Evaluation of a variant origin prediction (VOP) algorithm to distinguish clonal hematopoiesis (CH) variants from tumor-derived variants and to predict metastatic castrate resistant prostate cancer (mCRPC) clinical responses.

Abstract

201 Background: VOPis an algorithm that predicts the cellular origin of variants, currently available on FoundationOneLiquid CDx (F1LCDx) for research use only. IMbassador250 (IM250, NCT03016312) is a completed phase III trial that evaluated the safety and efficacy of atezolizumab in combination with enzalutamide for men with mCRPC who had prior progression on abiraterone. Here, we evaluated the prediction accuracy and demonstrated clinical validity of VOP with IM250 samples. We hypothesized that excluding predicted CH variants from maximum variant allele frequency (maxVAF) calculations would strengthen the association of reduction of maxVAF and clinical outcome and lead to better on treatment risk stratification. Methods: We developed VOP, a machine learning algorithm that classifies short variants into tumor somatic, CH, and germline categories based on fragmentomics and other features. We applied VOP to banked IM250 plasma samples from cycle 1 day 1 (C1D1) and cycle 3 day 1 (C3D1, 6 weeks on treatment) timepoints profiled by F1LCDx, and sequenced matched whole blood (WB) from a subgroup of patients to definevariant origin ground truth for an accuracy assessment. To assess clinical validity of VOP, we calculated maxVAF with and without filtering out CH variants predicted by VOP and assessed its association with clinical outcome. Results: Based on over 2,700 short variants in 221 patients with matched WB as truth, VOP achieved a positive percent agreement (PPA) of 92% and positive predicted value (PPV) of 94% for tumor somatic variant predictions (median VAF 3.6%). PPA and PPV were 90% and 88%, respectively, for CH variant predictions (median VAF 0.7%), and over 98% for germline variant predictions, consistent with past development data based on a pan-cancer cohort. To assess potential clinical impact, we applied VOP to 422 patients with F1LCDx results at both C1D1 and C3D1. Patients with at least 50% decrease in maxVAF were associated with longer overall survival. Importantly, CH-adjusted maxVAF led to better patient stratification (Hazard Ratio (HR) = 0.36 [0.28, 0.47], p = 0.0007) than non-CH-adjusted maxVAF (HR = 0.60 [0.45, 0.81], p < 0.0001). We observed similar results using cutoffs of 90% and 100% (ctDNA clearance) decrease in maxVAF and with radiographic progression-free survival as the endpoint. Conclusions: VOP had strong analytical concordance and clinical applicability in an independent mCRPC cohort (IM250). With the aid of VOP, CH-adjusted maxVAF more effectively identifies patients with better outcomes. The VOP algorithm is accurate, robust, and has potential clinical use in tumor monitoring, clinical outcomes and on treatment patient risk stratification.