Orthogonal disentanglement of single-cell multi-omics reveals private and shared drivers of tissue development and pathogenesis.

Characterizing gene expression and regulatory dynamics underlying both normal tissue function and disease progression requires an integrative analysis of single-cell multi-omics data. However, the asynchrony of gene regulation and the snapshot of single-cell multi-omics data give rise to private signals unique to each omics layer and shared signals reflecting cross-modality coordination. Here, we present Omics Separation Modeling using Domain Adaptation (OmiDos), a flexible annotation-free deep learning framework that disentangles omic-specific and interomic shared latent variables in multi-omics data with private-shared component analysis. Its modular architecture enables seamless extension to incorporate adversarial learning for unpaired data misalignment and to restructure its components to leverage the maximum mean discrepancy regularization, thereby minimizing interference with biological variability. Through this disentanglement, OmiDos enables the estimation of gene expression and regulatory dynamics at finer biological granularity and empowers various downstream analyses. We demonstrated the superior performance of OmiDos in terms of clustering accuracy, batch-effect correction, and misalignment resolution across datasets spanning diverse platforms and tissue types. In mouse secondary palate development, OmiDos precisely identified a cell type-specific unlinked distal enhancer, elucidating its essential role in the regulation of epithelial cell differentiation and migration. The application of OmiDos to medulloblastoma revealed a potential role deficiency in driving partial closure of the distal enhancer region of