News

AI Models Trained on Tumor Mutations Show Promise in Predicting Cancer Prognosis and Treatment Response

• By John K. Waters
• 08/01/2025

A new generation of artificial intelligence models trained on tumor-specific somatic mutations is demonstrating potential to transform cancer care by predicting prognosis and optimizing therapeutic choices based on individual patient profiles.

At the recent AACR Special Conference on Artificial Intelligence and Machine Learning in Cancer Research, investigators presented findings on a large language model (LLM) designed to analyze genomic data from thousands of tumors. The model, developed at Weill Cornell Medicine, was trained to interpret the "language" of cancer mutations in the context of next-generation sequencing (NGS) data.

Unlike traditional genomic tools that rely on a pre-defined list of actionable mutations, the AI model is designed to learn the biological meaning of novel or poorly understood mutations. It operates by embedding somatic alterations into a dual-attention transformer architecture, allowing it to evaluate both local mutation characteristics and global co-occurrence patterns. By doing so, it generates a unique mutational fingerprint for each tumor.

The model was initially trained using over 3 million somatic variants across more than 10,000 patients from The Cancer Genome Atlas. It was subsequently tested against the BeatAML2 dataset, which includes genomic and drug response data from patients with acute myeloid leukemia. In this cohort, the model demonstrated the ability to predict responsiveness to the drug cabozantinib, yielding statistically significant correlations between predicted and observed outcomes.

Using unsupervised clustering techniques, the model was also able to group patients by prognosis and generate Kaplan-Meier survival curves. Specifically, in colorectal cancer, it identified mutational dependencies consistent with the established Vogelstein model of carcinogenesis, capturing the sequence and interaction of key gene mutations such as TP53, KRAS, and APC.

Researchers highlighted the model’s capacity to classify variants of uncertain significance without relying on prior annotation, indicating that AI can offer clinically relevant insights into poorly characterized genomic regions. This ability to "read" cancer genomes without human curation could position AI as a valuable tool for identifying new targets, understanding resistance mechanisms, and guiding treatment in real time.

The implications extend into drug discovery and personalized medicine. Other researchers at the conference highlighted how AI tools are being used to evaluate neoantigens for cancer vaccines, analyze resistance pathways in prostate and pancreatic cancer, and design optimized antibodies and drug compounds using machine learning-directed evolution.

In a separate study presented by researchers from Memorial Sloan Kettering Cancer Center, AI models were used to identify personalized neoantigens in pancreatic tumors, guiding the design of custom vaccines that extended recurrence-free survival in early clinical trials.

Meanwhile, investigators at Stanford University and Dana-Farber Cancer Institute demonstrated how AI frameworks are aiding in the identification of drug targets by integrating genetic, epigenetic, and transcriptomic data. These approaches are enabling researchers to discover previously unrecognized therapeutic vulnerabilities and enhance drug selection.

As the cancer research community continues to generate and integrate multimodal biological data, AI is becoming increasingly central to efforts aimed at unraveling cancer’s complexity. While further validation and clinical integration remain ahead, early evidence suggests that transformer-based models trained on mutation data may be key to delivering on the promise of precision oncology.