Systematic evaluation of single-cell multimodal data integration enhances cell type resolution and discovery of clinically relevant states in complex tissues

Single-cell multimodal profiling is enabling increasingly comprehensive tissue reference atlases, yet the specific contribution of integrating distinct assays remains insufficiently characterized, particularly in complex organs. Here, we systematically evaluated multimodal single-cell data integration strategies in the human renal cortex using a benchmarking resource combining 3′ and 5′ scRNA-seq with paired snRNA-seq and snATAC-seq across matched and unmatched donor samples. To support consistent cross-modality comparison, we developed scOMM, an interpretable machine-learning framework for supervised cell-type mapping and benchmarking, and combined it with unsupervised integration approaches to assess horizontal, vertical, and global integration settings. We found that different modalities contribute complementary strengths to cell-type and cell-state identification. Horizontal integration of transcriptomic datasets improved cell-type definition, while vertical integration of snRNA-seq and snATAC-seq had an additive effect, enhancing resolution in homogeneous populations as well as in difficult-to-distinguish and transitional states. Global multimodal integration further refined adaptive and rare populations, enabling the identification of clinically relevant states, including WFDC2-expressing thick ascending limb cells and erythropoietin-producing Norn cells. Together, these results establish a systematic framework for multimodal reference atlas construction and provide practical guidance for improving cell-state resolution and disease-relevant discovery in complex tissues.