Molecular classifier for prostate cancer

Abstract

Described herein is a method for predicting the risk of disease progression in a subject with prostate cancer, the method comprising the steps of: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring values ​​of substantially all of the patient features listed for PRONTO-e or PRONTO-m in Table 6 and some or all of the reference or control features listed in Table 6; c) comparing the patient features with the reference or control features; and d) calculating a prediction score using a classifier that takes the patient feature values ​​as input data, wherein the classifier has been previously trained on samples from a population of early stage prostate cancer patients.

Description

[Technical Field]

[0001] [Related Applications] This application claims priority to U.S. Provisional Application No. 63 / 040,692, filed June 18, 2020, the contents of which are incorporated by reference in their entirety.

[0002] The present invention relates to molecular classifiers, and more particularly to classifiers for prostate cancer. [Background technology]

[0003] Although prostate cancer (CaP) is a leading cause of cancer death, the majority of biopsy-confirmed cases are indolent enough to allow safe observation without curative treatment [1,2]. The most powerful biomarker for advanced prostate cancer is the Gleason grade, determined by comprehensive pathological examination of the surgically resected prostate gland. Low-grade cancers, defined as Gleason grade 3 + 3 = 6 or WHO grade group (GG) 1 [3], carry a negligible risk of metastasis or death [4,5]. High-grade cancers (WHO grades GG2–GG5) require curative treatment. Unlike most cancer types, where grading schemes prioritize nuclear morphology and mitotic count, the GG grading scheme for prostate cancer focuses solely on glandular structure. Both benign prostate and GG1 prostate cancer glands are characterized by a single layer of luminal epithelial cells surrounding a single lumen. All cancer cells occupy a similar environment, directly contacting the lumen at their apical surface, the stroma at their base, and contacting other cancer cells on the remaining four sides. This arrangement is well-suited for similar intake of oxygen and nutrients from surrounding blood vessels. In contrast, high-grade cancers (GG2–GG5) form fused gland-like structures with multiple lumens or no lumens at all. This reflects much greater plasticity in cell-cell interactions, differentiation, and metabolism. The ability to grow in these different configurations corresponds to their ability to grow as metastatic deposits outside the prostate. Thus, cancer metabolism, epithelial plasticity, and epithelial-stromal interactions are important themes in prostate cancer progression [6–9]. Molecular support for the glandular structure associated with GG provides direction for the development of diagnostic biomarkers for advanced prostate cancer.

[0004] Active surveillance (AS) has become the standard of care for GG1 cancer in the United States, Canada, and Europe [10-13]. Patients are monitored with prostate-specific antigen (PSA) levels and serial core biopsies, with or without adjunctive imaging

[10] . While prostatectomy-based GG is highly informative, current methods cannot accurately separate GG1 from GG2 based on needle biopsy, presenting a major dilemma. Due to sampling error and interobserver variability in core biopsies, biopsy grading inaccurately reflects surgical GG in 36–67% of cases [14–17]. These inaccuracies result in men being placed in the wrong risk category. Patients eligible for AS may undergo aggressive surgical intervention (radical prostatectomy), resulting in excessive morbidity due to uncertainty regarding their true risk of developing advanced high-grade cancer. Conversely, some patients may not receive the necessary treatment in time to prevent the spread of incurable metastatic disease.

[0005] Inaccurate reporting of GG at biopsy has motivated molecular approaches to improve risk stratification based on core biopsy sampling for CaP

[18] . However, existing molecular classification methods for biopsy GG cannot accurately distinguish between GG1 and GG2 [19,20]. Summary of the Invention

[0006] In one aspect, a method of predicting risk of disease progression in a subject with prostate cancer is provided, comprising the steps of: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring values ​​of substantially all of 353 patient features, including mRNA and copy number aberration (CNA) features, listed for PRONTO-e in Table 6, as well as some or all of the reference or control features set forth in Table 6; c) comparing the patient features to the reference or control features; and d) calculating a prediction score using a classifier that takes the patient feature values ​​as input, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0007] In one aspect, a method is provided for predicting risk of disease progression in a subject with prostate cancer, the method comprising the steps of: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring substantially all of 94 patient features, including the mRNA, CNA, methylation and clinical features listed for PRONTO-m in Table 6, as well as some or all of the reference or control features set forth in Table 6; c) comparing said patient features to reference or control features; and d) calculating a prediction score using a classifier that takes said patient feature values ​​as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0008] In one aspect, a computer-implemented method for predicting risk of disease progression in a prostate cancer patient is provided, the method comprising the steps of: a) receiving, in at least one processor, data reflecting substantially all of the patient features defined in claim 1 or 7 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors and some or all of the reference or control features listed in Table 6; b) constructing, in the at least one processor, a patient profile based on the patient features; c) comparing, in the at least one processor, the patient profile to a reference or control; and d) calculating, in the at least one processor, a prediction score using a classifier that takes the patient profile as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0009] In one aspect, there is provided a computer program product for use with a general purpose computer having a processor and a memory connected to the processor, the computer program product comprising a computer readable storage medium having encoded thereon a computer program mechanism that may be loaded into the memory of the computer and that causes the computer to perform the method of any one of claims 13 to 15.

[0010] In one aspect, a computer readable storage medium having stored thereon a data structure for storing a computer program product according to claim 16 is provided.

[0011] In one aspect, a device is provided for predicting risk of disease progression in prostate cancer patients, the device comprising at least one processor and an electronic memory in communication with the at least one processor, the electronic memory storing processor-executable code that, when executed by the at least one processor, causes the at least one processor to: a) receive data reflecting substantially all of the patient features defined in claim 1 or 7 and some or all of the reference or control features set forth in Table 6 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors, b) compare the patient features with the reference or control features, and c) calculate, with the at least one processor, a prediction score using a classifier that takes the patient profile as input data, wherein the classifier has been pre-trained on samples from a population of early-stage prostate cancer patients.

[0012] These and other features of preferred embodiments of the present invention will become more apparent from the following detailed description, which refers to the accompanying drawings. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 is a schematic diagram of the approach. [Figure 2] FIG. 1 shows the performance of the top 25 classifiers from iterative cross-validation. [Figure 3(1)] Figure 1 shows the performance of multimodal classifiers PRONTO-e and PRONTO-m. [Figure 3(2)] Figure 1 shows the performance of multimodal classifiers PRONTO-e and PRONTO-m. [Figure 3(3)] Figure 1 shows the performance of multimodal classifiers PRONTO-e and PRONTO-m. [Figure 4] FIG. 1 shows molecular features with significant univariate association with GG (q-value <0.1). [Figure 5] FIG. 1 is a diagram of a computer device for implementing the method. [Figure 6] Schematic diagram of the GG classifier design. [Figure 7] FIG. 1 shows the results of PRONTO-e and PRONTO-m at different operating points. [Figure 8] FIG. 1 shows the similarity between the molecular profiles of low-grade and high-grade samples taken from the same case. [Figure 9] FIG. 1 illustrates the potential clinical impact of PRONTO-e. DETAILED DESCRIPTION OF THE INVENTION

[0014] Figure 1. Overview of the approach (A) Cases were divided into training and validation cohorts. Both high- and low-grade samples were extracted from each resected tumor (i.e., for each case). (B) 431 genes / locuses associated with GG were profiled. (C) A machine learning pipeline was used to develop a GG classifier. First, one or more data types were selected. Second, the associated data were partitioned for five-fold cross-validation. Third (optionally), features without significant univariate correlation with GG were discarded. Fourth, after selecting the machine learning algorithm, the classifier was trained on four partitions and tested on the fifth partition.

