Using AI to identify genetic variants in tumors with DeepSomatic (research.google)

Google Research and partners announced DeepSomatic, a convolutional neural network–based tool for detecting somatic (tumor-acquired) small variants that outperforms existing callers across major sequencing technologies. Described in Nature Biotechnology, DeepSomatic converts aligned sequencing reads into image-like representations and trains CNNs to distinguish reference sequence, germline variants, true somatic mutations and sequencing errors. The team also released a new high-quality multi-platform truth set, CASTLE, built from whole-genome data on six cancer cell lines (Illumina, PacBio, Oxford Nanopore) to support training and evaluation. Technically, DeepSomatic supports tumor-normal and tumor-only workflows, handles short- and long-read data, and tolerates challenging clinical inputs such as FFPE-preserved and whole-exome sequenced samples. In benchmarks it detected 329,011 somatic variants across test sets and substantially improved indel F1-scores (Illumina: 90% vs ~80% next-best; PacBio: >80% vs <50% next-best) compared with tools like MuTect2, Strelka2, SomaticSniper and ClairS. It generalized to glioblastoma and identified known plus 10 novel variants in pediatric leukemia samples without matched normal. By improving sensitivity and precision for somatic calls across platforms and sample qualities—and releasing both model and dataset—DeepSomatic aims to accelerate precision oncology research and rescue harder-to-analyze clinical or historical samples.