matrix was determined. Matrix RCC58 including the top 58 genes per subtype was chosen (marked in light red).

FIG. 5: Receiver operating characteristic (ROC) analysis. Proportional subtype assignments (PSA) were compared with the pathological classification of the TCGA RCC cohort (Ricketts et al., Cell Reports, 2018) (n=819). The various pRCC subtypes annotated in this study were here subsumed under “pRCC”. For each of the three proportional subtype assignments (below also referred to as scores), the ability to distinguish the tumors of the corresponding subtype from the pathologically differently classified tumors was investigated in the TCGA RCC cohort for different cut-offs. AUC: area under the curve.

FIG. 6: Kaplan-Meier curves showing cancer-specific survival (CSS) of pathological subtypes (C.J. Ricketts et al., loc. cit.) in cohort C3 (n=803).

FIG. 7: Restricted cubic spline estimates of the relationships between subtype scores and log relative hazards, based on a Cox PH model with endpoint CSS. 5, 4, and 5 knots were used for fitting ccRCC-score, pRCC-score, and chRCC-score to log relative hazards, respectively. The log relative hazards were shifted in such a way that patients whose tumor was assigned a score value of 100 had a log relative hazard of zero, respectively.

FIG. 8: Risk prediction using the ccRCC score (ClearScore) in the TCGA RCC cohort (C3). (A) Relationship between ClearScore and log relative hazard for 828 patients from C3 based on cubic polynomial Cox proportional hazard modelling with endpoint CSS. 36 of 864 patients were disregarded due to lack of survival data or non-validity of the deconvolution approach (as determined by a permutation P-value estimation approach). (B) Estimated 1-, 2-, and 5-year cancer-specific survival rate in dependence of the ClearScore. (C) Distribution of ClearScore values in the different RCC subtypes defined by Ricketts et al., Cell Reports, 2018. For 789 of 828 tumors a histological classification was available (T1=type 1 pRCC, T2=type 2 pRCC, Unc.=unclassified pRCC, MD=metabolically divergent chRCC). (D) Prognostic prediction by ClearScore significantly improved pathological classification of 789 patients. Here, CIMP cases were not considered as a separate subtype but were assigned to their pathological subtypes as defined in Ricketts et al., Cell Reports, 2018. Chi-square statistic values depict the improvement of the model likelihood when estimated log relative hazards from FIG. 8A were added to the Cox model initially including the pathological classification (left) or vice versa (right). Chi-squared test P-values are shown in the bars.

FIG. 9: Graph illustrating the estimated relationship between ClearScore and the hazard ratio for endpoint cancer-specific death, using a ClearScore of 0% as reference (i.e., a ClearScore of 0%, the hazard ratio was set to 1). The hazard ratio is calculated by taking the exponential of the log relative hazards from FIG. 8A. For example, a hazard ratio of 3 means that the risk of cancer-specific death is three times higher compared to patients having a ClearScore of 0%. Risk groups were formed by categorizing the log relative hazards from FIG. 8A using conditional inference trees with endpoint cancer-specific survival. Hence, the ClearScore allows a classification of the patients into a “high risk” (top area), “low risk” (bottom area) and “intermediate risk” (middle area) group. The dashed lines indicate the pointwise standard errors. The points indicate the actual ClearScore values occurring in C3.

FIG. 10: Analysis of random signature gene subsets. 2, 5, 10, and 20 genes per subtype were randomly drawn from the 3×58 signature genes. Random drawing was repeated 10,000 times for each subset size. PSA were determined with the reduced signature gene sets. The first row shows the log-rank test P-values from univariate survival analyses in the TCGA RCC cohort (n=847) with subtype proportions, i.e. ccRCC-score, pRCC-score or chRCC-score, as predictors. Restricted cubic splines were used to model the relationship between scores and CSS. The second row shows area under the curve (AUC) values from receiver operating characteristics (ROC) analyses for each score and subset size. PSA were compared with the histological classification of the TCGA RCC cohort published by Ricketts et al., Cell Reports, 2018, (n=819). The various pRCC subtypes annotated in this study were here subsumed under “pRCC”.

EXAMPLES
1. Overview

A subtype classification system based on gene expression data was developed for renal cell carcinoma (RCC). The basic idea was to model any RCC sample as a linear combination of clear cell RCC (ccRCC), chromophobe RCC (chRCC) and papillary RCC (pRCC). More than 95% of all RCC are assigned to one of these subtypes based on histological analysis and they represent both proximal and distal cell types as origin of kidney cancer evolution. Essentially, the inventors assumed a tumor not necessarily belonging to only one of these subtypes, but to carry parts of each of them. Therefore, rather than categorizing a tumor into one of the subtypes, the inventors intended to break down its composition through proportional subtype assignments (PSA).

Applying the linearity assumption (Y. Zhao and R. Simon, Gene Expression Deconvolution in Clinical Samples. Genome Med. 2(12), p. 93 (2010), the expression of each gene in the RCC sample to be analyzed can be modeled as weighted average of the expression of this gene in ccRCC, pRCC and chRCC.

