Google Research Introduces DeepSomatic for AI-Based Cancer Variant Detection

Google Research has introduced DeepSomatic, an AI-powered tool that identifies cancer-related genetic mutations in tumor cells. The model uses convolutional neural networks to distinguish between inherited and acquired genetic variants, improving the accuracy of cancer genome analysis across multiple sequencing technologies.

Developed in collaboration with the University of California, Santa Cruz Genomics Institute and other research partners, DeepSomatic processes sequencing data from tumor and normal cells to detect somatic variants that drive cancer growth. It can also operate in tumor-only mode, allowing analysis of cancers such as leukemia where normal cell samples are unavailable.

To train the system, researchers created the Cancer Standards Long-read Evaluation (CASTLE) dataset using six cancer samples sequenced with Illumina, PacBio, and Oxford Nanopore platforms. DeepSomatic outperformed existing tools in identifying complex mutations, achieving a 90% F1-score on Illumina data and over 80% on PacBio data for insertions and deletions.

The model also showed strong performance on challenging samples, including those preserved with formalin-fixed-paraffin-embedded methods and those analyzed through whole exome sequencing. DeepSomatic successfully generalized to new cancer types, identifying known and new variants in glioblastoma and pediatric leukemia samples.

Both the DeepSomatic tool and its training dataset have been made openly available by Google Research to support further cancer research and precision medicine development.