[0015] Figure 2. Performance of the top 25 classifiers from iterative cross-validation. Each column represents a classifier. The top panel shows the dataset used by the classifier, the machine learning algorithm used to train it, the sample weighting (i.e., envelope) scheme, and the type of training samples used (see Methods). In the AUC panel, each box summarizes the average AUC from 1000 iterations of cross-validation.

[0016]

number

[0017] The mean statistic is x mean =(x low +x high ) / 2, where x low and x high are statistics calculated from only low-grade or high-grade samples, respectively. Classifiers were ordered by decreasing AUC. Abbreviations: AUC - Area Under the Curve; BCR - Biochemical Recurrence; CAPRA - Cancer of the Prostate Risk Assessment; CN_MLPA - Copy Number, MLPA Platform; CN_NS - Copy Number, NanoString Platform; GG - Gleason Grade; MSP - Methylation-Specific PCR.

[0018] Figure 3. Performance of the multimodal classifiers PRONTO-e and PRONTO-m. (A-C) Multimodal classifiers, i.e., classifiers using different types of data, outperform single-modal classifiers in cross-validation. The TP rate (A), FP rate (B), and AUC (C) for each classifier were calculated from 1,000 cross-validation iterations (boxes summarize the iterations). In each iteration, each statistic was calculated using only high-grade or only low-grade samples for each case. The average of the high-grade and low-grade statistics is shown in the "Means" section. The type of input data used by a given classifier is shown in the key in (C), with CAPRA using only clinical data. The multimodal classifier is the top-performing classifier in cross-validation. (D) Validation performance of multimodal classifiers. For each case in the validation cohort, one sample was randomly selected, and statistics were calculated using a representative sample. This process was repeated 1000 times, and each point represents the median across iterations (i.e., sampling-based AUC), with the lower and upper error bars representing the first and third quartiles, respectively. (A-C) CNA refers to CNA data from MLPA, as PRONTO-e and PRONTO-m only use CNA data from MLPA. (E) Concordance between predicted classes for low- and high-grade samples from the same validation case. (F) Percentage of concordant cases that were correctly predicted.

[0019]

number

[0020] The total number of validation cases used to calculate each percentage is shown above the bar graph. Note that the numbers differ between PRONTO-e and PRONTO-m due to the different data requirements of the classifier for each sample.

[0021] Figure 4. Molecular features with significant univariate association with GG (q value < 0.1). For each significant molecular feature, the plot on the left shows

[0022]

number

[0023] This difference was calculated by randomly drawing 1,000 representative samples per case for each cohort, with the dot representing the median and the ends of the intersection line representing the first and third quartiles. The plot on the right shows the q-value (i.e., adjusted p) obtained by combining the training and validation cohort q-values, representing the significance of the univariate association between the feature and GG (see Methods). For mRNA feature analysis, 332 training and 200 validation cases were used, and for methylation feature analysis, 318 training and 202 validation cases were used. For target genes, preferential expression in the epithelial or stromal compartments has been shown

[54] .

[0024] Figure 5. Computer device for implementing the method A suitably configured computing device, and associated communications networks, devices, software and firmware, to provide a platform for enabling one or more of the embodiments described herein.

[0025] Figure 6. GG classifier design overview The GG classifier takes as input a patient profile, which potentially contains features of different data types (including clinical features, not shown).

[0026]

number

[0027] It is trained with one of several possible machine learning algorithms (see Methods), i.e., the final classifier output is yes or no.

[0028] Figure 7. PRONTO-e and PRONTO-m at different operating points (A) Validation ROC curves of the PRONTO-e and PRONTO-m classifiers for samples of only low-grade or only high-grade cases in each case. The prediction score is the numerical output of the classifier. If the operating point is x, then score >= x predicts a pathological GG >= 2, while score < x predicts a pathological GG1. The curves show the true positive rate and false positive rate at different operating points. (B) Prediction score distributions of the PRONTO-e and PRONTO-m classifiers. The boxes show the score distributions from the classifiers applied to all samples in the training cohort, separated by the GG of their source cases. As expected, for both classifiers, samples from cases with higher GG tend to have higher scores. The red line indicates the selected operating point of 0.5.

[0029] Figure 8. Similarity between the molecular profiles of low-grade and high-grade samples taken from the same case Since PRONTO-e and PRONTO-m use only CNA data from MLPA, CNA refers to CNA data from MLPA. Abbreviation: methyl-methylation.

[0030] Figure 9. Potential clinical impact of PRONTO-e Virtual performance of the PRONTO-e classifier when applied to diagnostic biopsies of 1000 patients for whom active surveillance is recommended. Assuming 1000 active surveillance patients and the predictive performance of PRONTO-e, the figure shows the hypothetical numbers of true positives and false positives, true negatives and false negatives, and how these patient subsets are affected by the test results. If the test result is positive, it triggers an early biopsy 3 or 6 months after diagnosis, resulting in an increase in malignancy and potentially subsequent treatment. If the test result is negative, a biopsy is instead performed 12 months after diagnosis.

[0031] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be understood that the present invention may be practiced without these specific details.

[0032] Cancer grade is the most powerful predictor of disease progression in early-stage prostate cancer (CaP). Intratumor heterogeneity and interobserver variability limit the accuracy of diagnostic biopsies and reduce their clinical utility. Using prostatectomy pathology as the gold standard, we developed and validated robust objective biomarkers of prostate cancer grade.

[0033] Radical prostatectomy patients with low- and intermediate-risk CaP were recruited and assigned to either a training cohort (n = 333) or a validation cohort (n = 202). To integrate intratumor heterogeneity, each case was sampled separately at two sites. We profiled 342 mRNAs enriched for CaP metabolism, stromal signaling, and epithelial plasticity, complemented by 100 copy number abnormalities (CNAs) and 14 DNA hypermethylation loci. Using 12 different machine learning algorithms, clinical, pathological, and molecular variables were applied to generate over 41,000 candidate classifiers (1 vs. ≥ 2) for pathological grade grouping using the training data. We selected two classifiers, PRONTO-e and PRONTO-m, for validation by prioritizing classifiers with higher true positive (TP) rates and areas under the receiver operating curve (AUC).

[0034] The PRONTO-e classifier includes 353 mRNA and CNA features, while the PRONTO-m classifier includes 94 mRNA, CNA, methylation, and clinical features. The classifiers (PRONTO-e and PRONTO-m) were independently validated with true positive rates of 0.802 and 0.810, false positive rates of 0.403 and 0.398, and AUCs of 0.799 and 0.786, respectively.

[0035] Two multigene classifiers were developed and validated in separate cohorts, each achieving superior performance by integrating different types of genomic data. The adoption of the classifiers could improve current active surveillance approaches without increasing patient morbidity.

