Abstract A023: SomAtt: A foundation model for the tumor genome

Abstract Somatic mutations in cancer cells can lead to diverse phenotypic consequences due to variations in mutation types and locations within the same gene. Interactions between mutations across different genes further contribute to tumor heterogeneity, affecting pathology, phenotype, and clinical outcomes. Current computational methods often treat all mutations within a gene as functionally equivalent and overlook combinatorial effects between genes. Additionally, the diversity of sequencing platforms and gene panels poses significant challenges for integrative genomic analysis across datasets. Here, we introduce SomAtt (Somatic Attention), a foundation model for tumor genomes that leverages self-attention mechanisms to model interactions among mutations within a tumor. Trained on 166,854 tumors from 149,831 unique cancer patients, SomAtt generates comprehensive tumor embeddings suitable for various downstream tasks. The model addresses three critical challenges in tumor genome analysis: (a) distinguishing the effects of different mutations within the same gene, (b) capturing complex interactions among multiple mutations within a tumor, and (c) integrating data from diverse targeted mutation panels without requiring uniformity. We demonstrate that SomAtt accurately classifies cancer types with an accuracy of 76.5% across 34 cancer types, outperforming the previous state of the art model by 2% on MSK-IMPACT panels. SomAtt also predicts patient survival outcomes, achieving a concordance index (C-index) of 0.76 across 89 gene panels and 108 cancer types. Notably, SomAtt also enables patient stratification beyond traditional subtyping methods by identifying previously unrecognized, molecularly defined patient subgroups with significantly different outcomes, suggesting a novel approach to cancer subtype classification based on tumor molecular profiles. Our model could enable better assessment of pre-cancerous lesions molecularly and aid in early detection. SomAtt produces a robust and scalable framework for comprehensive modeling of tumor genomes, addressing key challenges in cancer genomics and advancing the potential for precision oncology. Citation Format: David Arredondo, Luis Tafoya, Oladimeji Macaulay, Ala Jararweh, Avinash Sahu. SomAtt: A foundation model for the tumor genome [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A023 .