A Genomic Classifier Improves Prediction of Metastatic Disease Within 5 Years After Surgery in Node-negative High-risk Prostate Cancer Patients Managed by Radical Prostatectomy Without Adjuvant Therapy

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.

Adverse pathology on a radical prostatectomy (RP) specimen, as defined by high Gleason grade, extraprostatic extension (EPE), seminal vesicle invasion, or lymph node positivity, is known to increase the risk of tumor recurrence. However, not all patients with one or more of these features are destined to develop life-threatening disease, as demonstrated by a recent analysis of almost 24 000 men that showed approximately 70% with one or more of these features had not died of prostate cancer (PCa) 15 yr after surgery [1] . Although some of these men received adjuvant therapy soon after surgery and others were treated at the time of progression, the low rate of metastatic disease in this and other large series with long-term follow-up suggests that many men with adverse pathology are cured by surgery alone[2] and [3]. Currently available tools such as nomograms based on clinical and pathologic factors have imperfect ability to identify men with metastatic progression who may benefit from adjuvant therapy, closer monitoring for disease recurrence with early initiation of salvage therapy, or participation in clinical trials [4] . The addition of biomarkers to clinical risk factors may help identify high-risk men who are at risk for metastatic disease within 5 yr following RP [5] . Decipher, a genomic classifier that uses a whole-transcriptome microarray assay from formalin-fixed paraffin embedded (FFPE) PCa specimens, has been developed and validated as a specific predictor of metastases and PCa-specific mortality following RP in several cohorts of men with adverse pathologic features[6], [7], [8], and [9]. In this study, we aimed (1) to compare the performance of Decipher with standard risk-stratification tools (CAPRA-S and the Stephenson nomogram), (2) to assess performance and accuracy of adding Decipher to these tools for predicting metastatic disease within 5 yr after surgery in men with adverse pathologic features treated with RP, and (3) to compare Decipher with a panel of other reported gene biomarkers in predicting this end point.

The Cleveland Clinic institutional review board reviewed and approved the research protocol under which this validation study was conducted. The study met the PRoBE [10] and REMARK [11] criteria for prospective blinded evaluation and analysis of prognostic biomarkers.

Among the 169 patients, 15 experienced rapid metastasis (RM; metastasis within 5 yr of surgery) and 34 had late metastasis (LM; metastasis >5 yr after surgery). All clinical and pathologic factors at the time of treatment were not significantly different between the groups ( Table 2 ). Metastasis occurred at a median of 2.3 yr (interquartile range [IQR]: 1.7–3.3) in those with RM compared with 9.3 yr (IQR: 8.3–11.6) in the LM group. More patients in the LM group (28 of 34, 82.4%) received secondary therapy at the time of biochemical recurrence and prior to metastasis than those with RM (4 of 15, 26.6%) (p < 0.0019). Median PCa-specific mortality (PCSM) occurred at 7.3 yr in the RM group; LM patients had a PCSM rate of 25% by 15 yr after RP ( Fig. 1 ). The key difference between these groups was time to development of initial metastases, as no significant difference in PCSM rate was observed between the two groups after the development of metastasis (p = 0.87) (Supplementary Fig. 2).

Fig. 1 Cumulative incidence of prostate cancer specific mortality by (A) rapid and (B) late metastatic status of patients after radical prostatectomy. PCSM = prostate cancer–specific mortality; RP = radical prostatectomy.

The median Decipher score for all patients was 0.35 (range: 0.03–0.91; IQR: 0.22–0.59), with mean scores of 0.58 (IQR: 0.44–0.91), 0.54 (IQR: 0.33–0.87), and 0.33 (IQR: 0.18–0.41) for RM, LM, and control (nonmetastatic) patients, respectively. Decipher score was moderately correlated with pathologic Gleason score (Spearman ρ = 0.47). Decipher score was also moderately correlated with CAPRA-S score (which includes PSA and pathologic stage in addition to Gleason score; Spearman ρ = 0.35) but further stratified patients within each CAPRA-S risk category ( Fig. 2 ). Among the 47 patients with multiple adverse pathology features that had the highest CAPRA-S scores (corresponding to >50% probability of recurrence by this model), 15 (32%) were classified as Decipher low risk (according to previously reported cut points [7] ), and only one of these patients developed RM. Conversely, for the 23 patients with Gleason score 6 disease (19 with positive margins and 4 with EPE), all but 2 (who had borderline scores) were classified in the Decipher low-risk group, and none of these patients developed metastasis. Finally, in the subset of patients with metastasis, median Decipher score was 0.31 (IQR: 0.26–0.39; not different from controls) for those patients with primary Gleason grade 3 and 0.64 (IQR: 0.54–0.91) for those with primary Gleason pattern ≥4.