The inventors realized that signal separation, in particular computational deconvolution represented the method of choice for this problem. The weights correspond to the proportional composition and are estimated by computational deconvolution. The objective of deconvolution is to find a solution to the system of linear equations: m=f×S. Here, the unknown proportions of ccRCC, pRCC and chRCC in a sample A are modeled by the vector f of coefficients. m represents the vector containing the expression levels of signature genes in A. S is a signature matrix including the expression levels of the signature genes in ccRCC, pRCC and chRCC. Signature genes are defined based on a set of ccRCC, pRCC and chRCC samples that could be uniquely assigned by pathologists or previous analyses of molecular data. The matrix equation can be solved for f using standard linear least squares regression (Abbas et al., loc. cit.). To increase stability of subtype assignments we use robust linear regression as implemented in the “rlm” function from the R-package MASS. Deconvolution is performed on linear, i.e. non-log-transformed, expression data as suggested by (Y. Zhong and Z. Liu, Gene expression deconvolution in linear space. Nat. Methods 9(1), p. 8-9 (2011)). Further, linear expression levels are centered to zero mean and scaled to unit variance preceding deconvolution. Negative coefficients in f are set to zero and percentages are calculated by dividing the three estimated coefficients by their sum.

Gene expression deconvolution has been successfully applied to characterize the cell composition of heterogeneous samples, e.g. peripheral blood that includes many different immune cell types (S.S. Shen-Orr and R. Gaujoux, loc. cit.). Here, RCC were modeled as heterogeneous tissues that are composed of varying proportions of ccRCC, chRCC and pRCC.

This study did groundwork by attempting for the first time to comprehensively detect and quantify clear as well as composed signals in RCC samples indicative of the tumor type. Methodologically, an semi-supervised approach has been developed to utilize yet unknown patterns in gene expression profiles of RCC samples for subtype classification.

The inventors' approach is able to separate RCC tumors that can be unambiguously assigned to one of the main histological subtypes from those evading a clear histological classification. Unclear tumors were described as mixed types that combine features from different subtypes. Further, PSA enabled a new definition of RCC risk groups that is significantly stronger associated with patient survival than common pathological classification. Concluding, PSA as determined by the method according to the invention simplifies classification of RCC and specifies prognosis.

2. Material and Methods
Patient Cohorts

FIG. 1 shows the cohorts and their use in this work.

RCC cohort 1 (C1) consisted of 52 primary tumor samples with either clear cell (n=18), papillary (n=18), or chromophobe RCC histology (n=16); see FIG. 2A. All of these were collected from patients treated at the Department of Urology, University Hospital Tübingen, Germany. Use of the tissue was approved by the ethics committee of the University of Tuebingen and informed written consent was provided by each subject prior to surgical resection. Surgically resected ccRCC tissues were classified according to the seventh edition of the Union Internationale Contre Ie Cancer/American Joint Committee on Cancer system (2009). None of the patients received any kind of neoadjuvant therapy before surgery, neither immune- nor chemotherapy. Importantly, these samples have been independently evaluated by two teams of pathologists with special expertise in nephro-pathology to have maximum certainty regarding their RCC subtype. C1 was used to identify genes with RCC subtype-specific expression.

RCC cohort 2 (C2) is a combined cohort containing 143 RCC samples from five studies (K. A. Furge et al., Detection of DNA Copy Number Changes and Oncogenic Signaling Abnormalities from Gene Expression Data Reveals MYC Activation in High-Grade Papillary Renal Cell Carcinoma. Cancer Res. 67(7), p. 3171-3176 (2007); M.-H. Tan et al., Genomic Expression and Single-Nucleotide Polymorphism Profiling Discriminates Chromophobe Renal Cell Carcinoma and Oncocytoma. BMC Cancer, 10:196 (2010); S. Peña-Llopis et al., BAP1 Loss Defines a New Class of Renal Cell Carcinoma. Nat. Genet. 44(7), p. 751-759 (2012); M. V. Yusenko et al., High-resolution DNA Copy Number and Gene Expression Analyses Distinguish Chromophobe Renal Cell Carcinomas and Renal Oncocytomas. BMC Cancer 9, p. 152 (2009); T. H. Ho et al., Differential Gene Expression Profiling of Matched Primary Renal Cell Carcinoma and Metastases Reveals Upregulation of Extracellular Matrix Genes. Ann. Oncol. Off. J. Eur. Soc. Med. Oncol. 28(3), p. 604-10 (2017)); see FIG. 3. Common to these studies was the use of the Affymetrix GeneChip HG U133 Plus 2.0 for quantification of gene expression. Only samples from primary tumor tissue that were labeled ccRCC, chRCC or pRCC in the original studies were added to C2. Table S1 shows the numbers of samples per subtype obtained from each study. C2 was used to determine the signature.

TABLE S1

Cohort C2 (n = 143) includes RCC samples from

five studies providing gene expression data on Gene

expression omnibus. Expression measurements were

conducted with Affymetrix microarray HG U133 Plus 2.0.

