DoGMA Model Improves Cancer Multi-Omics Analysis

This paper introduces DoGMA, a novel foundation model designed for comprehensive multi-omics analysis across various cancer types. Unlike many existing multi-omics models that use unrestricted attention mechanisms, DoGMA integrates the fundamental biological principle of the central dogma, which dictates the directional flow of genetic information (DNA to RNA to protein). This domain-specific inductive bias is embedded into the model's Transformer-MoE architecture through directed attention, ensuring inter-omics communication aligns with biological reality. DoGMA is further pre-trained using masked hierarchical omics reconstruction, which reinforces the learning of central-dogma-consistent interactions. The model demonstrates strong predictive performance across a range of downstream tasks, including cancer representation learning, survival prediction, and metastasis prediction. Ablation studies confirm that the improved performance stems from the synergy between the central-dogma-guided directed attention and the reconstruction-based pre-training, leading to more biologically coherent cross-omics information exchange and enhanced robustness and transferability.