Abstract 6603: Analytical validation of a new plasma-only bioinformatics classifier to identify variants from clonal hematopoiesis (CH) in cfDNA liquid biopsy assays

Abstract Introduction: Distinguishing between tumor derived variants and those from clonal hematopoiesis (CH, or CHIP) can impact clinical decisions. It is essential to develop tools that can accurately differentiate tumor vs CH in clinical tests with expanded gene panels. One method to confirm CH involves sequencing both the buffy coat and plasma samples. However, this approach is cost-prohibitive and has two key limitations: detection sensitivity of buffy coat sequencing for low allele fractions, and the possibility that variants originate from tissues other than the immune cells present in the buffy coat. To address this, we developed a plasma only CH classifier to detect CH variants with high specificity in the Guardant360 test for late-stage cancer patients. Methods: First, limitations of the buffy coat sequencing were explored by sequencing over 60 healthy donor plasma and buffy coat samples. Second, data with more than 2000 variants observed in over 200 clinical samples across more than 15 cancer types (median age of 69) with paired buffy coats and plasma were utilized to train a boosted tree classifier. The final model was built using 14 features which included variant-level metrics derived from fragmentomics, methylation patterns, and variant allele frequencies, as well as metadata features aggregated from over 250, 000 Guardant Health samples and public databases such as COSMIC and GnomAD. The classifier was trained to predict CH variants with high specificity to ensure that informative tumor-derived variants are accurately reported. Results: In 62 healthy donors, we observed that 18% of variants detected in plasma at low allele frequency were not detected in buffy coat, highlighting the need for bioinformatic solutions to capture CH variants. Our bioinformatic approach was validated on an independent cohort of 173 Guardant360 late-stage, pan-cancer samples, each with paired buffy coat and plasma samples. In a set of clinically actionable genes, consisting of National Comprehensive Cancer Network (NCCN) recommended genes for biomarker testing and canonical CH-associated genes, the classifier successfully detected CH variants with a specificity of 98% and a positive predictive value (PPV) of 99%. Conclusion: We present here a plasma only CH classifier developed for late-stage cancer patients run on Guardant360 powered by Infinity. The performance of the classifier to detect non-tumor variants in the Guardant360 test achieves a specificity exceeding 98% in clinically relevant genes, with a PPV of 99% in detecting CH variants. Citation Format: Patrick C. Fiaux, Mingyang Cai, Pegah Safabakhsh, Jeffrey L. Werbin, Che-Yu Lee, Katie Quinn, Marisa Juntilla, Tingting Jiang, Reagan M. Barnett, Errin Lagow, Arielle Yablonovitch, Shile Zhang, Martina Lefterova, Darya Chudova. Analytical validation of a new plasma-only bioinformatics classifier to identify variants from clonal hematopoiesis (CH) in cfDNA liquid biopsy assays [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 6603.