Google has introduced DeepSomatic, a powerful new AI model capable of pinpointing cancer-related mutations in tumor genetic sequences with greater accuracy than ever before. This tool, developed in collaboration with the UC Santa Cruz Genomics Institute and the National Cancer Institute, marks a major leap forward in using artificial intelligence for precision oncology. Understanding the exact genetic mutations driving tumor growth is crucial for designing effective treatments. Today, doctors often sequence the genomes of tumor cells to guide therapy, but identifying true mutations among millions of data points — while filtering out sequencing errors — remains a challenge. DeepSomatic aims to solve this problem.
Why Somatic Variants Matter
Most cancers arise not from inherited genes but from somatic mutations — genetic changes acquired during a person’s lifetime due to environmental damage (like UV exposure) or random errors during DNA replication. These mutations can be rare within a tumor sample, sometimes occurring at frequencies lower than the sequencing error rate itself, making them notoriously difficult to detect. DeepSomatic’s design directly targets this problem by distinguishing true variants from background noise.
Inside the DeepSomatic Model
In typical cancer analysis, researchers compare DNA from a patient’s tumor to their normal tissue to identify unique mutations driving cancer growth. DeepSomatic takes this process a step further by converting the raw sequencing data into visual representations that capture how genetic information aligns across the genome. Using convolutional neural networks (CNNs), the AI analyzes these images to distinguish normal inherited variants from cancer-causing somatic mutations while filtering out artifacts. The end result: a precise list of genetic changes responsible for tumor development. Notably, DeepSomatic can also function in a “tumor-only” mode, where normal tissue samples are unavailable — a common scenario for blood cancers like leukemia. This flexibility makes it useful across a wide range of research and clinical settings.
Building the Dataset Behind the Breakthrough
To train the model, Google and its research partners created a benchmark dataset called CASTLE. It includes tumor and normal cell sequences from breast and lung cancer samples, analyzed across multiple sequencing platforms. By merging data from three industry-leading technologies and eliminating platform-specific noise, the researchers produced an exceptionally clean reference dataset. This not only trained DeepSomatic effectively but also provided new insights into how different cancers exhibit distinct mutational “signatures.”
Outperforming Existing Methods
In testing, DeepSomatic consistently outperformed leading mutation detection tools across all major sequencing platforms. It was especially strong in identifying complex mutations known as insertions and deletions (Indels). For example, on Illumina data, DeepSomatic achieved a 90% F1-score, compared to 80% for the next-best method. On Pacific Biosciences data, the gap was even wider — over 80% versus under 50%. The AI also performed exceptionally well on challenging or degraded samples, such as those preserved using formalin-fixed-paraffin-embedded (FFPE) methods or limited exome sequencing data.
Broad Applications Across Cancer Types
Although trained on breast and lung cancer data, DeepSomatic has shown the ability to generalize to other cancers. In tests on glioblastoma (a type of brain cancer) and pediatric leukemia samples, it accurately identified known genetic drivers and even discovered new mutations. This adaptability suggests enormous potential for both clinical use and research — from improving diagnostic precision to uncovering novel therapeutic targets.
Powering the Future of Precision Medicine
By making both DeepSomatic and the CASTLE dataset open-source, Google is inviting researchers worldwide to build upon its work. The company hopes clinicians and scientists will use this technology to better understand individual tumors and design more personalized treatment plans. If widely adopted, DeepSomatic could accelerate the transition toward AI-driven precision oncology, where every patient’s treatment is tailored to the genetic makeup of their cancer — ultimately improving outcomes and uncovering new paths for drug discovery.
Source: https://www.artificialintelligence-news.com/news/google-ai-tool-pinpoints-genetic-drivers-of-cancer/
