Multimodal-CanGen: A Multimodal Genome-Based Model for Cancer Type Classification

Cancer type classification is crucial for planning cancer treatment. Previous research has shown that genomic information is highly informative for classifying cancer types. However, the relationship between genomic features and cancer remains highly complex, and current genome-based models primarily use binary genome-level features or rely on comprehensive whole-genome sequencing (WGS) data, which limits their clinical feasibility. To address these complexities and improve clinical applicability, we introduce Multimodal-CanGen, which enhances cancer type classification using clinically accessible cancer ge-nomic sequencing data. Our approach includes a sequence-level encoder employing pre-trained models and graph neural networks (GNNs), a genome-level feature encoder, and a multimodal fusion module that integrates gated attention mechanisms and tensor fusion. The integration of sequence-level features provides insights typically unavailable from binary genome-level features alone, demonstrating superior performance and showing potential for predicting Cancer of Unknown Primary (CUP) cases. Overall, our work provides an advancement in accurately predicting cancer types in clinical practice and has the potential to pave the way for improved cancer treatment.