VariantMedium: sensitive and generalizable somatic point mutation calling with 3D DenseNets trained and evaluated on experimental data

BACKGROUND: Accurately identifying somatic variants from genomic sequencing is crucial for understanding and treating cancer. Previously, methods based on statistics and heuristics, as well as methods based on machine learning were proposed for somatic single nucleotide variant (SNV) calling from matched tumor-normal data, but they suffer from low sensitivity especially in certain genomic regions. METHODS: Here we present VariantMedium, a somatic variant caller that combines a tree-based classifier with a 3D densely connected convolutional network (DenseNet) architecture. We trained and evaluated our model on experimentally confirmed variant data and improved it with an active learning strategy by experimentally confirming the predicted variants via targeted deep sequencing experiments. Overall, we used 336,839 variants from 2,956 samples with whole exome or genome sequencing for training and validation, and 118,887 variants from two independent studies with deep sequencing data for evaluation and benchmarking. RESULTS: VariantMedium shows highest sensitivity amongst benchmarked callers and achieves similar or better F1 scores in SNV calling. Its performance is particularly pronounced on genomic regions characterized by high sequencing error rates, achieving higher F1 scores than Mutect2 and Strelka2. CONCLUSIONS: Our results demonstrate the strength of combining machine learning with high-quality experimental confirmation, enabling accurate somatic mutation detection even in low-mappability regions. We provide VariantMedium ( https://github.com/TRON-Bioinformatics/VariantMedium ) as an end-to-end pipeline to advance somatic mutation calling for precision medicine.