Abstract
PURPOSE: Genomic profiling of tumors by liquid biopsy (LBx) is a pragmatic alternative to profiling tissue. Despite recent methodologic advances, clonal hematopoiesis (CH) variants arising from hematopoietic stem cells may confound LBx results. Distinguishing the origin of variants detected by LBx will greatly enhance treatment decision-making for patients with cancer. EXPERIMENTAL DESIGN: We sequenced DNA isolated from paired plasma and white blood cells (WBC) at equal depth to train (n = 1977) and validate (n = 658) Variant Origin Prediction (VOP), a machine learning algorithm that leverages fragmentomics to generate probabilities that a short variant (SV) detected by LBx is tumor-somatic, germline, or CH in origin. The algorithm's classifications were validated for accuracy using paired WBC DNA and for reproducibility using LBx replicates. RESULTS: We show that 68% of LBx detected at least one reportable variant of CH origin. Our fragmentomic-based algorithm differentiated reportable tumor and CH variants with high sensitivity, high positive predictive value (PPA >93%, PPV >91%), and high reproducibility (>94%). Critically, VOP performs well for SVs with VAFs ≤1% (PPV >90%), as well as in genes known to harbor both CH and tumor-somatic SVs, such as TP53 (PPV >88%). In a longitudinal cohort of 422 metastatic castration-resistant prostate cancer (mCRPC) cases, VOP accurately predicted baseline variant origins, and allowed separate tracking of tumor-somatic and CH variants, including newly detected variants, at subsequent timepoints. CONCLUSIONS: VOP is a highly accurate and reproducible method to predict the origin of SVs detected in LBx without reliance on WBC sequencing. VOP can reduce inappropriate use of targeted therapies and their toxicities for patients with variants of CH origin and enables accurate tumor profiling and monitoring.