Fig. 2 Scatter plot comparing Decipher with (A) pathologic Gleason score and (B) CAPRA-S score. Decipher stratifies beyond Gleason score and CAPRA-S. Men with higher Decipher scores are at higher risk of developing metastasis (rapid or late).

As measured by Harrell's c-index for RM, Decipher (c-index: 0.77; 95% confidence interval [CI], 0.66–0.87) outperformed individual clinicopathologic variables (Supplementary Table 1) including the original pathologic Gleason score alone (c-index: 0.68; 95% CI, 0.52–0.84), expert-reviewed Gleason score regraded according to 2005 International Society of Urological Pathology criteria (c-index: 0.71; 95% CI, 0.59–0.84), CAPRA-S (c-index: 0.72; 95% CI, 0.60–0.84), and the Stephenson nomogram (c-index: 0.75; 95% CI, 0.65–0.85). The highest c-index was obtained by the combination of Decipher plus the Stephenson nomogram (c-index: 0.79; 95% CI, 0.68–0.89) ( Fig. 3 A). Moreover, among patients with low-risk Decipher scores based on cut points previously described [7] , 95% had RM-free survival, falling to 82% RM-free survival for those with high-risk scores.

Fig. 3 Discriminatory performance of Decipher compared with clinical risk factors and validated nomograms using different metrics. (A) Decipher has the highest concordance index of any individual risk factor. The performance of nomograms and Decipher is increased when integrated. (B) LASSO coefficient path of Decipher and clinical risk factors. As the penalty parameter, λ, is decreased, Decipher is the first variable to enter the model, followed by Gleason score. (C) Decision curve analysis shows the net benefit of nomograms and Decipher across probability thresholds compared withtreat allortreat nonescenarios (ie, for which no prediction model is required or used). The integrated models have the highest net benefit across threshold probabilities for rapid metastasis of 5–25%. C-index = concordance index; CI = confidence interval; EPE = extraprostatic extension; PathGS = pathologic Gleason score; PSA = prostate-specific antigen; RP = radical prostatectomy; SVI = seminal vesicle invasion; SMS = surgical margin status.

In a series of multivariable logistic regression analyses, Decipher was found to be the only significant variable for predicting RM ( Table 3 ). When adjusting for clinical risk factors, for every 0.1-unit increase in Decipher score, the odds of RM increased by 48% (odds ratio: 1.48; 95% CI, 1.07–2.05;p = 0.018). In multivariable models including only Decipher score and either the Stephenson nomogram or CAPRA-S, Decipher outperformed both ( Table 3 ). To ensure the robustness of the multivariable model, penalized LASSO regression for sparse data and rare events (ie, 15 RM events) was used to confirm the results of the logistic regression model and further demonstrate that Decipher is the most important variable for predicting RM ( Fig. 3 B). Even with large values of the penalty parameter λ, Decipher had a nonzero coefficient and is the first variable to enter the model, confirming its significance in predicting RM in multivariable analysis despite the few RM events. Following Decipher, the next two variables entering the model were EPE and high Gleason score (≥8) ( Fig. 3 B).

DCA further demonstrated increased discrimination of risk models that incorporated the genomic information captured by Decipher. Across a range of RM threshold probabilities from 5% to 25%, DCA showed that Decipher resulted in the highest net benefit compared withtreat allortreat nonescenarios (ie, in which no prediction model is required or used) and both the Stephenson and CAPRA-S clinical models ( Fig. 3 C). A combined Decipher plus Stephenson model and a combined Decipher plus CAPRA-S model each resulted in higher net benefits than individual models, demonstrating the potential clinical utility of Decipher in identifying patients with high-risk pathologic features at highest risk of RM. Because few men in this cohort developed metastatic disease even with conservative management after RP, DCA was then used to determine the number oftrue negativepatients that did not develop metastatic disease who could be spared unnecessary treatment compared with the scenario in which all patients with adverse pathology would be treated (eg, adjuvant radiation as per current guidelines). The number of interventions that could be reduced per 100 patients tested with the Decipher and Stephenson model predictions across a range of threshold probabilities for metastasis-free survival that would trigger intervention was calculated (Supplementary Fig. 3). Finally, Decipher also outperformed a panel of 19 individual biomarkers in predicting RM (Supplementary Table 2; Supplementary Fig. 4 and 5).

