Development and Clinical Validation of an In Situ Biopsy-Based Multimarker Assay for Risk Stratification in Prostate Cancer

In 2014, there will be an estimated 233,000 new diagnoses of prostate cancer in the United States (1). The majority of these patients will have early-stage, clinically localized disease (1–3). Given the marked genomic and pathologic heterogeneity of prostate cancer and concerns regarding its overtreatment (4, 5), significant effort has been devoted to the discovery and validation of biomarkers capable of distinguishing indolent cases with good prognosis from more aggressive cases with poor survival (6). Pathologic evaluation of needle biopsy tissue establishes a prostate cancer diagnosis and provides data to define a patient's risk category that will guide treatment decisions (6). Although a number of useful classification systems have been developed that combine available clinical and pathologic parameters (6, 7), it is widely recognized that improved classification methods are needed to ascribe an individualized risk score and help clinical management (8–10).

A significant proportion of patients undergoing biopsy ultimately will have their tumor pathology (as defined by Gleason score) upgraded or downgraded following assignment of a more accurate “surgical” or pathologic Gleason score after prostatectomy (11). For example, 25% to 30% of patients with low-risk disease at biopsy are subsequently upgraded upon pathologic review of their radical prostatectomy (10–12). These revisions may reflect initial biopsy sampling error (13), or pathologist discordance in tumor grading (14), both of which can contribute either to overtreatment or delayed treatment of disease (4). In particular, there are concerns around overcalling or undercalling Gleason pattern 4 in needle biopsy samples, because these cases usually fall in a gray zone between clear active surveillance and surgical intervention (11, 15–17). In addition, there is an ongoing need to determine whether a cancer is organ confined, or non-organ confined with ultimate metastatic potential in patients with low- to intermediate-grade disease.

Advances have been made in identifying genetic markers that may inform clinical risk prognostication in prostate cancer (18, 19). There has been focus on identifying in situ protein biomarkers on the cellular level that, under circumstances of intratumor cellular heterogeneity, would enable more accurate classification of the most aggressive tumor areas that may consist of only few cancer cells (20, 21). We recently reported the development and use of a robust quantitative multiplex proteomics imaging (QMPI) approach to identify and measure protein biomarker candidates resistant to such sampling error (see Supplementary Appendix; ref. 22). Starting with a large pool of potential candidates, this biopsy simulation study identified 12 biomarker candidates that predict both prostate pathology aggressiveness and lethal outcome despite sampling error (22) based on prostatectomy tissue samples (Supplementary Appendix). The 12 biomarkers are ACTN1, CUL2, DERL1, FUS, HSPA9, PDSS2, PLAG1, pS6, SMAD2, SMAD4, VDAC1, and YBX1.

In this study, we report further development and subsequent blinded validation of an 8-biomarker assay derived from 11 of these markers in two independent clinical biopsy studies (Figure 1). Specifically, we first designed a biomarker development study to define an optimal biomarker subset and develop a predictive mathematical model. This effort resulted in an 8-biomarker prognostic model that provided “risk scores” predictive of final disease pathology. Next, benchmarked against Gleason scores and TNM staging data of matched post-surgical prostatectomy tissue, the performance of this 8-biomarker biopsy assay to predict the clinically relevant endpoint of favorable versus nonfavorable pathology was evaluated in an independent clinical validation study. The risk score generated by the 8-biomarker assay for each patient was compared with two current risk stratification systems, the National Comprehensive Cancer Network (NCCN) guideline categories and D'Amico system (6, 7), and its potential to provide Supplementary Data of high-predictive value that could aid in treatment decision for individual patients was considered. The intended use of this biomarker assay is to supplement current biopsy-based prostate cancer risk assessment methods in cases where a clinical decision regarding active surveillance versus active treatment is not straightforward.

To develop a robust assay, multiple institutions were recruited representing typical North American patient cohorts (Table 1): Urology Austin, Chesapeake Urology Associates, Cleveland Clinic, Michigan Urology, and Folio Biosciences. Biopsy sample inclusion/exclusion criteria matched those that would be in place during routine clinical use of the assay. Because they are usually not candidates for active surveillance, patients with biopsy Gleason ≥4+3 were excluded, except for a limited number of biopsies that had been discordantly graded as both 3+4 and 4+3 by two expert pathologists. Annotations, including information on matched biopsy and prostatectomy pathology reports, were obtained. Researchers were blinded to sample identity during laboratory processing. The final number of cases included in the biomarker assay development cohort was 381.

