Attentive deep learning-based tumor-only somatic mutation classifier achieves high accuracy agnostic of tissue type and capture kit

In precision oncology, reliable identification of tumor-specific DNA mutations requires sequencing tumor DNA and non-tumor DNA (so-called “matched normal”) from the same patient. The normal sample allows researchers to distinguish acquired (somatic) and hereditary (germline) variants. The ability to distinguish somatic and germline variants facilitates estimation of tumor mutation burden (TMB), which is a recently FDA-approved pan-cancer marker for highly successful cancer immunotherapies; in tumor-only variant calling (i.e., without a matched normal), the difficulty in discriminating germline and somatic variants results in inflated and unreliable TMB estimates. We apply machine learning to the task of somatic vs germline classification in tumor-only samples using TabNet, a recently developed attentive deep learning model for tabular data that has achieved state of the art performance in multiple classification tasks (Arik and Pfister 2019). We constructed a training set for supervised classification using features derived from tumor-only variant calling and drawing somatic and germline truth-labels from an independent pipeline incorporating the patient-matched normal samples. Our trained model achieved state-of-the-art performance on two hold-out test datasets: a TCGA dataset including sarcoma, breast adenocarcinoma, and endometrial carcinoma samples (F1-score: 88.3), and a metastatic melanoma dataset, (F1-score 79.8). Concordance between matched-normal and tumor-only TMB improves from R 2 = 0.006 to 0.705 with the addition of our classifier. And importantly, this approach generalizes across tumor tissue types and capture kits and has a call rate of 100%. The interpretable feature masks of the attentive deep learning model explain the reasons for misclassified variants. We reproduce the recent finding that tumor-only TMB estimates for Black patients are extremely inflated relative to that of White patients due to the racial biases of germline databases. We show that our machine learning approach appreciably reduces this racial bias in tumor-only variant-calling.