Study
GEO accession
ccRCC
pRCC
chRCC

Yusenko et al.
GSE11151
26
19
4

Tan et al.
GSE19982
0
0
15

Peña-Llopis et al.
GSE36895
29
0
0

Furge et al.
GSE7023
0
35
0

Ho et al.
GSE85258
14
1
0

Using the established signature matrix, transcriptomes from the TCGA RCC cohort (C3) were deconvolved. Clinical information and gene expression data (“FPKM-UQ″”) generated by RNA-Seq from kidney cancer cohorts KIRC, KICH and KIRP from TCGA were downloaded on Sep. 25, 2019 from https://gdc.cancer.gov/ using R-package TCGAbiolinks. XML-structured clinical information was processed using R-package XML. Disease-specific survival outcome data for the TCGA RCC cohort was obtained from (Liu et al., Cell, 2018) and was referred to as cancer-specific survival (CSS) in this work. Four patients from the KIRC cohort were repre-sented by several samples in the expression data set. The sample with highest median expression was chosen, respectively. To ensure that only tumor samples were included, remaining tumor and non-tumor samples from the TCGA RCC cohort were hierarchically clustered using Ward's method. Three cases (TCGA-BQ-5889, TCGA-CJ-5683, TCGA-DV-5573) were wrongly assigned as tumor tissues were excluded. In case of TCGA-CW-5591 tumor and non-tumor data had been confused. Patients that received prior treatment were excluded. In total the C3 cohort included tumor samples from 864 patients (KIRC: 512, KIRP: 287, KICH: 65), with survival data available for 847 patients.

Analysis and Processing of Gene Expression Data

High quality total RNA was isolated from fresh-frozen RCC tissue from cohort C1 using the mirVana™ miRNA Isolation Kit (Life Technologies) as previously described (P. Fisel et al., DNA Methylation of the SLC16A3 Promoter Regulates Expression of the Human Lactate Transporter MCT4 in Renal Cancer with Consequences for Clinical Outcome. Clin. Cancer Res. 19(18), p. 5170-5181 (2013), S. Winter et al., Methylomes of Renal Cell Lines and Tumors or Metastases Differ Significantly with Impact on Pharmacogenes. Sci Rep. 6(1) (2018). Genome-wide transcriptome analyses were performed using the Human Transcriptome Array HTA 2.0 (Affymetrix) according to the manufacturer's protocol. Further processing of microarray data was performed as previously described (S. Winter et al., loc. cit.). Array quality control was conducted by Affymetrix Expression Console (Build 1.4.1.46). The microarrays from C1 were preprocessed together using the Robust Multiarray Average (RMA) implementation from the R-package oligo and probe sets were summarized on Entrez GeneID level using the annotation for the HTA 2.0 microarray provided by brainarray (http://brainarray.mbni.med.umich.edu, version 23).

Genome-wide transcriptome measurements performed by Affymetrix GeneChip HG U133 Plus 2.0 from 143 RCC patients included in C2 were downloaded from Gene Expression Omnibus (GEO) using R-package GEOquery (Table S1). Microarrays from C2 were normalized individually using the SCAN method from the R-package SCAN. UPC and probe sets were summarized on Entrez GeneID level using the annotation for the GeneChip HG U133 Plus 2.0 microarray provided by brainarray (http://brainarray.mbni.med.umich.edu, version 23).

Entrez GeneIDs were used as gene identifiers in this work. Probesets were summarized on Entrez GeneID level using annotations provided by brainarray (http://brainarray.mbni.med.umich.edu, version 23). Ensemble gene identifiers used in TCGA expression data were mapped to Entrez GeneIDs by means of the org.Hs.eg.db annotation package.

Statistical Tools

All statistical analyses were performed with R-3.6.1 including additional packages beanplot 1.2, MASS_7.3-51.4, partykit_1.2-5, pROC_1.15.3, RColorBrewer_1.1-2, rms_5.1-3.1 squash_1.0.8, survival_2.41.1, and XML_3.98-1.20. GEOquery_2.46.15, oligo_1.48.0, org.Hs.eg.db_3.8.2, SCAN.UPC 2.26.0 SummarizedExperiment_1.14.1, and TCGAbiolinks_2.12.62 are part of the Bioconductor software project (http://www.bioconductor.org). All statistical tests were two-sided. Statistical significance was defined as P-value<0.05.

In hierarchical cluster analyses Euclidean distance and Ward's method have been used if not stated otherwise.

Outcomes

Cancer-Specific Survival (CSS) was Used as an Endpoint in Survival analyses involving cohort C3. CSS time was defined as the time from initial diagnosis to death or last date of follow-up if alive. Data for patients who died from other causes than RCC disease were considered censored at the time of death.

Gene Expression Deconvolution Using Robust Linear Regression

In this work, RCC samples were considered as mixtures of ccRCC, chRCC and pRCC. Further, it was assumed that the proportional composition of the three main subtypes is reflected in the gene expression profile of the mixed samples. According to the linearity assumption (Y. Zhao and R. Simon, loc. cit.), the expression of each gene in a mixed RCC sample can thus be modeled as weighted average of the expression of this gene in ccRCC, chRCC and pRCC. The weights correspond to the respective proportional composition that can be estimated by gene expression deconvolution.

The objective of deconvolution is to find a solution to the system of linear equations: m=S·f. Here, the unknown proportions of ccRCC, pRCC and chRCC in a sample A are modeled by the vector f of coefficients. m represents the vector containing the expression levels of signature genes in A. S is a signature matrix including the expression levels of the signature genes in ccRCC, pRCC and chRCC. Signature genes were defined based on a set of ccRCC, chRCC and pRCC samples that could be uniquely assigned by pathologists or previous analyses of molecular data. The matrix equation can be solved for custom-character using standard linear least squares regression (A. R. Abbas et al., loc. cit.).

