COSIME: Cooperative multi-view integration and Scalable and Interpretable Model Explainer

Single-omics approaches often provide a limited view of complex biological systems, whereas multiomics integration offers a more comprehensive understanding by combining diverse data views. However, integrating heterogeneous data types and interpreting the intricate relationships between biological features-both within and across different data views-remains a bottleneck. To address these challenges, we introduce COSIME (Cooperative Multi-view Integration and Scalable Interpretable Model Explainer). COSIME uses backpropagation of Learnable Optimal Transport (LOT) to deep neural networks, enabling the learning of latent features from multiple views to predict disease phenotypes. In addition, COSIME incorporates Monte Carlo sampling to efficiently estimate Shapley values and Shapley-Taylor indices, enabling the assessment of both feature importance and their pairwise interactions-synergistically or antagonistically-in predicting disease phenotypes. We applied COSIME to both simulated data and real-world datasets, including single-cell transcriptomics, single-cell spatial transcriptomics, epigenomics, and metabolomics, specifically for Alzheimer's disease-related phenotypes. Our results demonstrate that COSIME significantly improves prediction performance while offering enhanced interpretability of feature relationships. For example, we identified that synergistic interactions between microglia and astrocyte genes associated with AD are more likely to be active at the edges of the middle temporal gyrus as indicated by spatial locations. Finally, COSIME is open-source and available for general use.