The validation study cohort (N = 276) was separate and independent from the assay development cohort, and comprised biopsy samples with matched prostatectomy annotation from patients managed at the University of Montreal Hospital Center, Canada (Table 1). Consent criteria and Institutional Review Board (IRB) approval steps were as for the assay development phase of the study, above. Inclusion criteria were biopsies with a centralized Gleason score 3+3 or 3+4 (biopsies with discordant grading by two expert pathologists of 3+4 and 4+3 were included as well), and matched prostatectomy with pathologic TNM staging, PSA level, and resulting surgical Gleason score. Central review was performed on biopsies, whereas for prostatectomies we had to rely on the original pathology annotation, due to practical challenges in accessing all the different locations where these had been conducted; however, all prostatectomy pathology annotation was done according to new ISUP classification. Sample sizes were statistically designed, with power greater than 95% for both cohorts.

The QMPI approach for protein in situ measurements was as described (Supplementary Appendix; ref. 23). Briefly, formalin-fixed, paraffin-embedded prostate cancer biopsy tissue slides were analyzed using a quantitative multiplex proteomics imaging (QMPI) platform for intact tissue that integrates morphologic object recognition and molecular biomarker measurements from tumor epithelium at the individual slide level. Four combinations of three (triplex) biomarkers each were used: (i) PLAG1, SMAD2, ACTN1; (ii) VDAC1, FUS, SMAD4; (iii) pS6, YBX1, DERL1; (iv) PDSS2, CUL2, HSPA9. When each of the primary antibodies used was validated for specificity, it was found that PLAG1 was insufficiently specific; it was, thus, excluded from the potential signature. Prostate tissue slides that passed initial quality control were subjected to the multiplex staining procedure with subsequent image acquisition, Definiens analysis, and bioinformatics analysis. (See Supplementary Appendix for further details of antibody validation, staining protocols, image acquisition, image analysis, and interexperimental controls.)

A noninterventional, retrospective clinical assay development study using biopsy case tissue samples was devised to define the best marker subset from those candidates previously shown to correlate with both prostate pathology aggressiveness and lethal outcome (22). The study goal was to define a model able to distinguish between prostate pathology usually recommended for active surveillance (“GS 6”): surgical Gleason 3+3 and ≤T3a versus those more likely to require prostatectomy (“non–GS 6”): Surgical Gleason ≥3+4 or nonlocalized >T3a or N or M. This GS 6 versus non–GS 6 definition was based on studies showing that tumors with surgical Gleason 3+3 at prostatectomy do not metastasize (15, 16). IRBs approved the biomarker development protocol, and patient consent was obtained or waived accordingly.

To rank our previously identified biomarkers in terms of their predictive importance, we evaluated each potential combination of biomarkers using 500 rounds of computation. In each round, a training set (65% of cases) was obtained by bootstrap and remaining samples (35% of cases) were used as a testing set. A logistic regression model was trained on the training set, and risk scores were obtained for samples in both the training and testing sets. The final marker coefficients were used in a logistic regression model for calculation of the risk score, provided as a continuous scale from 0 to 1. After computation of these statistics, we sorted the biomarker sets in three different ways. Biomarker sets were characterized by the area under the receiver operating characteristic (ROC) curve (AUC), and sorted by increasing value of Akaike information criterion, decreasing value of the AUC on the training set, and decreasing value of the AUC on the testing set. To finalize the set of biomarkers for our assay, we determined the frequency of usage of each biomarker in the 10% of most highly ranked sets. Optimal biomarker set size was likewise determined through examination of test-set AUC performance during the exhaustive model search. We defined a resulting risk score, a continuous number between 0 and 1, which estimated the likelihood of “non–GS 6” pathology. Sensitivity analyses confirm the defined, locked-down biomarker assay set. We also explored a number of sensitivity analyses to confirm performance robustness. These analyses identified an optimal set of 8 biomarkers with the best performance in distinguishing GS 6 from non–GS 6 cases (Fig. 2).

