Mixture-of-Experts Guided Multi-Omic Integration for Gastrointestinal Cancer Subtype Prediction

Accurate cancer subtype classification is a cornerstone of precision oncology, informing therapeutic decisions and improving prognostic assessment. Gastrointestinal adenocarcinoma (GIAC), however, presents a particularly challenging case due to its molecular heterogeneity and overlapping histological features. Traditional approaches based on single-omic biomarkers or naive multi-omic concatenation often fail to capture the complex interdependencies across genomic, epigenomic, and transcriptomic layers. We introduce \textbf{MoXGATE} (Mixture-of-Experts Guided Multi-Omic Integration), a deep learning framework that leverages modality-specific expert encoders, cross-attention fusion, and learnable modality weights to enable robust and interpretable integration of gene expression, DNA methylation, and miRNA profiles. By combining expert specialization with attention-driven fusion, MoXGATE effectively captures cross-omic dependencies while adaptively weighting each modality according to its predictive relevance. To address severe class imbalance in GIAC subtyping, we further incorporate focal loss, enhancing sensitivity to underrepresented subtypes. Comprehensive evaluation on TCGA GIAC demonstrates that MoXGATE achieves superior accuracy compared to state-of-the-art baselines, while ablation studies confirm the contributions of expert routing, cross-attention, and modality weighting. Moreover, transfer experiments on the TCGA BRCA cohort highlight the model’s adaptability beyond GIAC, underscoring its generalizability to other cancer types.