[0036] In one aspect, a method of predicting risk of disease progression in a subject having prostate cancer is provided, the method comprising the steps of: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring values ​​for substantially all of 353 patient features, including mRNA and copy number aberration (CNA) features listed for PRONTO-e in Table 6, and some or all of the reference or control features set forth in Table 6; c) comparing said patient features to reference or control features; and d) calculating a prediction score using a classifier that takes said patient feature values ​​as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0037] In some embodiments, substantially all of the 353 patient characteristics are all of the 353 patient characteristics.

[0038] As used herein, the term "control" refers to a particular value or data set that can be used for prognosis or classification. For example, patient characteristics including mRNA, copy number aberration (CNA) characteristics, or clinical characteristics obtained from the test sample associated with the outcome class, etc. One skilled in the art will understand that the comparison between the test sample and the control will depend on the control used.

[0039] The terms "low risk" or "low likelihood," as used herein with respect to cancer, refer to a statistically significantly lower risk of cancer compared to the general population or a control population. Correspondingly, the terms "high risk" or "high likelihood," as used herein with respect to cancer, refer to a statistically significantly higher risk of cancer compared to the general population or a control population.

[0040] The term "sample," as used herein, refers to any fluid, cell, or tissue specimen from a subject that can be assayed for DNA or RNA material as referred to herein.

[0041] In one aspect, a method of predicting risk of disease progression in a subject with prostate cancer is provided, the method comprising the steps of: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring substantially all of 94 patient features including mRNA, CNA, methylation and clinical features listed for PRONTO-m in Table 6, as well as some or all of the reference or control features set forth in Table 6; c) comparing the patient features with the reference or control features; and d) calculating a prediction score using a classifier that takes the patient feature values ​​as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0042] In some embodiments, substantially all of the 94 patient biomarkers are all 94 patient biomarkers.

[0043] In some embodiments, determining the predictive score comprises classifying the patient's tumor into a pathological Gleason Grade Group (GG) class.

[0044] In some embodiments, the patient's tumor is

[0045]

number

[0046] A score of <0.5 is classified as pathological GG1 class.

[0047] In some embodiments, if the patient is classified in the pathological GG1 class, the method further comprises managing the patient with active surveillance.

[0048]

number

[0049] It also includes treating the patient with surgery, endocrine therapy, chemotherapy, radiation therapy, hormone therapy, gene therapy, heat therapy, or ultrasound therapy.

[0050] The present system and method can be implemented in various embodiments. A suitably configured computing device and associated communications networks, devices, software, and firmware can provide a platform for enabling one or more of the above-described embodiments. By way of example, FIG. 5 illustrates a general-purpose computing device 100 that may include a central processing unit (“CPU”) 102 connected to storage 104 and random access memory 106. The CPU 102 can process an operating system 101, application programs 103, and data 123. The operating system 101, application programs 103, and data 123 may be stored in storage 104 and loaded into memory 106 as needed. The computing device 100 may further include a graphics processing unit (GPU) 122 operatively connected to the CPU 102 and memory 106 to offload intensive image processing calculations from the CPU 102 and perform these calculations in parallel with the CPU 102. An operator 107 can interact with the computing device 100 using various input / output devices, such as a video display 108 connected by a video interface 105 and a keyboard 115, a mouse 112, and a disk drive or solid state drive 114 connected by an I / O interface 109. In a known manner, the mouse 112 can be configured to control cursor movement within the video display 108 and to operate various graphical user interface (GUI) controls that appear within the video display 108 with mouse buttons. The disk drive or solid state drive 114 can be configured to accept computer-readable media 116. The computing device 100 can form part of a network via a network interface 111, enabling the computing device 100 to communicate with other suitably configured data processing systems (not shown). One or more different types of sensors 135 can be used to receive input from various sources.

[0051] The present system and method can be implemented on virtually any type of computing device, including a desktop computer, laptop computer, tablet computer, or wireless handheld. The present system and method can also be implemented as a computer-readable / usable medium containing computer program code that enables one or more computing devices to perform each of the various process steps in the method according to the present invention. When there are multiple computing devices that perform the entire operation, the computing devices are networked to distribute the various steps of the operation. It should be understood that the terms computer-readable medium or computer-usable medium include one or more of any type of physical embodiment of the program code. In particular, the computer-readable / usable medium can comprise program code embodied on one or more data storage portions of a computing device, such as memory associated with a computer and / or storage system, on one or more portable storage products (e.g., optical disks, magnetic disks, tapes, etc.).

[0052] In one aspect, a computer-implemented method for predicting risk of disease progression in prostate cancer patients is provided, the method comprising the steps of: a) receiving, with at least one processor, data reflecting substantially all of the patient features defined in claim 1 or 7 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors and some or all of the reference or control features listed in Table 6; b) constructing, with the at least one processor, a patient profile based on the patient features; c) comparing, with the at least one processor, the patient profile to a reference or control; and d) calculating, with the at least one processor, a prediction score using a classifier that takes the patient profile as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0053] In one aspect, there is provided a computer program product for use with a general purpose computer having a processor and a memory connected to the processor, the computer program product comprising a computer readable storage medium having encoded thereon a computer program mechanism that may be loaded into the memory of the computer to cause the computer to perform the method of any one of claims 13 to 15.

[0054] In one aspect, a computer readable medium having stored thereon a data structure for storing a computer program product according to claim 16 is provided.

[0055] In one aspect, a device for predicting risk of disease progression in prostate cancer patients is provided, the device comprising at least one processor and an electronic memory in communication with the at least one processor, the electronic memory storing processor-executable code that, when executed by the at least one processor, causes the at least one processor to: a) receive data reflecting substantially all of the patient features defined in claim 1 or 7 and some or all of the reference or control features set forth in Table 6 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors, b) compare said patient features with the reference or control features, and c) calculate, with the at least one processor, a prediction score using a classifier that takes said patient profile as input data, wherein the classifier has been pre-trained on samples from a population of early stage prostate cancer patients.

[0056] The advantages of the present invention are further illustrated by the following examples. The examples and their specific details described herein are presented for illustrative purposes only and should not be construed as limiting the scope of the invention claimed.

[0057] example material and method Patient samples:

[0058] To train and validate the classifier, radical prostatectomy samples were identified using local electronic medical records from Kingston General Hospital (diagnosis 1999–2012), Montreal General Hospital at McGill University Health Centre (1994–2013), and London Health Sciences Centre (LHSC) (2004–2009). Initial inclusion criteria were (i) core biopsy-confirmed diagnosis of GG1 or GG2, (ii) radical prostatectomy, and (iii) no prior treatment. Patients with clinical stage T3 or higher were excluded. Cases were assigned to either the training or validation cohort.

[0059] For all cases, central pathology review of both diagnostic core biopsies and radical prostatectomy was performed by expert pathologists (FB, MM, DB, TJ). When possible, DNA and RNA were extracted from punch cores obtained from two regions of the dominant tumor focus, enriched for relatively high and low GG regions, when present (Figure 1A)

[21] , using protocols optimized for this approach [22, 23]. All analyses performed were approved by the local ethical review board, which allowed for a waiver of informed consent (Table 3). Overall, we collected 633 samples from 333 cases for the training set and 346 samples from 202 cases for the validation set (see CONSORT data in Table 4).