A noninterventional, blinded, prospectively designed, retrospectively collected clinical study was conducted to validate the performance of the 8-biomarker assay to predict prostate pathology on its own and relative to current systems for patient risk categorization. This study used biopsy samples from a patient cohort independent of the one used for assay development (Table 1), and generated a “risk score” for each sample. Matched prostatectomy samples with annotated surgical Gleason scores ultimately classified the tumor as “favorable” or “nonfavorable” for the purposes of evaluating assay results. ROCs and corresponding AUCs determined for the 8-biomarker risk score quantitatively evaluated performance of the assay.

Favorable versus nonfavorable pathology was chosen for final patient categorization throughout the validation study. It reflects the increasing awareness that organ-confined disease with minimal Gleason 4 pattern is likely to remain indolent with a significantly better long-term prognosis than higher-grade (dominant Gleason 4 pattern) or non–organ-confined disease (16, 24, 25). Thus, one of our two coprimary endpoints was defined as “favorable” pathology: Surgical Gleason ≤3+4 and organ-confined disease (≤T2) versus “nonfavorable” pathology: Surgical Gleason ≥4+3 or non–organ-confined disease (T3a, T3b, N, or M). In addition, in recognition of clinical variability, a second coprimary endpoint was defined as “GS 6” pathology: Surgical Gleason of 3+3 and localized disease (≤T3a) versus “non–GS 6” pathology: Surgical Gleason ≥3+4 or nonlocalized disease (T3b, N, or M).

Using biopsy samples, the 8-biomarker assay–derived risk score validated these two coprimary endpoints for prostate pathology, as assessed by AUC. Secondary analyses included ORs for the highest quartile versus lowest quartile of risk score, and OR (point estimate) for the continuous scale. We compared the risk outcomes from our diagnostic test with current risk classification categories as defined by the NCCN and D'Amico (6, 7), using positive predictive values (PPV). Definition and statistical analysis of the Net Reclassification Improvement (NRI) was done as described by Pencina and colleagues (26).

See the Supplementary Appendix for the statistical plan for both clinical studies.

We first sought to identify a protein biomarker set and predictive model capable of correlating biopsy-based prognosis with final post-prostatectomy pathology. Biomarker assay development used prostate biopsies and post-prostatectomy annotation from 381 patients recruited from several U.S. institutions (Table 1). Eleven previously identified biomarker candidates were evaluated using a 5-step optimization process to determine the top-performing biomarker set from among all possible combinations (Supplementary Appendix, Supplementary Fig. S1). Individual quantitative biomarker measurements were correlated with prostate cancer pathology as an endpoint, and univariate ORs were calculated (Fig. 2A). The performance of all possible biomarker combinations were assessed and several high-performing sets were identified; the biomarker set with the highest performance had a test AUC of 0.79 [95% confidence interval (CI), 0.72–0.84]. Given that many biomarker sets had similar performance in the bootstrapped test AUC, both univariate performance and frequency of occurrence in top-performing models (Fig. 2B) were used to choose the final biomarker set for validation. Model coefficients for each marker were derived from logistic regression on the full dataset. The resulting locked-down biomarker assay set consisted of eight biomarkers (CUL2, DERL1, FUS, HSPA9, PDSS2, pS6, SMAD4, and YBX1) as shown in Fig. 2.

Clinical validation used biopsy specimens, and post-prostatectomy annotation from a second independent cohort of 276 patients. As shown in Table 2, the study met its two coprimary endpoints and validated the 8-biomarker assay for both endpoints: “favorable” versus “nonfavorable” and “GS 6” versus “non–GS 6” (see Materials and Methods for endpoint details). The analysis for “favorable” pathology yielded an AUC of 0.68 with 95% CI, 0.61–0.74. The associated P value was <0.0001, with an OR for risk score of 20.9 per unit change. “GS 6” pathology yielded an AUC of 0.65 with 95% CI, 0.58–0.72. The associated P value was <0.0001, with an OR for risk score of 12.6 per unit change. Further details are shown in Supplementary Figs. S2 and S3 in the Supplementary Appendix.