To increase stability of subtype assignments robust linear regression as implemented in the “rlm” function from the R-package MASS (parameter maxit was set to 200) was used in this work. Expression deconvolution was performed on linear, i.e. non-log-transformed, expression data as suggested by (Zhong et al., loc. cit.). Further, expression values were centered to zero mean and scaled to unit variance preceding deconvolution. Negative regression coefficients were set to zero and percentages were calculated by dividing the three estimates by their sum, such that c+p+h=100%, with c, p, and h representing the ccRCC, pRCC, and chRCC proportion, respectively.

An permutation P-value was calculated to assess the specificity of the signature for a certain RCC sample. Basically, the P-value computation was carried out in the same way as described in (A. M. Newman et al., Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods 12(5), p. 453-457 (2015).). Briefly, Pearson's correlation coefficient R between m and S·f was compared against a derived null distribution R* custom-character * for sample A. Expression levels in m were replaced by randomly drawn values from the complete transcriptomic data of A, denoted m_i*_i*. Subtype proportions f_i* were determined for m_i* by deconvolution and Pearson's correlation coefficient between m_i*_i* and S·f_i* was calculated. This process was repeated 9999 times, yielding R*, and the P-value was obtained by (|R*>R|+1)/(9999+1).

Selection of Candidate Genes

Samples from C1 were considered as clear cases since they could unambiguously be assigned to one of the main subtypes (FIG. 2A). Transcriptome-wide RNA expression in C1 was quantified using Human Transcriptome Array 2.0 microarray technology and annotated using the brainarray annotation resulting in expression levels for 32749 genes.

Expression data from C1 were sample-wise centered to the median expression level of C1. Genes with median expression below the cohort median in each of the three subtypes were removed. Further, genes not covered in TCGA RNA expression data (https://gdc.cancer.gov/) or by the Human Genome U133 Plus 2.0 Array were excluded. It has been demonstrated using TCGA data that RCC subtypes vary in tumor purity (Yoshihara et al., Nat Commun., 2013), see also http://bioinformatics.mdanderson.org/estimate/. This pattern could be observed in C1 as well (FIG. 2B). To minimize dependence on tumor purity, genes stronger related to tumor purity, as determined by the ESTIMATE method (Yoshihara et al., Nat Commun., 2013) than to tumor type were removed. To be precise, for each gene linear regression models were fit incorporating either tumor type or tumor purity as single predictor or both of them in a multiple regression model. In the latter case models with and without interaction effects were considered. The reduction in the residual sum of squares was compared by analysis of deviance tests. Per variable, i.e. purity or type, main and main+interaction effect were tested and the lower P-value was used. A gene was kept in case its expression in cohort C1 could be better explained by tumor type than by tumor purity (FIG. 2C). 11195 genes remained after these filtering steps (FIG. 2D) and were subsequently tested for subtype-specific expression by analysis of variance. 5881 genes showed significant variation between subtypes after correction for multiple testing using Holm's method. Subsequent pairwise comparisons between subtypes by Tukey's test revealed 3686 genes to be specifically expressed either in ccRCC (1379), pRCC (844) or chRCC (1463). Median expression per subtype was determined for the candidate genes and the minimum absolute log fold change compared to the two respective other entities was calculated using the absolute values of the log fold changes (FIG. 2E). Genes were ordered by decreasing minimum absolute log fold change per subtype and expression levels were transformed to linear space

Selection of the Signature Matrix

Given the set of 3686 tumor-specific genes, a sufficient subset was extracted to uncover the proportional composition of tested RCC samples. Besides distinguishing between histologically distinct ccRCC, chRCC or pRCC cases, the goal was to establish a method also being able to identify heterogeneous tumors. As in similar studies (A. R. Abbas et al., loc. cit., A. M. Newman et al., loc. cit., T. Gong et al., Optimal deconvolution of transcriptional profiling data using quadratic programming with application to complex clinical blood samples. PloS One 6(11), p. e27156 (2011), various signature matrices were created and compared (FIG. 4).

The top custom-character n genes with the highest log fold change per subtype were combined into a signature matrix S_n, i.e. S_nincluded 3×n different genes. Each matrix S_nwas used to perform a subtype prediction in cohort C2 (FIG. 3). n was iterated from 2 to 500 and for n>2 the difference in the subtype assignments between two consecutive signature matrices S_nand S_n-1were calculated for each sample. The maximum absolute difference (MAD) between two assignments for a sample, i.e. max(|c_n−c_n-1|,|p_n−p_n-1|,·|h_n−h_n-1|), was used for this.

FIG. 4A shows the 0.95-quantile MAD between consecutive signature matrices. The final signature matrix was determined using a heuristic approach. Based on the assumption that more included genes allow for a more precise estimation, the largest matrix was chosen, that led to a relevant MAD in the classification (MAD>5%) for a substantial portion of C2. To strengthen the decision basis, different cohort compositions were simulated by subset sampling. In total 10,000 times 50% subsets were randomly drawn from C2. FIG. 4B shows for each matrix S_n the proportions of the sampled subsets that experienced a MAD>5% compared to the previous matrix. S_58, including 174 genes, was the largest matrix significantly modifying a substantial portion of the samples relative to the predecessor matrix (on average 8.5% per sampled subset) and therefore has been chosen as signature matrix.

