Cancer treatment increasingly depends on finding the genetic changes that make an individual tumour grow. Sequencing can reveal those changes, but it also produces errors and inherited variants that are not driving the disease. The difficult task is separating the crucial signal from biological and technical noise.
Google’s DeepSomatic applies convolutional neural networks to that distinction. Published in Nature Biotechnology, the openly available tool analyses tumour sequencing data and produces a list of acquired, or somatic, mutations. Google has also released the high-quality dataset created to train and evaluate it.
The aim is practical precision medicine: identify known mutations that may guide an existing treatment and discover unfamiliar ones that could inform future therapies. Better detection does not select a therapy by itself, but it can give clinicians and researchers a clearer genetic description of the cancer they are confronting.
Somatic Variants Hide Below the Noise
Most cancer-driving changes are acquired after birth rather than inherited from a parent. Ultraviolet light can damage DNA, and random copying errors can occur as cells divide. When a somatic mutation disrupts the controls governing cell behaviour, uncontrolled replication can follow.
These variants may appear in only a small fraction of cells within a tumour sample. Their frequency can be lower than the sequencing platform’s own error rate, making a genuine driver difficult to distinguish from a technical mistake. Scientists usually compare a tumour biopsy with normal cells from the same patient to identify changes unique to the cancer.
DeepSomatic converts raw sequencing information from tumour and normal samples into visual representations. The images capture the reads and how they align along a chromosome. A neural network then distinguishes the reference genome, the patient’s inherited variants, somatic changes, and likely sequencing errors.
A Shared Dataset Makes Accuracy Measurable
Google worked with the UC Santa Cruz Genomics Institute and the National Cancer Institute to create CASTLE, a benchmark built from four breast-cancer and two lung-cancer samples. Each was analysed using three leading sequencing platforms, allowing the partners to combine results and filter platform-specific errors into a trusted reference.
DeepSomatic outperformed established methods across all three platforms. Its advantage was pronounced for insertions and deletions, complex mutations known as Indels. On Illumina data, it reached a 90 percent F1 score compared with 80 percent for the next-best method. On Pacific Biosciences data, it scored above 80 percent while the next alternative remained below 50 percent.
The model also handled difficult material. It performed strongly on a breast-cancer sample preserved through formalin-fixed, paraffin-embedded processing, which can damage DNA, and on lower-cost whole-exome sequencing that reads only the protein-coding one percent of the genome.
Tumour-Only Analysis Broadens the Reach
Normal tissue is not always available, particularly for blood cancers such as leukaemia. DeepSomatic includes a tumour-only mode for those cases. In work with Children’s Mercy in Kansas City, it analysed eight paediatric leukaemia samples, recovered the previously known variants, and identified ten new ones despite lacking matched normal samples.
The tool also generalised beyond cancers in its training set. When applied to glioblastoma, an aggressive brain cancer, it located the small number of variants already known to drive the disease. Those tests suggest the model has learned useful patterns rather than simply memorising the six CASTLE tumours.
Open access to both software and benchmark data lets laboratories examine that generalisation for themselves. Clinical adoption will still require careful validation within particular sequencing workflows and cancer populations, especially when a result could influence treatment.
DeepSomatic’s significance lies in strengthening a narrow but essential link in the care chain. It turns complicated sequencing evidence into a more accurate candidate list for expert interpretation. By reducing missed mutations and false signals, AI can help researchers see what is unusual about one tumour—and bring treatment one step closer to the biology of the individual patient.