[0060] The clinicopathological characteristics of the training and validation cohorts are summarized in Table 1.

[0061]

number

[0062] The power to validate the two classifiers (α = 0.01) was 89%.

[0063] Selecting candidate features for the classifier: We examined molecular features at the transcriptome (mRNA abundance), genome (DNA copy number alterations, CNAs), and epigenome levels (DNA methylation) for multiple functional aspects reflecting the biology of GG (Figure 1B). A list of 462 molecular features assessing 431 genes / locuses (each gene / locus may be assessed by multiple features) was compiled through a detailed literature search and input from numerous studies led by members of our research team [25-30] (see Methods; Table 6). We also included four clinical features assessed at diagnosis and a fifth clinical feature that combined them into Cancer of the Prostate Risk Assessment (CAPRA) risk groups

[31] . In total, 467 features were used to describe tumor samples (Table 6).

[0064] Centralized molecular profiling: We employed four molecular diagnostic platforms, three of which are currently in clinical use for the molecular diagnosis of cancer. mRNA analysis was performed using the Nanostring N-counter platform

[32] with specific code sets developed for this study. CNA analysis was performed using both multiplex ligation-dependent probe amplification (MLPA)-based assays and custom NanoString copy number code sets

[33]

[34] developed specifically for this project (Ebrahimizadeh et al., submitted). Finally, epigenetic profiling was performed using methylation-specific polymerase chain reaction (MSP)

[26] . All samples from both cohorts were profiled on as many platforms as possible given their RNA and DNA yields.

[0065] Development and validation of prognostic classifiers: Both the training and validation data were preprocessed as described in the Supplementary Methods. A supervised machine learning pipeline (Fig. 1C; Supplementary Methods) was created to develop a classifier using patient profiles (composed of feature values) as input and pathological prostatectomy GG as the endpoint.

[0066]

number

[0067] Using the training data, over 41,000 GG classifiers were evaluated by subjecting selected features to 12 different machine learning algorithms in a five-fold cross-validation. Specifically, the area under the receiver operating curve (AUC), TP, FP, and true negative (TN) rate were calculated for each classifier. This set of metrics was calculated using only low-grade or high-grade samples from each case, and the average of the low-grade and high-grade statistics was calculated. Two classifiers were selected for validation by prioritizing those with higher TP rates and AUCs through cross-validation.

[0068] We validated the classifier by calculating statistics as described above and by randomly selecting one sample (high-grade or low-grade) per patient in the validation cohort to calculate performance statistics, repeating this process 1,000 times. These sampling-based statistics better simulate clinical practice. All statistical analyses were performed using the R software framework (v3.4.3)

[35] , the machine learning package mlr (v2.15.0)

[36] , and the plotting package BoutrosLab.plotting general (v5.9.8)

[37] .

[0069] Ethical review All studies were conducted in accordance with the Tri-Council Policy Statement (TCPS2) and after obtaining ethical approval for the research protocol from the research ethics committee of each participating institution (Table 3).

[0070] Feature Selection CNA characteristics: MLPA assay

[0071] A multiplex ligation-dependent probe amplification (MLPA) assay was developed to evaluate 14 loci for copy number alterations (CNAs; Table 6 ) previously associated with clinical outcome in prostate cancer (CaP; Ebrahimizadeh et al., submitted). The assayed loci included the MYC oncogene [S1-3], PTEN [S4-7], TP53 [S2,8,9], CDKN1B [S10,11] and RB1 [S12,13] tumor suppressors, metastasis-related loci such as GABARAPL2 [S13,14] and PDPK1 [S15,16], loci related to maintaining genomic stability such as RWDD3 [S17-20], GTF2H2 [S21-24] and WRN [S13,25-27], and CaP subtype-associated genes CHD1 [S13,28,29], MAP3K7 [S13,28,30], NKX3-1 [S13] and PDZD2 [S31,32].

[0072] CNA Features: CPC-GENE NanoString Assay Using DNA CNA assays, the Canadian Prostate Cancer Genome Network (CPC-GENE) identified an association between genomic alteration rates and decreased biochemical recurrence-free survival in patients with low- to intermediate-risk CaP and developed a classifier that uses CNA features to predict patient outcomes [S33]. The NanoString CNA assay was designed to derive value for these features [S34], and herein we used an assay containing 92 CNA features: 85 loci (including 151 genes) and 7 additional genes associated with CaP in the literature (Table 6).

[0073] mRNA characteristics: We created an mRNA abundance gene panel (for the NanoString RNA assay) by combining gene lists from the following studies:

[0074] mRNA Features: CPC-GENE CPC-GENE performed RNA abundance profiling of samples from intermediate-risk patients [S35], and univariate analysis of these data identified 20 genes associated with poor prognosis. These genes were supplemented with 30 genes identified by Taylor et al. [S36] in a similar univariate analysis and predictive modeling of RNA data.

[0075] mRNA signature: stem cell signature The gene list was derived by "reprogramming" four androgen receptor (AR)+ CaP cell lines (LNCaP, LAPC4, CWR22rv1, and VCaP) to a stem-like phenotype [S37]. Agilent GeneChip analysis of each cell line revealed transcripts with significant abundance changes between parental and reprogrammed cells. These transcripts were then compared between cell lines to derive a ranked list of 132 commonly altered genes associated with reprogramming. From this signature, trends for recurrence, metastasis, and CaP-specific mortality were identified as described in [S37]. The top 50 genes from this list were included in an RNA panel.

[0076] mRNA signatures: hallmarks of epithelial-mesenchymal transition (EMT) Using the GEO2R program and the Benjamini-Hochberg method for multiple testing correction, we compared gene expression data from PC-3, PC-3M, ALVA-31, and RWPE-2-w99 cell lines undergoing invasive growth in 3D culture (GEO#GSE19426) [S38] and identified 1,669 genes dysregulated in at least three of the four cell lines. These genes were cross-referenced with EMT-related genes from the SABiosciences qRT-PCR array. The resulting 33 overlapping genes were used as a seed list for network construction using String v9.1 and the GeneMania algorithm [S39,40]. From the resulting network, 37 significant genes containing common nodal points connecting pathways were included in the RNA panel.

[0077] mRNA signature: Stroma influence on epithelial growth and differentiation. A list of 318 genes identified as enriched in embryonic prostate stroma [S41-43] was filtered to enrich for genes also expressed in cancer-associated fibroblasts and for association with clinical and pathological endpoints (recurrence, CaP death, and Gleason score) in four published datasets [S36, 44-46]. A list of 80 genes was created by prioritizing genes associated with Grade Group (GG) and / or recurrence across multiple datasets.

[0078] mRNA signature: Tumor cell metabolism Using String v9.1 and the GeneMania algorithm [S47], in silico gene network analysis linking the sterol regulatory element-binding protein 1 (SREBP1), insulin growth factor (IGF), AR, and suppressor of cytokine signaling 1 (SOCS1) signaling pathways identified 86 candidate genes related to CaP metabolism. Expression analysis of these genes was performed using the Nanostring nCounter assay in the discovery and validation cohorts. Each cohort contained 32 Gleason pattern 3 and 32 Gleason pattern 4 foci from individual tumors. Univariate analysis using the Mann-Whitney U test (p<0.05) identified 25 differentially expressed genes.