3. Results
Definition of Gene Signature Matrix for Deconvolution of RCC

Using the 52 RCC from cohort C1 comprising 18 ccRCC, 16 chRCC and 18 pRCC cases, each of which could be uniquely assigned by two independent teams of pathologists, candidate genes for the signature matrix were determined (FIG. 2A). To be shortlisted genes needed to be present in TCGA RNA-Seq data as well as on Affymetrix platforms HTA 2.0 and HG U133 Plus 2.0. Further, median expression had to be above the global median expression in C1 in at least one subtype. By means of TCGA RCC cohorts K. Yoshihara et al. loc. cit. have shown that tumor purity explicitly varies between RCC subtypes which could also be observed in C1 (FIG. 2B). The envisaged signature should be independent of tumor purity to be able to classify beside primary tumor tissue homogeneous tumor cells as well. Therefore, genes more related to tumor purity, as determined by the ESTIMATE method (K. Yoshihara et al. loc. cit.), than to tumor type were excluded (FIG. 2C).

11195 genes remained after filtering and were tested on differential expression between RCC subtypes by analysis of variance (FIG. 2D). Subsequent post-hoc testing using Tukey's method revealed 1379 genes to be specifically expressed in ccRCC, 1463 in chRCC, and 844 in pRCC (FIG. 2E). From this set of specifically expressed genes, signature genes have been selected by evaluating various signature gene matrices. Starting with the genes exhibiting highest minimal absolute log fold change, respectively, iteratively matrices with increasing number of genes were created and applied to deconvolve the 143 samples from cohort C2 (FIG. 3). Robust linear regression was employed for deconvolution. Matrices consisted of median expression values per RCC subtype calculated from C1. Matrix sizes ranged from 6, i.e. top two genes per subtype, to 1500 genes (FIG. 4A). Since RCC samples with defined histological composition were not available, supervised approaches could not be used to define the optimal signature. Therefore, a heuristic criterion was applied: Based on the assumption that resolution of deconvolution increases with the number of genes considered, the largest matrix achieving substantial change in subtype deconvolution compared to its predecessor matrix was chosen. If assigned proportional composition differed by more than 5% for a sample between consecutive matrices the change was considered substantial. Different cohort compositions were simulated by 50% subset sampling from C2 and the fraction of samples affected by substantial change was determined for each tested matrix. In this way, matrix RCC58 including the top 58 genes per RCC subtype was chosen as final signature (FIG. 4B).

Table 1 lists the top 58 genes for determining the pRCC subtype, Table 2 lists the top 58 genes for determining the ccRCC subtype, and Table 3 lists the top 58 genes for determining the chRCC subtype. ‘GeneID’ refers to the identifier that is assigned to a gene record in the ‘Entrez Gene’ database. The ‘Symbol’ column lists the HUGO Gene symbol of the genes. The columns headlined ‘ccRCC’, ‘chRCC’, and ‘pRCC’ list median expression values of the respective signature genes in the indicated RCC sub-types. The expression values are (non-log-transformed) processed signal intensities measured with the Affymetrix HTA2.0 array.

