Development and validation of computational histology artificial intelligence (CHAI)–powered prognostic and predictive biomarkers in metastatic hormone-sensitive prostate cancer (mHSPC) using ENZAMET and CHAARTED prospective randomized phase 3 trials (RCT).

Abstract

231 Background: Advanced biomarkers (BM) for mHSPC are needed to improve personalized treatment. The CHAI platform applies deep learning to extract quantitative histologic features from H&E stained digitized whole slide images (WSI). Using two RCTs, we aimed to apply the platform to separately develop (dev) and externally validate (val) three distinct BMs in mHSPC: 1) a prognostic risk stratifier (ProgPC), and predictors for treatment benefit with 2) docetaxel (PredDoce), and 3) androgen-receptor pathway inhibitor (ARPI) (PredARPI). Methods: Dev and val used all cases with available H&E-stained diagnostic specimen WSI and clinical data from CHAARTED, ENZAMET, and a real world dataset (RWD) of mHSPC from an NCI Center (Table). For dev, the CHAI platform extracted quantitative histomorphologic features. Features associated with progression-free survival (PFS) & overall survival (OS) were selected to produce continuous scores which were then dichotomized into BM+ (benefit) and BM- (less benefit) categories. BMs and thresholds were optimized on the dev sets and locked. Each val used an independent held-out cohort to assess performance. Predictive performance was assessed via BM-treatment interaction in Cox-proportional hazards models. Results: 1179 pts were available: 507 CHAARTED; 584 ENZAMET; and 88 RWD. In val, for ProgPC, unfavorable pts had worse PFS & OS, even after controlling for clinical variables on multivariable analysis (MVA) (Table). For PredDoce, BM+ pts had superior PFS & OS, while BM- pts had no difference with the addition of docetaxel. For PredARPI, BM+ pts had superior PFS & OS, while BM- had less benefit with the addition of ARPI. For both predictive BMs, the BM-treatment interaction was significant for PFS & OS even when controlling for clinical predictors. Conclusions: We separately developed three distinct and clinically relevant BMs for mHSPC using a deep learning-based computational histology platform. All three externally validated for PFS and OS using two independent prospective phase 3 RCTs, and independently of clinicopathologic risk factors, demonstrating Simon's level IB evidence for biomarker validation. Clinical trial information: NCT00309985 ; NCT02446405 . Biomarker Dev Val Biomarker N (%) OS HR (95%CI) P Interaction P ProgPC CHAARTED ENZAMET Favorable 465 (80) Unfavorable 119 (20) 2.9 (2.2-3.8) <0.01* PredDoce ENZAMET(no enzalutamide arm) CHAARTED Benefit 260 (60) 0.60 (0.41-0.87) <0.01 0.02** Less benefit 163 (40) 1.03 (0.69-1.55) 0.8 PredARPI RWD + ENZAMET(15%) ENZAMET(85%) Benefit 224 (65) 0.5 (0.32-0.97) <0.01 0.01 Less benefit 121 (35) 1.22 (0.70-2.12) 0.5 *Control for age, ECOG, Gleason, PSA, treatment, volume (low/high), timing (metachronous/synchronous). **Control for volume*treatment, and timing*treatment interaction.