DeepSomatic

Status: assess | Maturity: emerging | Last reviewed: 2026-03-20

What it does

DeepSomatic is a deep learning-based somatic variant caller from Google, designed to identify somatic mutations (variants present in tumour DNA but not germline) from whole-genome or whole-exome sequencing data. It extends the earlier DeepVariant framework, applying convolutional neural networks to classify candidate variants in cancer genomics datasets.

It supports multiple sequencing platforms including Illumina short-read, PacBio HiFi, and Oxford Nanopore long-read data.

Key capabilities

• Somatic variant calling from tumour-normal paired sequencing
• Support for short-read and long-read sequencing platforms
• Deep learning approach reduces false positive rate versus traditional heuristic callers
• Open-source code deployable on local HPC or cloud
• Docker and Singularity container support for reproducible pipelines

Fit for Sheffield researchers

Relevant to cancer genomics and oncology research groups within the School of Medicine and Dentistry (now Faculty of Health) and Sheffield Institute for Translational Neuroscience. Also relevant to groups in Biosciences working on somatic evolution.

Requires bioinformatics expertise to deploy and interpret — not a point-and-click tool. Should be evaluated by Sheffield Bioinformatics Hub or similar technical research support. Local deployment on Stanage HPC would be the recommended route to avoid uploading sensitive patient-derived data to cloud infrastructure.

Licensing and access

Open source under BSD 3-Clause licence. Available on GitHub. Containerised via Docker and Singularity. Google Cloud Batch execution option available for teams with cloud compute budgets.

Trust and governance notes

Patient-derived genomic data requires careful governance. Local deployment (HPC) is strongly recommended over cloud execution for any data derived from NHS/clinical samples. Ensure compliance with Sheffield/NHS data sharing agreements before processing clinical data. DeepSomatic itself does not phone home when run locally — governance risk lies in the underlying data, not the tool.

GDPR status marked unclear as no explicit vendor compliance statement found; local deployment mitigates most exposure.

Observed limitations

• Requires matched tumour-normal pairs for optimal performance (tumour-only mode available but less accurate)
• High compute requirements for large cohorts
• Primarily validated on standard sequencing protocols — exotic library preps may degrade performance
• Requires bioinformatics pipeline expertise (not self-service)
• Relatively new tool — limited independent benchmarking outside Google-published evaluations

Related tools

alphafold · labos

Sources and evidence

• DeepSomatic GitHub repository: https://github.com/google/deepsomatic (accessed 2026-03-20)
• DeepSomatic preprint: Nattestad et al., 2024: https://www.biorxiv.org/content/10.1101/2024.01.16.575956v1
• DeepVariant background: https://github.com/google/deepvariant (accessed 2026-03-20)