Google has launched “DeepSomatic,” a groundbreaking artificial intelligence tool designed to significantly improve the identification of mutations associated with cancer in tumor DNA sequences, offering potentially transformative implications for the treatment and management of cancer. Developed by Google researchers and detailed in a publication in Nature Biotechnology, this tool leverages convolutional neural networks (CNN) for analyzing genetic data—offering an accuracy superior to existing technologies.
Cancer arises due to mutations in genes that regulate cell growth and division, which can occur spontaneously or from environmental factors such as ultraviolet light. Current methods of cancer treatment often involve sequencing the cancerous tissue’s genome to identify these mutations, which are critical for tailoring personalized treatment strategies. However, separating genuinely causative mutations from background noise and errors in DNA sequencing has presented a significant challenge. Most concerning are ‘somatic’ mutations that are not inherited but acquired throughout a person’s life, often presenting in low frequencies that are difficult to detect amid the noise of data errors.
DeepSomatic addresses these difficulties by analyzing genetic sequencing data from both cancerous (tumor) and normal tissues to pinpoint the exact mutations driving cancer growth. The tool uses a CNN to convert sequencing data into image-like representations, which are then analyzed to differentiate between normal genetic variants and those specific to cancer, while also filtering out sequencing errors. This process results in a highly precise list of tumor-specific mutations.
Remarkably, DeepSomatic still functions effectively even in scenarios where only tumor samples are available without corresponding normal tissue references. This feature is especially crucial for types of cancer like leukemia, where obtaining normal tissue samples may be challenging. This capability ensures that DeepSomatic is applicable in a wide range of clinical and research settings.
To build such a robust model, the developers created an extensive training dataset in collaboration with the UC Santa Cruz Genomics Institute and the National Cancer Institute. Named CASTLE, this dataset includes both tumor and normal cell sequences from samples of breast and lung cancer, derived using three major sequencing platforms. This approach allowed the consolidation of data into a comprehensive, error-minimized benchmark dataset, reflecting the diversity of mutational signatures across different instances of the same cancer types. Understanding these signatures can help predict how a patient might respond to specific treatments.
In comparative tests with other established methods, DeepSomatic displayed superior performance in identifying complex types of mutations such as insertions and deletions—achieving a 90% F1-score in tests using Illumina sequencing data versus 80% for its closest competitor. The improvement was even more pronounced in data from Pacific Biosciences, with DeepSomatic scoring over 80% compared to less than 50% by other methods. It also demonstrated strong results in challenging situations such as samples preserved by formalin-fixed-paraffin-embedded (FFPE) methods, which can damage DNA, and in whole exome sequencing (WES), which targets only the protein-coding regions of the genome.
Additionally, DeepSomatic proved adaptable to new, unstudied types of cancer during its evaluations. For instance, it effectively identified known driver mutations in glioblastoma, a severe form of brain cancer, and in pediatric leukemia samples, where it not only detected known variants but also discovered ten new potential mutations.
Google aims for DeepSomatic to be widely adopted in clinical and research environments, enhancing the understanding of individual tumors and aiding in the development of tailored treatments. By accurately identifying both known and novel cancer-driving mutations, DeepSomatic has the potential to significantly advance precision medicine, leading to more effective and personalized cancer therapies. This initiative reflects a broader goal within the healthcare and technology sectors to integrate advanced computational tools like AI into practical, patient-centered applications, promising a future where technology and medicine converge for better patient outcomes.
Read the full post on artificialintelligence-news.com
