Google AI Maps Genetic Mutations Driving Cancer Development

Google has introduced DeepSomatic, an AI-powered tool designed to detect cancer-related mutations in tumor genetic sequences with improved accuracy.

Cancer begins when the mechanisms controlling cell division break down. Identifying the specific genetic mutations responsible for a tumor's growth is critical for developing effective therapies. Today, physicians routinely sequence the genomes of tumor cells from biopsies to guide treatments that target how a particular cancer develops and spreads.

Published in Nature Biotechnology, this research introduces a tool that uses convolutional neural networks to spot genetic variants in cancer cells more precisely than current approaches. Google has released both DeepSomatic and the high-quality training dataset developed for it as open resources.

The challenge of somatic variants

Cancer genetics is a complex field. Although genomic sequencing identifies cancer-related genetic alterations, telling true variants apart from sequencing errors remains difficult—a challenge where AI tools can offer valuable support. Most cancers are driven by 'somatic' variations, which are acquired after birth, rather than inherited 'germline' mutations passed down from parents.

Somatic mutations occur when environmental factors—such as UV radiation—damage DNA, or when random mistakes arise during DNA replication. When these variations disrupt normal cellular processes, they may trigger uncontrolled growth, fueling cancer initiation and progression.

Detecting somatic mutations is more difficult than identifying inherited ones, since they can appear at low frequencies within tumor cells—sometimes even below the sequencing error threshold.

How DeepSomatic works

In clinical practice, scientists sequence both tumor cells from a biopsy and the patient’s normal cells. DeepSomatic compares these sequences to detect changes present only in the tumor—variations not inherited. These unique changes help reveal what is driving the tumor's development.

The model transforms raw genetic sequencing data from tumor and normal samples into images that represent multiple data types, including sequencing reads and their alignment along chromosomes. A convolutional neural network then examines these images to distinguish the standard reference genome, inherited variants, and cancer-causing somatic mutations—while filtering out sequencing noise. The final output is a list of cancer-associated mutations.

DeepSomatic can also operate in 'tumor-only' mode when normal tissue samples are unavailable, a situation common with blood cancers like leukemia. This flexibility makes the tool applicable across a wide range of research and clinical settings.

Training a more precise AI cancer research tool

Building an accurate AI model requires high-quality training data. For this tool, Google collaborated with the UC Santa Cruz Genomics Institute and the National Cancer Institute to create a benchmark dataset called CASTLE. The team sequenced tumor and normal cells from four breast cancer and two lung cancer samples.

These samples were profiled using three leading sequencing platforms, and the outputs were combined to create a unified, high-confidence reference dataset, free of platform-specific artifacts. This data illustrates how even cancers of the same type can have very different mutation patterns—insights that may help predict how individual patients respond to certain treatments.

Across all three major sequencing platforms, DeepSomatic models outperformed other established techniques. The tool proved especially effective at detecting complex mutations known as insertions and deletions, or 'Indels.' For these variants, DeepSomatic achieved a 90% F1-score on Illumina sequencing data, compared to 80% for the next best tool. Performance gains were even more significant on Pacific Biosciences data, where DeepSomatic scored above 80%, while the runner-up reached less than 50%.

The AI also performed well with challenging samples. Tests included a breast cancer sample preserved via formalin-fixed-paraffin-embedding (FFPE)—a routine method that can cause DNA damage and complicate analysis. DeepSomatic was also evaluated on whole exome sequencing (WES) data, a cost-effective technique that sequences only the protein-coding 1% of the genome. In both situations, it surpassed competing tools, indicating its usefulness for analyzing lower-quality or archival samples.

An AI tool for all cancers

The tool has demonstrated it can generalize its learning to cancer types it was not explicitly trained on. When applied to a glioblastoma sample—an aggressive brain cancer—it successfully identified the small number of driver mutations known to cause the disease. In collaboration with Children’s Mercy Kansas City, DeepSomatic analyzed eight pediatric leukemia samples and not only confirmed previously known mutations but also discovered 10 new ones—even when working only with tumor data.

Google hopes research labs and clinicians will adopt this tool to gain deeper insights into individual tumors. By recognizing known cancer mutations, it could help inform choices among existing treatments. By uncovering new ones, it might open doors to novel therapies. The ultimate aim is to advance precision medicine and bring more effective care to patients.

See also: MHRA fast-tracks next wave of AI tools for patient care

Interested in learning more about AI and big data from industry experts? Check out the AI & Big Data Expo happening in Amsterdam, California, and London. This comprehensive event is part of TechEx and runs alongside other top tech events including the Cyber Security Expo. Click here for further details.

AI News is powered by TechForge Media. Discover other upcoming enterprise technology events and webinars here.