Disentangling Over-Integration of Single-Cell and Spatial Multiomics Data with Transfer Learning

Integrating data from various single-cell sequencing technologies provides deeper insights than analyzing individual assays alone. Although several methods have been developed to integrate single-cell multiomics data, they often focus heavily on the degree of mixing between different modalities, resulting in over-integration that can obscure distinctions among cell subtypes and biological conditions. To address this limitation, we introduce an unsupervised method for integrating single-cell RNA sequencing (scRNA-seq) and Assay for Transposase-Accessible Chromatin (ATAC) data by bridging shared and specific patterns via transfer learning (namely scRabbit). scRabbit maintains the integrity of distinct data sources by using domain-specific batch normalization to reconstruct original expression profiles while preserving unique modality-specific features, thereby reducing over-integration. Additionally, scRabbit utilizes a modality classifier to capture shared features across modalities, enhancing integration and improving model training efficiency. Comprehensive benchmarking analyses demonstrate that scRabbit outperforms current state-of-the-art methods across complex scenarios, including paired and unpaired datasets, cell type imbalances, and biological conditions. Furthermore, we evaluate the effectiveness of scRabbit in integrating single-cell RNA data with spatial omics data, such as MERFISH and seqFISH.