[0079] mRNA signature: prostate homeostasis This study utilized benign prostate homeostasis as a model for steroid hormone-driven growth and differentiation and the dysregulation of these pathways in CaP. Transcripts represented in this series of studies included FER, PTK2, FLT1, LYN, SRC, JAK1, JAK3, MARK3, STAT3, STAT5A, EDF1, WNT11, ITGAV, ITGA2, and ITGV5.

[0080] Methylation and mRNA characteristics: CpG island hypermethylation Genes (n=14) whose CpG islands were hypermethylated in CaP were identified from the literature, and DNA methylation of these genes was assayed using methylation-specific PCR as described in [S48] to obtain values ​​for their methylation signature (Table 6). These genes (except UCHL1) were also added to the RNA panel, along with seven additional epigenetic modification and regulatory genes: DNMT1, EZH2, HDAC1, HIC1, KCNK2, SRP14, and TERT.

[0081] In summary, collating genes from each of these studies yielded a novel NanoString mRNA panel containing 342 genes (see Table 6) with additional housekeeping genes (see Supplemental Methods). NanoString assays were used to measure the mRNA abundance of each gene and determine mRNA signature values.

[0082] Clinical characteristics The Cancer of the Prostate Risk Assessment (CAPRA) score is calculated using five clinical features: 1) age at diagnosis, 2) PSA at diagnosis (units: ng / ml), 3) biopsy GG (i.e., clinical GG), 4) clinical T stage, and 5) percentage of biopsy core involved in cancer [S49]. A patient's CAPRA score can be used in turn to assign a CAPRA risk group (low, intermediate, or high), and our candidate prognostic classifier optionally used features from this group. Alternatively, the first four clinical features can be used directly by the classifier. If age at diagnosis was unavailable, age at radical prostatectomy (if available) was used. If PSA at diagnosis was unavailable, preoperative PSA (if available) was used.

[0083]

number

[0084] Clinical stage T was simplified to two values, T1 and T2, which were represented to the classifier as 0 and 1, respectively.

[0085] Preprocessing training and validation data mRNA abundance data To select which normalization method to use, we tried 96 different methods supported by the NanoStringNorm R package (v1.1.22; [S50]) with different combinations of parameter values: Background={none, mean.2sd, max}, CodeCount={none, sum, geo.mean}, SampleContent={none, housekeeping.sum, housekeeping.geo.mean, total.sum, top.mean}, OtherNorm={none, rank.normal}. The remaining parameters were left at their default values: round.values=FALSE, take.log=TRUE. To evaluate each normalization method, we calculated several metrics using the resulting normalized data. These metrics include:

[0086] 1) Pass if the normalized counts of low-abundance housekeeping genes are significantly lower than the normalized counts of medium-abundance housekeeping genes and are similarly lower for medium-abundance genes compared to high-abundance genes (one-tailed Student's t-test P<0.05); otherwise fail.

[0087] 2) Dynamic range is measured as the percentage increase in mean normalized counts of high abundance housekeeping genes relative to the mean value of low abundance housekeeping genes.

[0088] 3) Agreement between normalized counts of replicate control samples between cartridges, where higher values ​​suggest less batch effect.

[0089] 4) Number of non-normal samples. If the distribution of normalized counts across endogenous genes does not pass the Shapiro-Wilk test for normality (FDR adjusted q<0.1), the sample is non-normally distributed.

[0090] 5) Number of significant cohort covariates, i.e., genes for which patient origin (Kingston General Hospital / Montreal Hospital at McGill University Health Centre) is a significant covariate in a linear model predicting normalized counts, where GG and biochemical recurrence status are other covariates (FDR-adjusted p<0.1).

[0091] 6) Correlation of the normalized total count of a sample with the age of its source tissue block.

[0092] 7) Percentage of rejected samples; a sample may be rejected if: a) Normalized counts of housekeeping genes = 0. b) After calculating Z-scores on the normalized counts of housekeeping genes, any |Z|>5. c) If CodeCount normalization is performed, the normalization factor is <0.3 or >3. d) The sample has abnormal background levels (|Z|>5). e) RNA content value <1 when SampleCount normalization was performed. f) When SampleCount normalization is performed, the sample has an outlier RNA content value (|Z|>5). g) The deletion rate of the endogenous gene is greater than 0.9 (>0.9),

[0093]

number

[0094] We ranked methods by first ranking them individually by metrics 2–7, considering only methods that passed metric 1, had intercartridge agreement greater than 0.9, and failed less than 10% of the training samples. We then generated a consensus ranking using the DECOR method (ConsRank package v2.0.1; [S51]). Based on this ranking, we selected the following normalization method: Background=none, CodeCount=none, SampleContent=housekeeping.sum, target value=5000 (roughly estimated based on the training data), and OtherNorm=none.

[0095] MLPA CNA data One or two probes targeted each gene, and each test sample was assayed in duplicate. For each replicate, the signal from each test probe was divided by the signal from each of the 10 reference probes, resulting in a set of seven ratios. A probe was considered positive for a CNA if the 95% confidence interval of the replicate ratios fell outside the 95% confidence interval of that probe in at least two of the three reference samples (fresh healthy female genome, normal FFPE kidney tissue, and normal FFPE breast lymph node tissue) (Promega). A probe was considered positive for a test sample if it was positive for both of its replicates. If there was a discrepancy between replicates, the probe was considered negative for a CNA. If none of the replicates passed quality control (Ebrahimizadeh, submitted), no CNA status was assigned to a given probe in a given test sample. If all probes for a gene were positive, the gene was considered positive for a CNA in the test sample; if there was a discrepancy, the gene was considered negative; otherwise, no CNA status was assigned. Only deletions were considered for the RWDD3, GTF2H2, CHD1, MAP3K7, NKX3-1, WRN, PTEN, CDKN1B, RB1, GABARAPL2 and TP53 genes, and only gains were considered for the MYC, PDPK1 and PDZD2 genes.

[0096] NanoString CNA Data Data were preprocessed as previously described [S34].

[0097] Methylation data C q was calculated as previously described [S48]. For a given test sample t and target gene g, the methylation level was calculated as follows: m t,g,i,j,k,l =(C q p,g,i -C q p,r,j )-(C q t,g,k -C q t,r,l ) where: p indicates a positive control sample on the same plate as the test sample, r denotes the reference sequence (ALU), i, j, k, l indicate the number of replicates.

[0098] The normalized methylation level was then defined as follows: m t,g =center i,j,k,l (m t,g,i,j,k,l )

[0099] Machine learning pipeline for development of prognostic classifiers We have constructed a pipeline to comprehensively evaluate various methodologies for developing prognostic classifiers. Specifically, the pipeline uses supervised machine learning techniques to develop classifiers that predict good or poor prognosis (i.e., test negative and positive, respectively) using patient profiles as input data. In this study, we binarize GG into prostatectomy specimens (i.e., pathological GG), and patients with only GG1 are considered the negative gold standard, and

[0100]

number

[0101] The true classes of patients were defined (Supplementary Figure 1).

[0102] The pipeline consists of four main stages: 1) dataset, 2) partition, 3) feature reduction, and 4) cross-validation (Figure 1C).

