Abstract Background: Liquid biopsy has been established as a powerful, non-invasive means to profile tumors in and identify clinically relevant variants. However, the presence of clonal hematopoiesis (CH) variants, or biological noise, due to aging or therapy has potential to confound biomarker interpretation. Currently, comprehensive methods to filter out non-tumor variants require genotyping the white blood cell (WBC) fraction of the paired plasma sample, which is a costly, complicated workflow. A plasma-only, bioinformatics solution to identify non-tumor variants is needed for accurate biomarker assessments in the cell-free DNA (cfDNA). Method: An ensemble model was trained on a database of >250,000 plasma samples comprising healthy donor, early and late-stage cancer patients sequenced on the Guardant360TM , GuardantREVEALTM, and GuardantOMNITM liquid biopsy panels as well as public tissue datasets. The model was optimized with 5 fold cross-validation and hyperparameter tuning to produce a non-tumor and tumor variant classifier. To validate these calls, 116 paired plasma and WBC advanced cancer samples were selected for high prevalence of putative CH variants and genotyped using an in-house bioinformatics pipeline. In the validation cohort, cfDNA variants were determined to be of non-tumor or CH origin if there was molecule support in the WBC; cfDNA variants above 0.6% (limit of detection in the gDNA) with no support in the WBC were determined to be from the tumor. Results: The validation cohort consisted of 2150 somatic SNV and Indels, 956 of which were confirmed in the WBC and 1194 confirmed as plasma-only. Half of confirmed CH variants (48%, 458/956) occurred in known CH genes, while the other half occurred in genes such as TP53, ATM, NOTCH4, FAT1, SRSF2. No clinically actionable variants were confirmed in the WBC. Non-tumor or CH predictions were made for 624 somatic variants: 515/624 correctly identified CH, for a positive predictive value (PPV) of 83%. Of all CH variants confirmed in the WBC, 54% (553/956) had a CH or non-CH prediction; CH predictions had 91% (515/553) positive percent agreement (PPA) with the WBC. Remaining variants with no CH prediction (403/956) were low or no prevalence across datasets and occurred predominantly in LRP1B, TET2, TP53, KMT2D. Nearly half (67%, n=109) of CH predictions not in WBC occurred in a CH gene. For non-CH gene variants, 16% of false positive predictions occurred in 6 variants across 4 genes (ACVR2A, RNF43, B2M, FLT3). Conclusion: We present a plasma-only method that has high PPA and PPV with WBC genotyping for classifying non-tumor, CH variants in the cfDNA. Further investigation is underway to improve the sensitivity of annotating rare CH variants. Accurate CH identification is critical for treatment selection across targeted therapies, particularly for loss of function variants in DNA repair genes that may confer sensitivity to PARPi or ATRi therapies. Citation Format: Jennifer Yen, Yu Fu, Jeff Werbin, Katie Quinn, Robert Foley, Minh Trahn, Carin Espenschied, Scott Higdon, Brett Kennedy, Han-Yu Chuang. Validation of a bioinformatic model for classifying non-tumor variants in a cell-free DNA liquid biopsy assay [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 3135.