We had sufficient prostatectomy annotation to classify 256 cases within our validation cohort according to NCCN and D'Amico criteria, two current systems for risk classification. The performance for “favorable” pathology of the 8-biomarker assay on associated biopsy samples from this cohort is shown in Supplementary Fig. S4 and was similar to the full cohort of 276 cases (AUC, 0.69; 95% CI, 0.63–0.76; P < 0.0001; OR for risk score, 26.2/unit change).

The AUC for identification of the “favorable” group based on NCCN risk stratification alone was 0.69 (95% CI, 0.62–0.75) and 0.65 (95% CI, 0.59–0.71) based on D'Amico alone. The AUCs for the combined models of the 8-biomarker assay with NCCN and D'Amico were 0.75 (95% CI, 0.69–0.81) and 0.75 (95% CI, 0.69–0.81), respectively.

Multivariate analysis was also performed during development of the combined models. The OR for aggressive disease per 0.25 change in the 8-biomarker assay risk score, with NCCN held as a constant, was 3.5 (pv < 10⁻³; 95% CI, 2.6–4.7). The OR for aggressive disease per 0.25 change in 8-biomarker assay risk score, when the D'Amico status was held constant, was also 3.5 (pv < 10⁻³, 95% CI, 2.6–4.7). These results are consistent with the notion that the 8-biomarker assay, the results of which are entirely based on integrated molecular and morphologic information, is capturing information that is partly independent of the prognostic features represented in the NCCN and D'Amico models, both of which are entirely based on clinical and pathologic parameters.

The 8-marker proteomic assay was also considered in the context of standard clinical parameters, including PSA, the percentage of cores positive as a surrogate for tumor volume, and biopsy Gleason pattern. These parameters were available to the investigators for the clinical development (‘train-test’) cohort and held as blinded data by the independent validating statistician for the validation cohort. A logistic regression model was built for the “favorable” endpoint on the train–test cohort using the clinical nomogram parameters alone as well as for the integrated model combining the clinical and 8-marker assays. The clinical parameters alone resulted in an AUC of 0.67 (95% CI, 0.61–0.74), and the combined model resulted in an AUC of 0.71 (95% CI, 0.64–0.77). Multivariate analysis of the combined model showed that the OR for aggressive disease per 0.25 change in 8-biomarker assay risk score, when clinical parameters were held constant, was 3.1 (P < 10⁻³; 95% CI, 2.2–4.4). Again, these results are in accordance with the molecular information captured by the 8-biomarker assay is partly independent of the prognostic features based on the clinical and pathologic parameters in the clinical nomogram.

Next, to define the risk score ranges (cutoff points) for favorable versus nonfavorable pathology, we generated sensitivity and specificity curves for the 8-biomarker assay (Fig. 3A and B, respectively). Figure 3A shows an example of a favorable risk score category identified in this study population on the basis of the 8-biomarker assay. Here, a sensitivity of 90% (95% CI, 82%–94%) corresponds to a biomarker assay risk score of 0.33 [sensitivity: P(risk score > 0.33 | nonfavorable pathology)]. Below this threshold of risk score, the false-negative rate among patients with nonfavorable pathology is limited to 10% (95% CI, 6%–18%). Stated differently, among patients with risk score ≤0.33, the chance of having nonfavorable pathology is less than or equal to 1 in 10. Conversely, in Fig. 3B, a nonfavorable category was defined by a biomarker assay risk score of 0.80, which in this study population corresponded to a Specificity of 95% (95% CI, 90%–98%). Above this threshold of risk score, the false-positive rate among patients with favorable pathology is limited to 5% (95% CI, 2%–10%), where Specificity = P(risk score ≤ 0.80| favorable pathology). Stated differently, among patients with risk score greater than 0.80, the likelihood of having a favorable pathology is less than 5 in 100.

