Introduction

To develop and validate novel genomic signatures to predict adverse pathology (AP) after radical prostatectomy (RP).

Methods

We developed three classifiers aimed at predicting 1) high-grade disease on final pathology (primary Gleason grade 4/5) and 2) AP (seminal vesicle or bladder neck invasion or lymph node positive disease). For the development of the Grade Group (GG) classifier, we performed genome-wide differential expression analysis between high- (primary pattern 4/5) and low- (primary pattern 3) Gleason grade disease on 967 high-risk prostate cancer (PCa) patients from the Decipher genomic resource information database (GRID®, NCT02609269). For the development of the high-grade disease 1 (HGD) classifier, we used differential expression analysis comparing high (≥7) and low (≤4) CAPRA-S scores in a retrospective cohort of 425 high-risk RP patients treated at Mayo Clinic and for HGD 2 classifier, we compared patients with a Decipher score ≥0.7 and Gleason score ≥8 to those with lower-risk scores in _x0007__x0008_a prospective cohort of 1817 high-risk RP patients available in GRID. The GG classifier was trained using a deep neural network model with 3 hidden layers consisting of 79 genes using the &[Prime]h2o&[Prime] package in R v3.1. The HGD 1 and HGD 2 classifiers were based on an elastic net algorithm consisting of _x0007_109 _x0008_and _x0007_60 _x0008_genes, respectively. Two validation cohorts were used: 1) 6,739 Decipher RP profiles (not used in training) in the GRID ordered by 285 centers over 2015-2016 and 2) a retrospective biopsy (Bx) cohort consisting of 107 Decipher biopsy profiles from 4 academic institutions. Discrimination performance of the classifiers was evaluated via the c-index and logistic regression.

Results

On the RP validation cohort, the GG, HGD 1 and HGD 2 classifiers had c-indices of 0.65 (95% confidence interval [CI] 0.64-0.67), 0.67 (95% CI 0.65-0.69) and 0.69 (95% CI 0.67-0.70) for prediction of the AP endpoint, respectively. On the Bx cohort, the GG, HGD 1 and HGD 2 classifiers had c-indices of 0.80 (95% CI 0.70-0.90), 0.82 (95% CI 0.71-0.93) and 0.84 (95% CI 0.74-0.93), respectively, for the prediction of AP after surgery.

Conclusions

GG and HGD classifiers were all predictive of adverse pathology on RP using two validation cohorts of RP and Bx specimens with high accuracy. This study shows how prospective data may be used to develop and validate novel genomic classifiers that provide an independent assessment of high-risk disease at the time of initial diagnosis that complements clinical risk factors and prognostic test to further help optimize treatment decision-making.

Funding

none