TABLE 1

pRCC-specific

GeneID
Symbol
ccRCC
pRCC
chRCC

164312
LRRN4
39
815
40

6590
SLPI
68
1142
69

10568
SLC34A2
120
1901
51

8842
PROM1
34
628
42

5649
RELN
23
331
24

25928
SOSTDC1
12
210
16

5284
PIGR
69
2130
174

27255
CNTN6
14
157
16

80736
SLC44A4
102
998
65

7348
UPK1B
23
184
23

5950
RBP4
74
558
70

1365
CLDN3
219
1396
215

4316
MMP7
258
1397
114

10103
TSPAN1
231
1215
238

54102
CLIC6
41
184
32

54716
SLC6A20
52
241
69

25805
BAMBI
82
286
67

3595
IL12RB2
32
110
26

346389
MACC1
145
489
76

6372
CXCL6
20
68
21

2239
GPC4
116
378
69

5789
PTPRD
71
256
81

3489
IGFBP6
77
238
71

3918
LAMC2
53
182
60

63917
GALNT11
225
659
209

135932
TMEM139
99
280
88

5069
PAPPA
63
174
55

30811
HUNK
60
164
43

6662
SOX9
180
492
117

83543
AIF1L
105
283
77

54825
CDHR2
64
169
61

144165
PRICKLE1
72
179
69

6659
SOX4
216
534
174

27445
PCLO
51
141
57

54756
IL17RD
53
130
51

4634
MYL3
65
160
66

400566
C17orf97
54
133
55

108
ADCY2
31
75
30

5745
PTH1R
79
184
71

80162
PGGHG
127
293
114

166647
ADGRA3
116
266
90

1999
ELF3
175
398
164

3934
LCN2
55
243
109

81615
TMEM163
54
119
48

3912
LAMB1
429
1017
462

23359
FAM189A1
56
134
62

51435
SCARA3
113
241
88

10317
B3GALT5
32
74
35

5076
PAX2
173
369
143

6565
SLC15A2
41
87
30

54768
HYDIN
33
69
29

94234
FOXQ1
79
166
76

122481
AK7
27
56
22

5364
PLXNB1
106
260
126

131566
DCBLD2
240
489
171

56245
C21orf62
52
105
52

222236
NAPEPLD
82
164
65

152330
CNTN4
44
88
33

TABLE 2

ccRCC-specific

GeneID
Symbol
ccRCC
pRCC
chRCC

1356
CP
1572
18
15

4015
LOX
1128
40
44

112399
EGLN3
1629
126
77

1573
CYP2J2
255
23
19

1038
CDR1
554
51
37

56901
NDUFA4L2
1018
84
101

51129
ANGPTL4
717
72
69

159963
SLC5A12
248
26
26

2487
FRZB
651
69
57

2938
GSTA1
175
18
19

8701
DNAH11
168
17
20

10050
SLC17A4
248
29
20

5350
PLN
161
18
20

23236
PLCB1
414
52
44

29923
HILPDA
805
101
76

1244
ABCC2
167
21
19

29974
A1CF
169
22
16

5054
SERPINE1
581
71
76

6515
SLC2A3
806
107
85

115361
GBP4
448
48
59

4311
MME
221
32
26

2157
F8
315
43
47

2327
FMO2
246
38
21

10203
CALCRL
359
57
56

54578
UGT1A6
345
55
22

8564
KMO
103
18
12

9457
FHL5
132
18
23

7903
ST8SIA4
246
43
37

6513
SLC2A1
818
154
144

80243
PREX2
184
26
36

3486
IGFBP3
3445
683
635

230
ALDOC
320
61
64

5166
PDK4
756
156
115

1009
CDH11
266
55
51

3625
INHBB
247
38
52

1906
EDN1
347
74
50

55076
TMEM45A
110
24
21

6531
SLC6A3
334
67
73

168537
GIMAP7
366
68
80

10964
IFI44L
250
57
29

57493
HEG1
632
133
144

5033
P4HA1
711
162
147

55303
GIMAP4
358
82
73

2113
ETS1
476
111
91

10186
LHFPL6
906
170
212

8519
IFITM1
905
192
222

57561
ARRDC3
1315
326
262

768
CA9
251
63
62

6518
SLC2A5
332
83
55

664
BNIP3
798
201
164

3678
ITGA5
450
114
107

23516
SLC39A14
599
155
133

11067
DEPP1
655
121
170

3910
LAMA4
618
93
160

5256
PHKA2
421
104
110

6925
TCF4
344
81
90

5139
PDE3A
117
29
31

2
A2M
2275
549
597

TABLE 3

chRCC-specific

GeneID
Symbol
ccRCC
pRCC
chRCC

245972
ATP6V0D2
16
18
5252

115111
SLC26A7
12
11
1241

5816
PVALB
26
29
2834

26228
STAP1
12
11
1107

1080
CFTR
11
13
805

253012
HEPACAM2
9
10
578

121506
ERP27
78
67
4154

116449
CLNK
18
18
943

51458
RHCG
44
44
1984

9073
CLDN8
8
8
228

127124
ATP6V1G3
6
6
134

2299
FOXI1
32
37
782

245973
ATP6V1C2
28
29
532

158401
SHOC1
9
9
134

120939
TMEM52B
34
38
593

155006
TMEM213
34
46
698

658
BMPR1B
20
18
300

200958
MUC20
158
113
2314

6521
SLC4A1
59
68
995

7113
TMPRSS2
34
42
560

100820829
MYZAP
21
20
273

525
ATP6V1B1
59
68
865

80157
CWH43
20
21
258

1950
EGF
35
25
421

1160
CKMT2
40
44
464

2888
GRB14
57
57
571

222545
GPRC6A
10
11
111

10512
SEMA3C
94
72
891

9194
SLC16A7
68
78
730

6549
SLC9A2
13
17
155

129684
CNTNAP5
25
28
254

3816
KLK1
47
53
468

27199
OXGR1
14
16
144

23327
NEDD4L
201
213
1651

5618
PRLR
38
30
293

266722
HS6ST3
32
30
240

154091
SLC2A12
19
20
150

51635
DHRS7
203
198
1485

199920
FYB2
12
11
90

55026
TMEM255A
68
91
635

10655
DMRT2
29
33
228

84803
GPAT3
67
74
515

7069
THRSP
41
45
308

53828
FXYD4
46
52
352

202374
STK32A
19
20
135

202151
RANBP3L
13
11
83

26049
FAM169A
16
17
112

51760
SYT17
48
66
429

26249
KLHL3
44
39
277

57533
TBC1D14
220
187
1362

7368
UGT8
93
84
574

54899
PXK
87
97
585

65267
WNK3
21
19
126

64409
GALNT17
37
43
250

135138
PACRG
28
36
205

3606
IL18
114
118
675

5569
PKIA
29
31
179

9068
ANGPTL1
14
14
75

Deconvolution of the TCGA RCC Cohort

RCC58 was used to perform proportional subtype assignment (PSA) by deconvolution of 864 tumor transcriptomes from the combined TCGA RCC cohort, including the KIRC, KIRP, and KICH cohorts. Receiver operating characteristic (ROC) analyses showed very good agreement between PSA and the most recent histological classification of the TCGA RCC cohort (Ricketts et al., Cell Reports, 2018) (FIG. 5). Note that histological classification can still contain errors.

