Abstract 1959: Variant Origin Prediction (VOP) determines the tumor-somatic, germline, or clonal hematopoiesis (CH) origin of short variants (SVs) detected in liquid biopsy (LBx)

Abstract Introduction: Distinguishing the origin of variants detected by LBx is essential as tumor-somatic and germline, but not CH variants, can provide prognostic insights and inform treatment decisions. We describe a machine learning model that generates variant origin predictions for SVs (including base substitutions and short indels) detected by LBx. Method: For all SVs detected in the FoundationOne®Liquid CDx (F1LCDx) assay, VOP generates a 3-way probability that the variant originated as germline, CH, or tumor-somatic. Tumor-somatic and germline probabilities can be summed to predict that the SV is in the tumor. This model was trained using 1977 F1LCDx results for which matched buffy coat (BC) DNA was also sequenced to equal depth. A variety of features, including fragmentomics and variant allele frequency, but no patient demographics, were used to train the classification model. Probability cutoffs were applied to pathogenic SVs and validated for accuracy from paired BC (658) and reproducibility using LBx replicates. Results: VOP made accurate best call predictions (>50% probability cutoff) for pathogenic SVs across all origin types, with high sensitivity and high positive predictive value of tumor and CH (PPA > 93%, PPV > 91%). 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%). For a use case requiring higher confidence in tumor origin, an 80% tumor cutoff maintained high PPA (90%) while increasing PPV overall (99%), including low VAF and TP53 SVs (PPV > 96%). These cutoffs were assessed in 31 unique samples across 546 replicates and showed high reproducibility (Table). Conclusion: VOP is a highly accurate and reproducible method to predict the origin of SVs from LBx alone. VOP performs well even when VAF ≤ 1% and in heavily confounded genes such as TP53. Increased confidence in variant origin is valuable for improving clinical care. Citation Format: Derek W. Brown, Alexander D. Fine, Daokun Sun, Hanna Tukachinsky, Michael McDevitt, Kerriann Pontbriand, Eliana Polisecki, Angela Kou, Mingyue Li, Shumeng Zhang, Zheng Kuang, Aparna Aiyer, Bahar Yilmazel, Brennan Decker, Jason Hughes, Chang Xu. Variant Origin Prediction (VOP) determines the tumor-somatic, germline, or clonal hematopoiesis (CH) origin of short variants (SVs) detected in liquid biopsy (LBx) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 1959.