scCMIA: Mutual Information-Guided Decoupled Learning for Robust Single-Cell Cross-Modal Integration

Recent advances in single-cell multimodal omics sequencing enable the joint profiling of multiple molecular layers within individual cells. Despite this progress, computational integration remains challenging because cross-modal alignment must be achieved without discarding modality-specific information. This paper introduces scCMIA, a mutual-information-guided framework for robust single-cell cross-modal integration. scCMIA decomposes the representation of each modality into a semantic latent variable for shared cellular states and a modality-specific latent variable for non-shared information required for reconstruction. The framework combines contrastive cross-modal alignment, mutual-information-guided decoupling, and a unified CrossVQ codebook to support both accurate reconstruction and interpretable discrete representation learning. Benchmarking across paired single-cell multi-omics datasets demonstrates that scCMIA achieves strong alignment and reconstruction performance, improves downstream label transfer and cell-type classification, and enables code-level analysis of cross-modal coupling patterns across cell types. These results show that scCMIA provides an effective and interpretable framework for single-cell cross-modal integration.