Prognostic Classification of RCC Based on PSA

RCC subtypes vary in prognosis (C. J. Ricketts et al., loc. cit.) (FIG. 6). Hence, the inventors were wondering if PSA estimated by deconvolution were predictive of patient survival as well. Deconvolution assigns three estimates (scores) to each sample representing the proportions of ccRCC, pRCC, and chRCC. The terms “proportions” and “scores” are used interchangeably in the following. Univariate Cox proportional hazard regression with subtype scores as continuous predictors were performed in C3. The scores were modelled via restricted cubic spline functions to detect possible non-linear associations. Highly significant, non-linear relationships to CSS were found for the ccRCC-score and the pRCC-score (FIG. 7). The ccRCC-score exhibited the strongest relationship to patient survival and therefore will be presented here in more detail. Analysis of the fitted curve in FIG. 7 suggested a cubic relationship between ccRCC-score and log relative hazard. This observation could be confirmed by the use of a cubic polynomial, which enabled a similarly good fit (FIG. 8A). FIG. 8B shows the estimated 1-, 2-, and 5-year survival rates in dependence of the ccRCC-score (“ClearScore”). Patients with ClearScore between 20 and 70 had worst prognosis. In particular CIMP but also pRCC type 2 and a few of the ccRCC tumors were in this interval (FIG. 8C). Comparing the prognostic value of ClearScore and histological classification revealed that both provide independent information, however, the ClearScore outperformed the pathological classification in cohort C3 (FIG. 8D).

In FIG. 9 the graph illustrates the estimated relationship between ClearScore and the hazard ratio for endpoint cancer-specific death, using a ClearScore of 0% as reference (i.e., a ClearScore of 0%, the hazard ratio was set to 1). The hazard ratio is calculated by taking the exponential of the log relative hazards from FIG. 8A. For example, a hazard ratio of 3 means that the risk of cancer-specific death is three times higher compared to patients having a ClearScore of 0%. The ClearScore allows a classification of the patients into a “high risk” (top area), “low risk” (bottom area) and “intermediate risk” (middle area) group.

Association Between Survival and Proportional Subtype Assignments Using Different Signature Gene Subsets

A question that arose was whether subsets of the 3×58 (=174) signature genes are already sufficient for determining in a subject's biological sample the relative proportions of ccRCC, pRCC, and chRCC and whether PSA based on these subsets are significantly associated to survival. The inventors proceeded as follows: The 174 genes are composed of the 58 top-specific genes per subtype. Random subsets of size 3×2 (i.e. 6 genes in total), 3×5, 3×10, and 3×20 were drawn from the set of 3×58 signature genes, i.e. the number of randomly drawn signature genes per subtype were identical with each subset. Random sampling was repeated 10,000 times. With each subset a deconvolution of the TCGA cohort (n=864) was performed, followed by ROC analysis (n=819) and survival time analysis (n=847) as done for the complete signature. Log-rank P-values from survival time analysis and area under the curve (AUC) values are shown in FIG. 10. A clear tendency was observed. With increasing subset size P-values from survival time analysis were decreasing and AUC values were increasing. Even for the majority of the 3×2 gene subsets the association with survival (CSS) was significant as well as the AUC value above 0.9. The inventors assume that non-tested subset sizes would match the trend described here.

Computational gene expression deconvolution is performed by solving a linear system of equations using regression methods such as least square regression, support vector regression, or preferably robust linear regression. In order to derive estimates of the three proportions (pRCC, ccRCC, and chRCC), at least three equations in the linear system are necessary, corresponding to three genes in the signature matrix. In case of three genes in the signature matrix, a sufficient condition for the linear system to have a solution is that these equations (i.e. the rows of the matrix) are linear independent. In our method, this condition can be satisfied by an appropriate selection of three genes, each specific in exactly one of the subtypes.

As a consequence, even with only one gene per subgroup type, i.e. from each of Table 1, Table 2, and Table 3, preferably with two genes per subgroup type a reliable subtype classification can be made.

Clustering a RCC Cohort by Using Principal Component Analysis (PCA)

The inventors tested whether signal separation methods other than deconvolution can be used to carry out the invention on the basis of the 3×58 (=174) signature genes. Deconvolution makes it possible to analyze individual samples; a forecast is then made on the basis of the relative proportions. Alternatively, the 174 genes can be used to cluster a comprehensive RCC cohort or make a principal component analysis (PCA) or use other techniques from the field of machine learning to group the data. The clustering obtained could then be used as a reference for new, unknown samples: One measures the 174 genes in a new sample, uses their expression levels to cluster the new sample together with the reference cohort, and finally determines the cluster of the new sample. This can be illustrated by a PCA plot. It shows the result of a PCA with the TCGA cohort based on the 174 signature genes according to the invention. The samples are colored according to their relative hazard ratio. One can spot that the samples with similar hazard ratio cluster together.

