scSNOT: Single-cell multi-omics clustering method based on minibatch unbalanced optimal transport and study-specific adaptive normalization

The rapid development of single-cell multi-omics technology has enabled the simultaneous capture of multiple aspects of cell information. Comprehensive analysis of these data helps deepen our understanding of the biological processes underlying complex diseases and plays a crucial role in precision medicine. To address challenges such as the curse of dimensionality, batch effects, and heterogeneity in single-cell multi-omics clustering, we propose scSNOT, an effective clustering method. ScSNOT utilizes coupled variational autoencoders to process highly variable common gene features from scRNA-seq and scATAC-seq datasets, mapping them into a unified low-dimensional latent space. Through a decoder with a study-specific adaptive normalization layer, scSNOT captures the shared information between the two data types while preserving their specific characteristics. To mitigate batch effects and adjust distributional differences, scSNOT introduces a minibatch imbalanced optimal transport loss function, ensuring effective clustering. Finally, scSNOT combines a clustering loss function with spliced low-dimensional embeddings for enhanced clustering performance. Experimental results on six real single-cell multi-omics datasets demonstrate that scSNOT outperforms existing methods, achieving significant improvements in clustering accuracy.