Abstract 107: An expanded negative prediction algorithm for actionable mutations utilizing genomic and epigenomic profiling in cfDNA.

Abstract Introduction: When using cell-free DNA (cfDNA) for comprehensive genomic profiling, it is important to differentiate between true biomarker-negative status and absence of the biomarker due to low tumor shed. This could inform whether treatment decisions can be made in the absence of tissue testing, potentially impacting time to treatment and cost of diagnostic workup. Here we expand our ‘negative prediction’ algorithm to eleven tumor types, leveraging both epigenomic and genomic signals to provide a posterior confidence of a sample being truly negative for specific clinically actionable biomarkers. Methods: The algorithm utilizes population prevalence of specific alterations, locus-specific sequencing coverage, sample-level epigenomic tumor fraction, and sub-detection level genomic evidence to inform the estimate of an alteration and posterior probability for the absence of clonal, somatic FDA-approved biomarkers. Leveraging a clinical cohort of >80,000 cfDNA samples (Guardant360 Liquid, Guardant Health, Palo Alto, CA) across 11 tumor types, we compiled prior likelihood tables for actionable variants spanning diverse classes. This builds on the previous release on colorectal carcinoma and non-small lung cancer to include breast, prostate, pancreatic, bladder, endometrial, ovarian, gastric/gastroesophageal, cholangiocarcinoma, and melanoma tumor types. We expanded our probabilistic approach to model the likelihood of homozygous deletions, and refined our approach towards assessing microsatellite instability, fusions, and focal amplifications. Results: Positivity rates for clinically actionable biomarkers varied across tumor types from 5% in pancreatic cancer to 47% in breast cancer. Among negative samples, at least 49% of samples were associated with either a biomarker-positive or a ‘confident negative’ status (confidence >90%). Variability in confidence of negative samples was associated with tumor shedding profiles, as well as biomarker prevalence and composition, with more complex variants, such as homozygous deletions and gene fusions, reducing confidence levels. As an initial accuracy assessment, we compiled a cohort of >350 subjects with matched tissue and plasma genotyping results, among which 81/84 (96%) of confident negative samples were confirmed as negative in tissue. Finally we assessed discordances between plasma and tissue, which highlighted clonal/resistance dynamics as the major source of discrepancies. Conclusion: Here we demonstrate our ability to assess the confidence of biomarker negative samples across diverse cancer types. This helps address a known potential limitation for liquid biopsy and may be able to assist in therapy making decisions and accelerate time to treatment initiation, particularly in cases where tissue NGS is not available. Citation Format: Andrew M. Gross, Hao Wang, Brandy Freschi, Keelia Clemens, Marisa Juntilla, Martina Lefterova, Justin Odegaard, Darya Chudova. An expanded negative prediction algorithm for actionable mutations utilizing genomic and epigenomic profiling in cfDNA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 107.