[0103] The first stage focuses on preparing a training dataset. The training dataset includes a matrix of patient-sample features (i.e., each row represents a patient profile) and a set of true class values, one value for each sample in the matrix. The pipeline can take input data generated by different platforms. In this application, we have clinical / CAPRA, RNA abundance, MLPA / NanoString CNA, and methylation data. For each platform, this stage reduces the dataset to samples with no missing data. If multiple platforms are desired, the dataset is also reduced to samples with data from each platform of interest. Finally, invariant features, i.e., features that have the same value across all remaining samples, are removed from the dataset.

[0104] The second stage focuses on dividing the training dataset for repeated cross-validation. Depending on the desired option, the dataset is reduced to only low-grade samples, only high-grade samples, or randomly selected samples for each patient. By default, this stage prepares a five-fold cross-validation that is repeated 1000 times, thus creating 1000 divisions of the dataset into five equally sized subsets. For each candidate partitioning, each sample is initially randomly assigned to one of the five subsets. Because our training samples were obtained from different institutions (i.e., Kingston General Hospital and Montreal Hospital of McGill University Health Center), the partitioning will be retained if it is balanced with respect to true class, biochemical recurrence status (which may be related to the true class in this application), and sample origin. Specifically, for each pair of subsets in the partitioning, a two-tailed Fisher's exact test is used to test the association with each feature. If any of the potential associations is significant (p<0.05), another candidate partitioning is generated until a balanced one is obtained.

[0105] The third stage focuses on feature reduction. For x-fold cross-validation, each partition allows for x training subsets. In this stage, invariant features, i.e., features with the same value across all samples, are removed from each training subset. If desired, each remaining feature is then tested for univariate association with the true class (e.g., using a two-tailed Mann-Whitney U test). Features with significant association (e.g., P<0.01 or 0.05) are retained.

[0106] The fourth step involves performing iterative x-fold cross-validation using the desired machine learning algorithm using the mlr package v2.15.0 [S52] (Figure 6). Algorithm options (mlr implementation identifier in parentheses) include decision trees (classif.rpart), flexible discriminant analysis (classif.earth), GLM with lasso or elastic net regularization (lasso or elastic net regularization), cross-validation lambda (classif.cvglmnet), k-nearest neighbors (classif.kknn), linear discriminant analysis (classif.lda), logistic regression (classif.logreg), naive Bayes (classif.naiveBayes), nearest neighbor shrinkage centroid (classif.pamr), quadratic discriminant analysis (classif.qda), random forests (classif.ranger), regularized discriminant analysis (classif.rda), and support vector machines (classif.svm). Regardless of the choice of algorithm, cross-validation iterations are performed using unweighted samples (i.e., all samples are weighted equally by default).

[0107] For algorithms that support sample weighting, this stage cross-validates different weightings of the negative / positive gold standard classes, i.e., 30% / 70%, 40% / 60%, 50% / 50%, 60% / 40%, 70% / 30%. Specifically, w n % / (100-w n )% weighting is given to negative and positive samples, respectively. n / p n and (100-w n ) / (1-p n ) weights are assigned, where p n is the proportion of samples in the negative gold standard class. Therefore, the total weight of all negative samples is the overall w n %, and the total weight of all positive samples is the overall (100-w n )%. For all other machine learning algorithm parameters, default values ​​are used.

[0108] In cross-validation, a classifier is trained on x-fold (x-1) using a given machine learning algorithm, dataset (prepared in the previous step), and sample weights. If this training fails after three attempts, the pipeline skips to training on the next (x-1) fold of data. If successful, the resulting classifier is tested against two perspectives: i) only low-grade samples from each case, and ii) only high-grade samples from each case on the remaining fold of data. For each perspective, the pipeline calculates the area under the receiver operating curve (AUC) averaged over the x folds, using an operating point of 0.5.

[0109]

number

[0110] The proportion of true positives (TP), false positives (FP), and true negatives (TN) are calculated for all patients in the x-category. Furthermore, for each of these statistics, the pipeline evaluates two aspects [e.g., AUC mean =(AUC low +AUC high ) / 2]. Finally, the pipeline further summarizes by computing a median statistic across the cross-validation iterations (e.g., across 1000 partitions).

[0111] The validation pipeline for the Grade Group classifiers PRONTO-e and PRONTO-m allowed us to thoroughly test all possible methods supported by the pipeline, thereby exhaustively searching for the optimal method. The selection of methods for validation was driven by two main factors. First, we desired methods with large AUC values ​​from cross-validation, as they suggest a better overall performance of the corresponding classifier. Second, following consultation with clinicians who prioritized earlier intervention for some GG1 cases (quantified by FP rate) at the expense of overtreating these cases, we chose methods with larger AUC values ​​from cross-validation, as they suggest a better overall performance of the corresponding classifier. Second, following consultation with clinicians who prioritized earlier intervention for some GG1 cases (quantified by FP rate), we chose methods with larger AUC values ​​from cross-validation, as they suggest a better overall performance of the corresponding classifier.

[0112]

number

[0113] The top 25 performing classifiers have AUCs ranging from 0.772 to 0.790 (Figure 2), with the majority of them using either regularized discriminant analysis or support vector machines. PRONTO-m is the only classifier among the top 25 that satisfies the TP rate constraint (TP rate = 0.800, AUC = 0.774). PRONTO-e (TP rate = 0.833, AUC = 0.770) was selected for validation. Table 5 describes the methods used to generate these two classifiers.

[0114] Each selected method was then used to train a classifier using an undivided training cohort restricted to patients with the required sample and feature data. As with cross-validation, the mean AUC, TP rate, and FP rate were calculated. Here, the mean values ​​are for low-grade samples only and high-grade samples only. Despite known intratumor heterogeneity [S53], it is unclear to what extent the grade of a biopsy sample at diagnosis represents the overall grade of the entire tumor. To better mimic this clinical scenario, for each patient in the validation cohort, one sample was randomly selected, statistics were calculated using a representative sample, and this process was repeated 1,000 times. Across these repetitions, the median AUC, TP rate, and FP rate were calculated (i.e., sampling-based statistics).

[0115] Similarity between molecular profiles In this analysis, we calculated the similarity between molecular profiles of samples from the same patient (i.e., the similarity between low-grade and high-grade sample profiles), so patients with only one sample were excluded. For all platforms, we considered only profiles with no missing values ​​(for any feature). For CNA profiles, we first limited the profiles to features from the MLPA platform because the validated classifier only used CNA features from this platform. We defined pairwise similarity between CNA profiles as the proportion of features in both samples with the same CNA status (i.e., altered or unchanged). For RNA abundance and methylation profiles, we defined pairwise similarity as the concordance coefficient between features.

[0116] Univariate feature analysis

[0117]

number

[0118] The p-values ​​from statistical tests were adjusted across all features from the same platform using the Benjamini-Hochberg method (resulting in q-values). The sampling procedure and subsequent statistical calculations were repeated 1000 times, allowing the calculation of the median, first quartile, and third quartile values ​​across the repetitions. This feature analysis was performed separately for the training and validation data. To estimate the significance of the univariate association of a given feature across both cohorts, the weighted Z method was used to combine the median q-values ​​from each cohort, weighting each q-value by the number of cases used in the calculation [S54].