Similar to prognostic assay such as OncotypeDx, which stratifies an individual patient to a particular risk level through cutoff points that define risk categories (27), the intended use of this 8-biomarker assay is to categorize an individual patient to favorable versus nonfavorable disease pathology based on where his risk score is on the specificity curves of Fig. 3. To this end, we first calculated the predictive value of the risk scores. The PPV for identifying favorable disease with a risk score of ≤0.33 was 83.6% (specificity, 90%; Table 3). Conversely, at a risk score of >0.80, 76.9% had nonfavorable disease (i.e., only 23.1% of patients in the high-risk score category had favorable disease). This translates to 39% of patients in this study population having risk scores ≤0.33 or >0.8, of which 81% were correctly identified by our 8-biomarker assay (Table 3). Of patients with intermediate risk scores (0.33 < risk score ≤ 0.8), 58.3% had favorable disease.

Next, we compared the PPV of this 8-biomarker assay with those of the NCCN and D'Amico risk categories. Here, within each NCCN category, we analyzed the distribution of patients categorized as having favorable disease by our biomarker risk score. The PPV for favorable pathology, at a risk score of ≤0.33, was 75% for patients in the NCCN intermediate-risk category; 81.5% for NCCN low-risk patients; and 95% for NCCN very low-risk, with an overall PPV for favorable pathology of 85% (Table 4, in green, and Fig. 3C). This contrasts favorably with the PPVs obtained for the NCCN risk categories alone, which were 40.9% for intermediate-risk, 63.8% for low-risk, and 80.3% for very low-risk (Table 4, blue). In other words, when combined with the current prognostic standard, this 8-biomarker assay can increase the confidence that a patient categorized to the NCCN very low-risk group indeed has a favorable pathology based on his biopsy from 80.3% to 95%. Conversely, the predictive value for nonfavorable pathology was 76.9% at a risk score >0.8, but reached 100% at a risk score cutoff point of >0.9. Correspondingly, a higher biomarker risk score correlated with decreased frequency of favorable cases within each NCCN category (Fig. 3D). Taken together, we conclude that the risk score generated by our 8-biomarker assay offered additional prognostic value for individual patients relative to NCCN risk categories alone. Similar results were obtained when comparing with the D'Amico categories (Supplementary Fig. S5).

To further confirm the benefit of having the 8-biomarker risk score in the context of current risk stratification systems, we performed an NRI analysis for the favorable and nonfavorable categories determined by our 8-biomarker assay, relative to NCCN and D'Amico categories (Supplementary Fig. S6). Using the underlying data shown in Tables 3 and 4, we found an NRI for NCCN of 0.34 (P < 0.00001; 95% CI, 0.20–0.48) and for D'Amico of 0.24 (P = 0.0001; 95% CI, 0.12–0.35). This provides strong evidence that this 8-biomarker assay provides additional discriminatory capability to that offered by current risk stratification systems alone.

We also performed decision curve analysis, as this provides another method for characterizing performance of different risk systems and risk cutoff points (28). Supplementary Fig. S7 provides an example of such an analysis and describes the net benefit for a number of treatment regimes based on the 8-marker model, NCCN, and the combined model NCCN and 8-marker model. Using the illustrated cutoff points, we see that the combined model provides benefit through the span of risk thresholds. The range of incremental net benefit improves from small (0.005 at risk threshold of 20%) to significant (0.06 at a risk threshold of 50%) across the scale of risk thresholds. This further demonstrates the added prognostic information achieved by combining the 8-marker model with a standard-of-care risk stratification system.

Recent studies indicate that long-term survival for patients with organ-confined Gleason 3+4 disease is significantly better than for patients with non–organ-confined disease or for tumor with dominant Gleason pattern 4 or higher (17, 24, 25), and that deferred therapy for the former group does not significantly change long-term outcome (29–31). Most risk stratification systems do not discriminate between Gleason 7 biopsies, so typically patients considered candidates for active surveillance belong to “very low-risk” or “low-risk” groups, which only contain biopsy Gleason score ≤6 (2, 6). However, around 25% of Gleason grade 3+4 biopsies are “downgraded” and a similar percentage of Gleason grade 3+3 biopsies are “upgraded” when comparing with the surgical Gleason scores, primarily owing to biopsy sampling error and pathologist discordance (11, 15). Therefore, there is a clear unmet clinical need for a molecular evidence-based prognostic assay that can complement the current pathology-based stratification system(s) to more accurately predict disease aggressiveness of patients with Gleason grade 3+3 and 3+4 biopsies (16).

