Google AI Research Releases DeepSomatic: A New AI Model that Identifies Cancer Cell Genetic Variants

A team of researchers from Google Research and UC Santa Cruz has introduced DeepSomatic, an AI model designed to identify genetic variants in cancer cells. In collaboration with Children’s Mercy, DeepSomatic successfully identified 10 genetic variants in pediatric leukemia cells that were overlooked by existing tools. This innovative model features a somatic small variant caller optimized for cancer genomes, compatible with various sequencing technologies including Illumina short reads, PacBio HiFi long reads, and Oxford Nanopore long reads.

How It Works

DeepSomatic transforms aligned reads into image-like tensors, which encapsulate pileups, base qualities, and alignment context. A convolutional neural network (CNN) is employed to classify candidate sites as somatic or non-somatic, producing outputs in VCF or gVCF formats. This design is platform-agnostic, as the tensor summarizes local haplotype and error patterns across different sequencing technologies. The researchers emphasize the model’s ability to differentiate between inherited and acquired variants, particularly in challenging samples such as glioblastoma and pediatric leukemia.

Datasets and Benchmarking

The model was trained and evaluated using the CASTLE dataset, which contains six matched tumor and normal cell line pairs sequenced across Illumina, PacBio HiFi, and Oxford Nanopore technologies. The research team has released benchmark sets and accessions for further reuse, addressing a gap in multi-technology somatic training and testing resources.

Reported Results

The research findings indicate that DeepSomatic consistently outperforms widely used methods in detecting both single nucleotide variants (SNVs) and insertions/deletions (indels). For instance, on Illumina indels, the next best method achieves an F1 score of approximately 80%, while DeepSomatic surpasses this with about 90%. For PacBio indels, the next best method falls below 50%, whereas DeepSomatic achieves over 80%. The study identified a total of 329,011 somatic variants across the reference lines and an additional preserved sample, demonstrating DeepSomatic’s strength in indel detection.

Generalization to Real Samples

The research team also evaluated the model’s effectiveness on cancers beyond those included in the training set. For example, a glioblastoma sample demonstrated the recovery of known driver mutations, and pediatric leukemia samples tested the tumor-only workflow where no clean normal sample was available. The tool successfully recovered known calls and identified additional variants in the pediatric leukemia cohort. These evaluations suggest that the representation and training methodology generalize well to new disease contexts and scenarios without matched normals.

Key Takeaways

• DeepSomatic identifies somatic SNVs and indels across Illumina, PacBio HiFi, and Oxford Nanopore, building on the DeepVariant methodology.
• The pipeline supports both tumor-normal and tumor-only workflows, including formalin-fixed, paraffin-embedded (FFPE) whole genome sequencing (WGS) and whole exome sequencing (WES) models.
• It encodes read pileups as image-like tensors and utilizes a CNN to classify somatic sites and produce VCF or gVCF outputs.
• The training and evaluation utilized the CASTLE dataset with six matched tumor-normal cell line pairs sequenced on three platforms, with benchmarks and accessions available for reuse.
• Reported results show approximately 90% F1 score on Illumina indels and above 80% on PacBio, outperforming common baseline methods, with a total of 329,011 somatic variants identified.

Editorial Comments

DeepSomatic represents a significant advancement in somatic variant calling across multiple sequencing platforms. The model retains DeepVariant’s image tensor representation and CNN architecture, allowing consistent scaling from Illumina to PacBio HiFi and Oxford Nanopore technologies, while maintaining uniform preprocessing and outputs. The introduction of the CASTLE dataset strengthens training and benchmarking, enhancing reproducibility within the field. The reported results highlight a marked improvement in indel accuracy, effectively addressing a longstanding challenge in indel detection.

For additional technical details, datasets, and access to the GitHub repository, please refer to the original research article.