Abstract 4272: A novel approach to differentiation of somatic vs. germline variants in liquid biopsies using a betabinomial model

Background: A recent large study found that over 2% of advanced cancer patients have unidentified germline alterations found incidentally during next-generation sequencing (NGS) for targetable somatic alterations. However, tissue-based NGS cannot definitively distinguish germline from somatic mutations without comparison to normal tissue. Because somatic variants typically occur in plasma at allele fractions 1-2 orders of magnitude lower than germline, liquid biopsy often enables differentiation of incidental germline mutations. Not uncommonly, high tumor burden or allelic imbalance from copy number variation or loss of heterozygosity complicates simplistic differentiation based on assumption of germline 50% allele frequency. We developed a novel statistical method for discriminating between somatic and germline variants in deep sequencing data, and applied to data generated from the Guardant 360 targeted cfDNA assay. Our model dynamically incorporates prior knowledge of germline SNP variants and observed allele frequencies within a sample, and gains power in larger panels like the GuardantOmni 500-gene assay. Methods: Our method uses common heterozygous SNPs to model local germline allele count behavior, and calls variants somatic if they deviate significantly from the observed germline mutant allele fraction (MAF). A betabinomial model is well suited to this problem, because it models both the mean and variance of mutant allele counts at common SNPs. This is important since simpler methods like fixed MAF cutoffs or Poisson models may not represent the variance in molecule counts appropriately. Results: We evaluated our method against variants from 361 clinical samples that were manually annotated by reviewers with expertise in cancer genomics. Of the 11,679 heterozygous variants considered in these samples, 7,221 (62%) had two or more common SNPs in the same gene, so that a local germline heterozygous MAF could be estimated and the betabinomial model could be applied. Using the manual annotation as ground truth, we created ROC curves for a classifier based on the betabinomial model with varying p-value cutoffs, as well as a classifier based on MAF cutoffs. The AUC for the betabinomial model was 0.996, as compared with an AUC of 0.985 for MAF cutoffs. The ROC curve allowed us to choose a betabinomial p-value cutoff that maintained a similar level of specificity as a 15% MAF cutoff (99.2% for MAF cutoff, 99% for betabinomial), but with significantly increased sensitivity (89% for MAF cutoff, 97% for betabinomial). Conclusions: We have developed a novel method for differentiation of somatic versus germline variants in a single plasma sample. Identifying incidental germline alterations in advanced cancer patients may identify new opportunities to apply targeted therapy such as PARP inhibitors for BRCA1/2, and may also serve to alert physicians and their patients to familial risk. Citation Format: Tracy Nance, Elena Helman, Carlo Artieri, Jennifer Yen, Thomas P. Slavin, Darya Chudova, Richard B. Lanman, AmirAli Talasaz. A novel approach to differentiation of somatic vs. germline variants in liquid biopsies using a betabinomial model [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 4272.