Fast, accurate, and racially unbiased pan-cancer tumor-only variant calling with tabular machine learning

Accurately identifying somatic mutations is essential for precision oncology and crucial for calculating tumor-mutational burden (TMB), an important predictor of response to immunotherapy. For tumor-only variant calling, accurately distinguishing somatic mutations from germline variants is challenging and often leads to unreliable, biased, and inflated TMB estimates. This study applies machine learning to classify somatic versus germline variants in tumor-only whole-exome sequencing samples using TabNet, XGBoost, and LightGBM. Training used features derived exclusively from tumor-only pipelines with truth labels from matched-normal analysis. All models achieved state-of-the-art performance on two holdout test datasets: TCGA samples (AUC > 94%) and a metastatic melanoma dataset (AUC > 85%). Concordance between matched-normal and tumor-only TMB improves from R² = 0.006 to 0.71–0.76 with machine-learning classification, with LightGBM performing best. The models generalize across cancer subtypes and capture kits, with a 100% call rate. The study reproduces the finding that tumor-only TMB is extremely inflated for Black patients due to racially biased germline databases and shows that XGBoost and LightGBM eliminate this significant racial bias.