Advancing the drug development process in metastatic prostate cancer through machine learning.

Abstract

192 Background: Despite advances in the development of systemic therapeutics, metastatic prostate cancer (mPC) remains a lethal disease. The drug approval process is slow, and delays in approving active agents translate into lives lost, and increased morbidity for men living with mPC. Clinical benefit end points proximate to survival that could support regulatory approval have been sought. Methods that accelerate the prediction of overall survival (OS) from baseline and short-term clinical data could dramatically accelerate the "readout" timeline for registrational studies. Methods: Curated, anonymized data from 5,679 mPC patients (pts) across six completed clinical trials were used: COU-AA-301, COU-AA302, ACIS, GALAHAD, LATITUDE, and TITAN. These trials comprise heterogeneous pts across castrate sensitive (TITAN, LATITUDE) and castrate resistant (COU-AA-301, COU-AA-302, ACIS, GALAHAD) populations. Transcriptomic profiles of primary tumors were available from a subset of these pts (N = 605) in COU-AA-302, ACIS, TITAN using the DECIPHER platform (Decipher Biosciences, Inc.). We harmonized pts study data and are developing machine learning models in an attempt to predict long term outcomes using baseline clinical features, short term PSA kinetics, and transcriptional profiling data. Associations between clinical and transcriptional signature profiles and overall survival were assessed with the Cox proportional hazards model. Results: Pts were assigned to training (N = 3167) and test (N = 2,512) cohorts. Cohorts were balanced by age, self-reported race, ECOG performance status, extent of disease, and disease site, and had statistically indistinguishable survival curves. At month 3 on study, 97% of pts had PSA data available; 88% had ≥ 2 PSA values. Matched genomic data was evaluated in 504 and 101 pts in training and test sets respectively. Training cohort pts with a 50% reduction in PSA by month 3 (m3-PSA50) had significantly longer OS (HR 0.51 [0.38-0.67]; p = 1.2 x 106, median 26 vs. 38 months). ANOVA analysis identified expression signatures in the Decipher GRID that significantly improved OS model performance when added to a model including only m3-PSA50. We stratified pts by a combined m3-PSA50 and signature score and identified signatures that improved model fit over m3-PSA50 alone and significantly stratified pts in the test cohort (P < 0.05). Conclusions: In this early analysis (prior to applying machine learning), m3-PSA50 was significantly associated with OS in a heterogeneous group of mPC pts. Although the transcriptomic analysis was conducted on primary tumors obtained many years earlier, we identified transcriptional signatures that significantly improved OS models (independent of m3-PSA50) in men who subsequently developed mPC many years later. These results suggest adding genomic signature data to models for predicting outcome using PSA kinetics may provide improved predictive power.