To address this unmet need, we have embarked on an assay development effort, building on prior work that had led to the identification of 11 biomarker candidates shown to be predictive of both prostate cancer pathology and lethal outcome despite tissue sampling error (22). Of importance is that, we have used a multiplex proteomics imaging platform (23) for quantitative measurements of eight independent biomarkers in intact tissue to generate the risk score (Supplementary Appendix). This in situ approach has the advantage compared with other tests using homogenized tissue (19, 32), as it requires relatively few cancer cells (which is common of the biopsy specimens) and is not sensitive to variations in the ratios of benign tissue relative to tumor tissue.

Here, we report the results demonstrating that we have successfully developed an 8-biomarker biopsy-based prognostic assay that was validated in an independent blinded study and shown to complement current risk stratification systems. Specifically, our study shows that, at a risk score of ≤0.33, our biomarker assay can identify patients with favorable pathology in the very low-risk and low-risk NCCN and low-risk D'Amico groups with PPVs of 95%, 81.5%, and 87.2%, respectively, that represents a significant improvement over NCCN or D'Amico stratification system alone. We suggest that information provided by the biomarker assay in addition to existing pathologic parameters can help both the physician and patient to consider active surveillance option with added confidence.

Furthermore, our 8-biomarker assay is also able to identify with high-confidence patients with nonfavorable pathology who are arguably unsuitable for active surveillance. Although a patient with a risk score >0.8 has a PPV of 76.9% for nonfavorable pathology, the PPV reaches 100% if the risk score is >0.9 across all risk groups for both risk stratification systems tested. This shows how this biomarker assay can be used in combination with current stratification system(s) to personalize the risk level for each individual patient, enabling rational and evidence-based decision by both physician and patient. Note, that the settings used for clinical decision-making thresholds are a matter of risk tolerance in the patient and physician. Because our assay provides continuous specificity and sensitivity curves, and individualized risk for each patient, other thresholds that reflect different individual risk tolerance could be chosen. For example, if 80% specificity is sufficient for a particular patient, then a risk score of 0.6 or higher would favor treatment over active surveillance.

Because of the heterogeneous and multifocal nature of prostate cancer, prostate biopsies frequently contain only lower-grade morphologic features with higher-grade molecular features that are not reflected in morphology. Our 8 biomarkers were specifically selected and evaluated to equally well-predict pathology outcome regardless whether they were measured in low-grade or high-grade regions from the same patient (22). Accordingly, by measuring proteins only from intact tumor regions, our test can accurately and sensitively assess such high-grade molecular features in situ, even in tissue samples with variable amounts of tumor versus benign components. This is a significant advantage over previously reported gene expression-based technologies that rely on tissue homogenization (18, 32–34), and where molecular (genetic) information from higher-grade regions is variably diluted by the values of the individual markers in intermixed benign tissue regions that inadvertently are part of the tissue homogenate. Genomic Health's “Oncotype DX prostate cancer test” is a 17-gene PCR-based assay, similar to the biopsy-based test described in this study, that provides genomic prostate score (GPS) for predicting cancer aggressiveness in prostate cancer patients based on the same endpoints we have developed (32, 33). Although a stand-alone AUC for the “Oncotype Dx” test has not been reported, the combined AUC for GPS and CAPRA is reported to be 0.67, whereas the AUC for CAPRA alone was reported as 0.63 (for favorable pathology). Obviously, a direct comparison between these studies is difficult, as a head-to-head study would be required to fully establish relative performance. Myriad's “Prolaris” test is another gene expression–based assay that measures activity of a panel of cell-cycle progression (CCP) genes and provides risk of prostate cancer–specific progression and mortality (18, 34). Although this test was originally developed on prostatectomy samples, it was assessed on needle biopsy material for its ability to predict lethal outcome from prostate cancer. Unfortunately, in this study (19) patients with biopsy Gleason scores of 4+3, 8, 9, and 10 were included and made up approximately 26% of the study population. Arguably, these patients are in need of aggressive therapy, and therefore not included in our intended use, clinical need population. The “Prolaris” CCP score showed an HR of 1.65 in multivariate analysis, and although the majority of the performance of the test was derived from the patients with advanced disease, no specific measure of the performance on patients with biopsy scores of 3+3 or 3+4 was reported. Finally, the “Prolaris” test provides a prostate cancer risk score for patients by combining the CCP score with standard clinicopathologic parameters, it is not a stand-alone test (18, 32–34).