Implementation of the Invention into Daily Clinical Routine

Tissue (either fresh, fresh-frozen or FFPE) or body fluids like blood plasma or urine of a patient with RCC is obtained. It may be subjected to the hospital's laboratory or delivered from the hospital or out-patient center to a specialized laboratory. Nucleic acids (total RNA) will be prepared by standard methods. Quantification of the expression levels of candidate genes will be performed using state-of-the art methods. Here different methods like RNA sequencing, microarray or chip based technology or RT-PCR etc. can be used. Based on the established gene signature, (deconvolution-) analysis using well-established algorithms (e.g. robust linear regression) will be performed to determine the proportions of ccRCC, pRCC and chRCC in the sample, subsequently resulting in the outcome classification of RCC patients (low, intermediate, high risk). This report will be delivered to the respective physician, who requested the analyses of RCC specimen.

Example for Making a Diagnosis by the Method According to the Invention

In the following an example is provided in which on the basis of two genes per RCC subgroup a diagnosis is made for a patient suffering from RCC. The tissue sample of the patient is referred to as TCGA-BQ-5894-01A-11R-1592-07.

Random selection of the two top specific genes per subtype results in the reduced signature matrix RCC2:

RCC2
pRCC
ccRCC
chRCC

IL17RD
130
53
51

GALNT11
659
225
209

NDUFA4L2
84
1018
101

KMO
18
103
12

WNK3
19
21
126

RANBP3L
11
13
83

For deconvolution, columns of _RCC2RCC2 are centered to zero mean and scaled to unit variance, resulting in the scaled signature matrix RCC2_z:

RCC2_z
pRCC
ccRCC
chRCC

IL17RD
−0.09
−0.48
−0.68

GALNT11
2.01
−0.04
1.65

NDUFA4L2
−0.28
2.00
0.06

KMO
−0.54
−0.35
−1.25

WNK3
−0.53
−0.56
0.43

RANBP3L
−0.57
−0.58
−0.21

PSA are calculated for sample TCGA-BQ-5894-01A-11R-1592-07 from TCGA KIRC cohort. FPKM-UQ expression values were obtained from https://portal.gdc.cancer.gov/. For the six genes in RCC2_z, the vector m containing these values is given by:

m
TCGA- BQ-5894-01A-11R-1592-07

IL17RD
37063.62

GALNT11
200584.18

NDUFA4L2
230892.48

KMO
5837.22

WNK3
6013.57

RANBP3L
2038.63

As for the signature matrix, values in custom-character are centered to zero mean and scaled to unit variance, resulting in the scaled expression profile m_z:

m_z
TCGA-BQ-5894-01A-11R-1592-07

IL17RD
−0.41

GALNT11
1.13

NDUFA4L2
1.42

KMO
−0.70

WNK3
−0.70

RANBP3L
−0.74

The assumption is that the expression level of gene custom-character in is the sum of its expression in the ccRCC, pRCC and chRCC proportions of sample TCGA-BQ-5894-01A-11R-1592-07. Vector ″ shall denote these unknown proportions of ccRCC, pRCC and chRCC. f is estimated by solving the linear system m_z=RCC2_z·f using robust linear regression. The function “rlm” from the R-package MASS performs a robust linear regression and is applied as follows:

$fit = rlm (m_z \sim RCC2_z, maxit = 200)$

Resulting regression coefficients are accessed via fit$ coefficients:

intercept
pRCC
ccRCC
chRCC

0.02
0.55
0.77
0.03

The intercept is discarded and proportions are calculated by dividing the three estimates by their sum resulting in the following predicted subtype composition of TCGA-BQ-5894-01A-11R-1592-07:

pRCC
ccRCC
chRCC

0.41%
0.57%
0.02

Using the coefficients from the Cox model including the ccRCC-score (ClearScore) as well as its square and cube as predictors the prognostic index custom-character (i.e. log relative hazard) for TCGA-BQ-5894-01A-11R-1592-07 can be calculated, with x=0.57:

$PI = x \times 14.71 - x^{2} \times 25.46 + x^{3} \times 12.21 - 1.46$

Subtraction of 1.46 causes tumors with a ccRCC proportion of 100% to obtain a pi of 0. Hence, the obtained pi of 0.91 denotes the log relative hazard compared to tumors with ccRCC proportion of 100%. e^pi(here 2.48) gives the hazard ratio between TCGA-BQ-5894-01A-11R-1592-07 and the group of tumors with ccRCC proportion of 100%. In addition, with the baseline survival function the probability of survival at a certain time point can be calculated. For the given example the predicted cancer-specific 1-year survival probability is 84% (SE: 81%-87%), the 2-years survival probability is 73% (SE: 67%-77%), and the 5-year survival probability 51% (SE: 42%-58%).

Conclusion and Miscellaneous

The inventors provide for the very first time an objective and reference-free subtype classification or a proportional subtype assignment method for RCC which provides reliable results and is easily applicable in clinical settings.

Although the invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims.

All publications, patents and patent applications mentioned in this specification are herein incorporated in their entirety by reference into the specification, to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention. To the extent that section headings are used, they should not be construed as necessarily limiting.

	Number	Date	Country
Parent	PCT/EP2020/056398	Mar 2020	US
Child	17496641		US

METHOD FOR DETERMINING RCC SUBTYPES

Information

Publication Number

Date Filed

Date Published

Inventors

Original Assignees

CPC

International Classifications

Abstract

Description

Claims

Priority Claims (1)

CROSS-REFERENCES TO RELATED APPLICATIONS

Continuations (1)