Recent randomized clinical trials have demonstrated the efficacy and survival benefit of RP over watchful waiting, particularly for men with intermediate- and high-risk disease[33] and [34]. The vast majority of men treated with surgery for PCa will have excellent metastasis- and disease-free survival. However, men with so-called adverse pathology that includes high-grade or non–organ-confined disease are, by current clinical practice guidelines, considered at risk for disease recurrence, metastasis, and cause-specific mortality [35] . Despite this recognized risk and level 1 evidence of clinical benefit for adjuvant radiotherapy, and unlike common practice for other prevalent cancers such as breast and colon, the use of adjuvant therapy after RP is generally deferred in the absence of detectable PSA. This approach is largely based on the belief that even in high-risk patients, as defined by adverse pathology, undetectable PSA soon after surgery indicates a low risk of developing lethal disease rapidly; it is also based on the desire to limit exposure to potentially harmful secondary therapy.

In this study, we investigated whether a genomic classifier can identify a subset of patients with a highly lethal form of metastatic disease (characterized by median onset at about 2 yr and 50% cause-specific mortality by 7 yr) who are otherwise indistinguishable by standard clinical and pathologic features, including undetectable PSA postoperatively, from those with a more indolent clinical course (ie, no recurrence or metastasis >5 yr with only 25% PCSM at 15 yr). Identification of such patients is clinically relevant because those at highest risk for early metastasis and death are most likely to benefit from effective adjuvant therapy. Notably, all of the RM patients had preoperative PSA <20 ng/ml and were clinical stage T1 or T2, and nearly half had only Gleason 7 disease on final pathology. All achieved undetectable PSA after RP, but only 27% received secondary treatment prior to the onset of metastasis. In contrast, 82% of the late metastasis group received treatment prior to metastasis. The development of novel biomarkers to identify the subset with rapidly lethal disease is clearly needed [36] because they are a clinically homogenous group (ie, all had one adverse pathology feature or more) for which nomograms are recognized to have reduced discrimination ability [37] . As such, on univariable analysis, no clinical or pathologic risk factors could be used to distinguish RM patients from patients with no or late metastasis. Consequently, it is not surprising that we observed that the Decipher genomic classifier remained the only independent variable on multivariable analyses when compared with both individual clinical risk factors or with commonly used clinical risk-assessment tools (CAPRA-S and the Stephenson nomogram). More important, DCA demonstrated that the clinical utility of Decipher in this cohort may have the most impact with its ability to identify (beyond nomograms) more patients who, despite having multiple adverse pathology findings, have a high probability of metastasis-free survival even when managed conservatively after surgery. Finally, the biologic robustness of Decipher was demonstrated by the observation that it outperforms other biomarkers, includingSPOP, Ki-67, andFOXP1, that have been associated with aggressive disease.

Unlike PSA measurements for disease monitoring, which may not indicate true disease aggressiveness until long after surgery, results of a genomic classifier derived from tumor sampled at the time of prostatectomy are available soon after surgery and thus may be used immediately for clinical decision making, including early institution of adjuvant radiotherapy or more frequent monitoring for recurrence with earlier institution of salvage therapy. Such a tool would also address the need to accurately identify patients at high risk of recurrence for clinical trials of newer adjuvant therapies, thereby enriching the event rate and potentially allowing smaller and shorter trials. The results of this study suggest that Decipher can be used as a standard tool to better risk-stratify patients based on clinical and pathologic risk factors. Recent studies show that Decipher influences physician postoperative decision making in the adjuvant setting and lowers the discrepancy in decision making between urologists and radiation oncologists[38], [39], and [40].

This study has several strengths. First, it meets both PRoBE and REMARK criteria for the prospective validation of biomarkers. Second, all results were derived and reported blind to clinical outcome. Third, it used a robust and extensively validated expression assay. Fourth, there was long clinical follow-up in the censored and LM patients. The results build on other recent studies using commercially available genomic or gene expression platforms that have shown value in improving clinical decision making for men with early stage PCa[41], [42], and [43].

There are also several limitations in this study. There were few patients with RM in this cohort and further validation may be needed, although penalized LASSO regression demonstrated the robustness of these results even after taking the rarity of RM events into account. The information on extent of positive margin was available only for about one-third of those patients with positive margins and limited us from determining whether the performance of Decipher is different among those with focal versus extensive margin positivity. In addition, tumor specimens for this study were sampled only from the primary Gleason pattern present in RP specimens, although the specified method for the commercial Decipher assay is to sample tumor cells with the highest Gleason grade of the index lesion. Ongoing studies by this group aim to more fully characterize the performance and consistency of genomic signatures among primary, secondary, and tertiary lesions and gain a better understanding of the relationship between pathologic and genomic heterogeneity and more detailed assessment of multifocality.