[0119] result Cohort / Sample Overview: Across the training and validation cohorts, we successfully generated 954 mRNA, 845 NanoString-CNAs, 794 MLPA-CNAs, and 847 methylation profiles for samples from 535 prostatectomy cases. We also generated CAPRA scores for 492 cases.

[0120] Development and validation of the GG classifier The classifier was trained on 333 cases from two sites, with 202 cases from a third site reserved for independent validation (Table 4).

[0121]

number

[0122] The clinical need for early intervention took priority over specificity, resulting in the selection of the top two best-performing classifiers, PRONTO-e and PRONTO-m, for validation (Table 5). For cases with GG > 2 samples, both classifiers were trained using only the high-grade samples from that case. Performance statistics (by AUC) for the top 25 best-performing classifiers are shown in Figure 2. PRONTO-e uses 353 features, including 342 mRNA abundance and 11 CNA features (Table 6), as well as a random forest. PARSE-m uses fewer features (94 in total) but has more available data categories (64 mRNA, 14 CNA, 12 methylation, and 4 clinical (Table 6)) and uses a support vector machine. Performance statistics calculated using only low-grade or high-grade samples from each case, as well as the average low-grade and high-grade statistics, are shown in Figure 3A-C and Table 2.

[0123] Despite reported intratumor heterogeneity in prostate cancer

[38] , performance statistics were remarkably stable when calculated using one randomly selected sample per case (Figure 3D). This process mimics sampling error in biopsies, and the validation performance of both classifiers exceeded that of previously validated adverse pathological biomarkers [19, 20] (Table 2).

[0124] The validated classifier frequently provided consistent GG classification between paired samples from the same case, i.e., 70.8% for PRONTO-e and 73.9% for PRONTO-m, demonstrating a high degree of tolerance to sampling error.

[0125]

number

[0126] This tendency was also observed for PRONTO-m (Figure 3F).

[0127] Molecular Features of Grade Groups We investigated which molecular features were most strongly associated with GG. Univariate analysis revealed that the abundance of 22 transcripts and methylation at 9 loci were significantly associated with GG (adjusted p < 0.1, see Methods; Figure 4). Where cell-type-specific expression patterns could be identified, some transcripts were associated with preferential expression in the epithelium or stroma

[39] . Similar preferential expression rates were observed for the stromal and epithelial compartments. Similarly, the associations between each molecular feature and high GG were similar in both positive and negative proportions. Interestingly, although no significant univariate associations with GG were identified for CNA features, their inclusion in the multivariate classifier for GG improved performance (Figure 3C).

[0128] Multimodal classifiers outperform CAPRA in cross-validation The CAPRA score represents the current clinical standard for prostate cancer prognosis and is calculated using only non-molecular features such as age at diagnosis and biopsy GG [S49]. Importantly, both the PRONTO-e and PRONTO-m classifiers outperformed the CAPRA classifier in cross-validation, with higher TP rates and AUCs (Figures 3A,C).

[0129] GG classifier and intratumor heterogeneity ROC curves calculated using only low-grade or high-grade samples from each case in the validation cohort show differences in classifier performance depending on the grade of the sample relative to the overall tumor grade (Figure 7A). The ROC curves for the PRONTO-m classifier are more divergent than those for the PRONTO-e classifier.

[0130]

number

[0131] It is wider in PRONTO-m versus PRONTO-e ( Figure 7B ).

[0132] The potential impact of intratumor heterogeneity on the validated classifier was examined by comparing input profiles (DNA, RNA) of samples taken from the same case.

[0133]

number

[0134] However, for both CNA and RNA data, the median similarity was greater than 0.9, regardless of the GG subset (Figure 8), indicating that these molecular input profiles were fairly consistent within cases.

[0135] Discussion Here, we report the development of a GG classifier and validation of the PRONTO-e and PRONTO-m classifiers in an independent patient population. These results suggest that incorporating diverse molecular (e.g., mRNA and CNA) features can add significant value (Figure 3C).

[0136]

number

[0137] Both PRONTO-e and PRONTO-m represent significant improvements over current approaches. Three commercially available biomarker tests are designed to inform the management of early-stage CaP at the time of diagnosis on biopsy tissue

[40] . Prolaris uses RNA expression data of cell cycle progression genes in combination with clinical / pathological parameters (Myriad Genetics) to report 10-year risk of prostate-specific mortality

[41] . Given that CaPs are typically diagnosed between the ages of 50 and 65, and the majority of deaths occur 20–25 years after diagnosis

[42] , Prolaris may not be well suited to determining pre- and post-AS. OncotypeDXprostate (Genomic Health), a 17-gene qPCR-based test, and ProMark (Metamark Genetics), a quantitative in situ proteomics test [22, 43], are

[0138]

number

[0139] These intermediate-risk patients fall into a gray area when selecting AS. Adding the OncotypeDx Genomic Prostate Score (GPS) to the CAPRA clinical and pathological nomogram resulted in a slight improvement in the AUC for adverse pathology (AUC = 0.67) compared with CAPRA alone (AUC = 0.63) [20, 44]. ProMark performed slightly better, increasing the AUC for "favorable pathology" alone from 0.69 at biopsy

[19] to 0.75 when used only in patients classified as favorable by the NCCN (National Comprehensive Cancer Network) guidelines [2, 45].

[0140]

number

[0141] Both OncotypeDx and ProMark have reported robustness to tumor heterogeneity [19,20]. These results suggest the presence of measurable clonal changes that mediate CaP aggressiveness, reflect tumor-wide GG, and are consistently present across regions of phenotypic tumor heterogeneity [46,47]. In this study, we derived and independently validated two novel GG classification methods that demonstrated robustness to tumor heterogeneity and sampling-based AUCs of 0.799 (PRONTO-e) and 0.786 (PRONTO-m).

[0142]

number

[0143] PRONTO-e includes 353 features divided into mRNA abundance and DNA CNA type. The more compact PARSE-m includes 94 features divided into mRNA abundance, DNA CNA, and DNA methylation type, and includes preoperative clinical and pathological features (age, clinical stage, PSA, and biopsy GG). Both classifiers are more tolerant to sampling error, although GG is derived from prostatectomy tissue, which is the most accurate. Therefore, when used on biopsy tissue, they are likely to provide better information for AS versus clinical management decisions. Work is currently underway to validate the classifiers using biopsy samples from a statistically powered cohort.

[0144] When OncotypeDx and Prolaris are administered to the same patient, they often yield conflicting recommendations.

[48] Nevertheless, the tests have shown potential to reduce biopsy frequency and overtreatment.

[40] This suggests that more accurate testing has similar, if not superior, potential impact. If the performance of PRONTO-e and PRONTO-m is validated on core biopsies, these assays could dramatically improve this impact. It is relatively straightforward to model each validated classifier on diagnostic biopsies of 1,000 hypothetical men selected for AS, with an estimated 33% of these men expected to upgrade during AS.

