Abstract 3744: Learning the language of somatic mutations: A large language model approach to precision oncology

Abstract The interpretation of somatic mutations in cancer remains challenging due to complex patterns of co-occurring variants and their clinical implications. Traditional approaches often analyze mutations in isolation, missing crucial contextual relationships that influence tumor behavior and treatment response. Large language models, having revolutionized natural language processing, offer a promising framework for understanding mutations in context. Just as words derive meaning from surrounding text, variants are better understood through their co-occurrence patterns, making this approach particularly powerful for precision oncology where each patient's unique combination of mutations tells a distinct disease story. We developed a transformer-based model that learns representations of somatic variants by considering both local sequence context and global co-occurrence patterns. Each variant is defined by its reference sequence, alteration, and flanking genomic context. The model architecture incorporates dual attention mechanisms: local attention between variants and their sequence context, and global attention between co-occurring variants within patients. Training utilized a masked prediction task where 20% of variant alterations were masked, and patient-level representations were generated by computing weighted averages of variant embeddings, with weights determined by variant allele frequencies (VAF) to capture tumor-specific mutational signatures. Sample-level representations proved highly effective for tumor type classification across both whole exome sequencing data from the TCGA (AUC=0.911, 33 cancer types) and targeted panel sequencing data from AACR Project GENIE's MSK-IMPACT cohort (AUC=0.880, 117 cancer types). These representations stratified patient progression-free survival in multiple TCGA cancer subtypes (p-values: colon=0.021, breast=0.033, pancreatic=0.002; multivariate log-rank), and early-stage colorectal (p = 0.037) and lung (p = 0.001) cancers in GENIE BPC cohorts, identifying distinct prognostic groups. Finally, analysis of learned attention weights through graph-based community detection revealed novel patterns of mutational dependencies, providing insights into how key mutations drive mutagenic processes. This work demonstrates that treating cancer mutations as a language enables powerful representations of personal cancer genomics, with implications spanning tumor classification, prognostic stratification, and biological discovery. Our approach, which leverages both local sequence context and global mutation co-occurrence patterns, generalizes from whole exome to panel sequencing data, facilitating integration into clinical workflows. The model's ability to identify prognostic subgroups and reveal mutational dependencies suggests its potential utility in advancing personalized treatment strategies. Citation Format: John-William Sidhom, Alexander S. Baras. Learning the language of somatic mutations: A large language model approach to precision oncology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3744.