In this study, Gleason scores and TNM staging of matched post-prostatectomy annotations were used as benchmark to evaluate performance of the 8-biomarker assay, instead of metastatic disease or death as endpoints. This is due to relatively low grade and early stage of the target population of our test, rendering it necessary to have very long follow-up of very large and conservatively managed cohorts to be sufficiently powered for such study. Indeed, only few biopsy studies have had metastatic disease or lethal outcome as endpoints. As expected, to compensate for the power, these studies have included a significant proportion of high-risk cases (biopsies with Gleason score ≥4+3) in which time to death is much shorter; however, in such cases, the clinical utility of a complementary molecular test is arguably more limited (19).

Finally, our methodology was based on a binary endpoint derived from prostate pathology that was shown in a previous clinical study to correlate with lethal outcome (22). The binary endpoint was chosen to match clinical practice, in which patients are considered for active surveillance or aggressive treatment. For future investigations, assays could be developed on the basis of either an ordered categorical (e.g., low-, medium-, high-risk prostate pathology), or a continuous endpoint by combining inherently categorical prostate pathology assessment with other clinical variables. Although development of such models is straightforward, they would inherently be based on subjective weighting and combination of parameters that would require a new assessment of correlation with long-term outcome before general acceptance.

In conclusion, this study supports further evaluation of this biopsy-based prognostic biomarker assay for personalized prognostication of prostate cancer and its impact on therapeutic choice. The ability to provide differential information for the individual patient relative to current risk stratification systems, in which prognostic values are more limited, makes it a potentially useful addition in practice to improve accuracy of clinical decision.

P. Blume-Jensen, L. Coupal, E.A. Klein, and F. Saad are consultants/advisory board members for Metamark Genetics, Inc. D. Berman reports receiving a commercial research grant from Myriad Genetics and is a consultant/advisory board member for Metamark Genetics, Inc. D.L. Rimm and P. Kantoff have ownership interest (including patents) and are consultants/advisory board members for Metamark Genetics, Inc. No potential conflicts of interest were disclosed by the other authors.

Conception and design: P. Blume-Jensen, D.M. Berman, D.L. Rimm, T.P. Nifong, A. Hurley, E.A. Klein, J.I. Epstein, P. Kantoff, F. Saad

Development of methodology: P. Blume-Jensen, M. Shipitsin, T.P. Nifong, C. Small, S. Choudhury, A. Hurley, A. Kaprelyants, F. Saad

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): P. Blume-Jensen, M. Shipitsin, M. Putzi, T.P. Nifong, S. Choudhury, T. Capela, L. Coupal, C. Ernst, A. Kaprelyants, H. Chang, M. Loda, E.A. Klein, C. Magi-Galluzzi, M. Latour, F. Saad

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): P. Blume-Jensen, D.M. Berman, T.P. Nifong, C. Small, L. Coupal, E. Giladi, J. Dunyak, M. Loda, M. Latour, J.I. Epstein, F. Saad

Writing, review, and/or revision of the manuscript: P. Blume-Jensen, D.M. Berman, D.L. Rimm, M. Shipitsin, M. Putzi, T.P. Nifong, S. Choudhury, L. Coupal, A. Hurley, J. Dunyak, E.A. Klein, C. Magi-Galluzzi, M. Latour, J.I. Epstein, P. Kantoff, F. Saad

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): P. Blume-Jensen, C. Small, A. Hurley, A. Kaprelyants, H. Chang, J. Nardone, F. Saad

Study supervision: P. Blume-Jensen, D.M. Berman, M. Shipitsin, A. Hurley, F. Saad

The authors thank Drs. Giovanni Parmigiani, Lynda Chin, Raju Kucherlapati, and Greg Critchfield and collaborators at Metamark Genetics, Inc. for insightful comments and suggestions. The authors thank Rowena Hughes and Winnie McFadzean from Oxford PharmaGenesis Ltd., and Denise Spring for editorial support in collating comments from authors and finalization of the article for submission.

This project was funded by Metamark Genetics, Inc.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.