[49]

[0145]

number

[0146] Of those who test positive (534 / 1000 men), 267 were TP, likely benefiting from early repeat biopsy and treatment. Of the 466 men who test negative, only 13.5% (63) would be false negatives. For the 26.7% of all FP-identified cases, we suggest performing an earlier initial AS biopsy rather than additional biopsies. Early biopsy for these patients would provide pathological reassurance of low-GG disease without further morbidity. Hypothetical results for PRONTO-e are similar (Figure 9). Over time, use of such a test could ease surveillance for the majority of patients identified as low-risk, potentially reducing the number of biopsies performed on a population basis.

[0147] This study established PRONTO-e and PRONTO-m as molecular biomarkers of GG that are robust to sampling error and therefore likely perform well in diagnostic biopsies. Further studies are needed and ongoing to fully validate their clinical performance. Multifocal CaP is a potential pitfall of any biopsy test in that biopsy may sample less significant low-grade lesions while failing to sample high-grade "dominant" or "index" lesions. This phenomenon is estimated to explain 20–30% of cases that are upgraded between biopsy and prostatectomy [15,50]. Classifier performance on biopsy tissue may also be compromised by limited nucleic acid yield from small biopsy tissue samples. This limitation should be balanced by factors expected to improve classifier performance in biopsies compared to surgical specimens, including the higher quality of nucleic acids observed in biopsy tissue

[51] and the opportunity to employ more sensitive and accurate massively parallel sequencing technologies

[52] in clinical assays.

[0148] Although several studies have correlated biopsy classification with postoperative outcomes, little information is available linking test results to outcomes in men undergoing AS. Further validation of PRONTO-e and PRONTO-m in biopsies from AS patients is warranted. Overall, these results indicate that combining transcriptomic, epigenomic, and genomic features can improve the performance of clinically relevant biomarkers for CaP tissue. The results suggest that other biological sample types (e.g., blood or urine) and tumor sites may be advantageous.

[0149] While preferred embodiments of the present invention have been described herein, it will be understood by those skilled in the art that variations can be made thereto without departing from the spirit of the invention or the scope of the appended claims. All documents disclosed herein, including those in the reference list below, are incorporated by reference.

[0150] Table 1

[0151] Table 2

[0152] Table 3

[0153] Table 4

[0154] Table 5

[0155] Table 6(1)

[0156]

Table 6(2)

[0157] Table 6(3)

[0158] Table 6(4)

[0159] Table 6(5)

[0160] Table 6(6)

[0161] Table 6(7)

[0162] Table 6(8)

[0163]

Table 6(9)

[0164] Table 6(10)

[0165] Table 6(11)

[0166]

Table 6(12)

[0167] Table 6(13)

[0168] Table 6(14)

[0169] Table 7(1)

[0170] Table 7(2)

[0171] Table 7(3)

[0172] Table 7(4)

[0173] Table 7(5)

[0174] Table 7(6)

[0175] Table 7(7)

[0176] Table 7(8)

[0177] Table 7(9)

[0178] Table 7(10)

[0179] Table 7(11)

Claims

1. 1. A method for predicting the risk of disease progression in a subject with prostate cancer, said method comprising: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring the values ​​of some or all of the 353 patient features, including mRNA and copy number aberration (CNA) features listed for PRONTO-e in Table 6, as well as the reference or control features listed in Table 6; c) comparing said patient characteristics with reference or control characteristics; d) calculating a prediction score using a classifier that takes patient feature values ​​as input data, said classifier having been pre-trained on training samples from a population of early stage prostate cancer patients, and correlating said patient features with known Gleason Grade Group (GG) classes of said training samples using a mean area under the curve (AUC); e) classifying the patient's tumor into a pathological Gleason Grade Group (GG) class.

2. The patient's tumor, [Equation 1] Alternatively, the method of claim 1, wherein a score of <0.5 is classified into the pathological GG1 class.

3. 3. The method of claim 2, further comprising recommending active surveillance for the patient if the patient is classified in the pathological GG1 class. [Request Item 4] [Number 2] 3. The method of claim 2, further comprising recommending surgery, endocrine therapy, chemotherapy, radiation therapy, hormone therapy, gene therapy, heat therapy, or ultrasound therapy to the patient.

5. 1. A method for predicting the risk of disease progression in a subject with prostate cancer, said method comprising: a) providing a sample containing RNA and DNA material from tumor cells; b) determining or measuring some or all of the 94 patient features, including mRNA, CNA, methylation and clinical features listed for PRONTO-m in Table 6, as well as the reference or control features listed in Table 6; c) comparing said patient characteristics with reference or control characteristics; d) calculating a prediction score using a classifier that takes the values ​​of the patient features as input data, the classifier having been pre-trained on training samples from a population of early stage prostate cancer patients, and correlating the patient features with the known Gleason Grade Group (GG) classes of the training samples using a mean area under the curve (AUC); e) classifying the patient's tumor into a pathological Gleason Grade Group (GG) class.

6. The patient's tumor, [Equation 3] The method of claim 5, wherein a score of <0.5 is classified into pathological GG1 class.

7. 7. The method of claim 6, further comprising recommending active surveillance for the patient if the patient is classified into the pathological GG1 class. [Request Item 8] [Number 4] 10. The method of claim 6, further comprising recommending surgery, endocrine therapy, chemotherapy, radiation therapy, hormone therapy, gene therapy, heat therapy, or ultrasound therapy to the patient.

9. 1. A computer-implemented method for predicting risk of disease progression in a prostate cancer patient, said method comprising: a) receiving, in at least one processor, data reflecting patient characteristics as defined in claims 1 or 5 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors, and some or all of the reference or control characteristics listed in Table 6; b) constructing, in at least one processor, a patient profile based on the patient characteristics; c) comparing said patient profile with a reference or control in at least one processor; d) calculating, in at least one processor, a prediction score using a classifier that takes the patient profile as input data, the classifier having been pre-trained on training samples from a population of early stage prostate cancer patients, and correlating the patient characteristics with known Gleason Grade Group (GG) classes of the training samples using a mean area under the curve (AUC); e) classifying the patient's tumor into a pathological Gleason Grade Group (GG) class.

10. 10. A computer program product for use with a general purpose computer having a processor and a memory connected to said processor, said computer program product comprising a computer readable storage medium having encoded thereon a computer program mechanism that may be loaded into the memory of the computer and that can cause said computer to perform the method of claim 9.

11. 11. A computer readable medium having stored thereon a data structure for storing the computer program product of claim 10.

12. 1. A device for predicting risk of disease progression in a prostate cancer patient, the device comprising: at least one processor; and an electronic memory in communication with at least one processor, the electronic memory, when executing the at least one processor, transmitting to the at least one processor: a) receiving data reflecting some or all of the patient characteristics defined in claim 1 or 5 and the reference or control characteristics set forth in Table 6 corresponding to a PRONTO-e or PRONTO-m classifier for prostate cancer tumors; b) comparing said patient characteristics to reference or control characteristics; c) calculating, with at least one processor, a prediction score using a classifier that takes the patient profile as input data, said classifier having been pre-trained on training samples from a population of early stage prostate cancer patients, and correlating said patient characteristics with known Gleason Grade Group (GG) classes of said training samples using a mean area under the curve (AUC); d) classifying the patient's tumor into a pathological Gleason Grade Group (GG) class; A device that stores processor-executable code.

Images