SMODA: Interpretable Multimodal Omics Integration for Disease Classification and Subtype Discovery via Heterogeneous Transfer Learning

Disease heterogeneity leads to various clinical molecular subtypes, further limiting the development of precision medicine. Multimodal data integration shows promise for addressing this challenge, but existing methods are affected by omics noise, data imbalance, and limited interpretability. Here, we propose SMODA (Semi-Supervised Multimodal Omics Data Analysis), which is a flexible framework to integrate multimodal omics data by combining heterogeneous transfer learning and semisupervised modeling. SMODA learns shared latent representations across different modalities to reduce the cross-modal heterogeneity. Systematic benchmarking demonstrates that SMODA outperforms existing multiomics integration methods both in disease classification and subtype identification. The application of SMODA in a multimodal esophageal cancer data set still shows better classification performance. A previously unrecognized disease subtype is also identified. This subtype shows altered lipid metabolism, inflammatory responses, and distinct exposure features, which are also linked to poor clinical outcomes. SMODA provides a reliable and interpretable framework for multimodal data integration and supports clinically relevant disease stratification. The SMODA framework is available at https://github.com/zhaoxiaoqi0714/SMODA.git.