Improved tumor-only variant calling and mutation burden estimation with VarNet-T

Abstract

Abstract Somatic variant calling algorithms typically detect mutations in cancer genomes by comparing sequence data from a tumor sample against a matched normal sample. However, matched normal samples are often unavailable in clinical diagnostics or retrospective analyses of archival tumor samples in biobanks, compromising variant calling accuracy due to the difficulty in distinguishing somatic mutations from germline mutations or sequencing artifacts. Here, we introduce VarNet-T, an end-to-end weakly supervised deep learning framework for accurately identifying somatic variants from aligned tumor reads without a matched normal sample. VarNet-T is trained using millions of high-confidence variants and benchmarked using public datasets, demonstrating 20-33% performance improvement over existing methods. We assess the accuracy of tumor mutation burden (TMB) estimation on 1000 tumor samples spanning 10 solid cancer types. Compared to existing methods, VarNet-T demonstrates >3x higher accuracy in TMB-high status classification, suggesting significant potential to improve patient selection for immunotherapy. Overall, the improved accuracy of VarNet-T has the potential to enhance the utility of tumor-only sequencing in cancer research and clinical molecular diagnostics.