Contrastive Learning in Single-cell Multiomics Clustering

Recent advancements in single-cell multiomics sequencing technology present new opportunities for researchers. However, the integrative analysis of the multiomics data poses new challenges, especially in cell clustering, a crucial step for any downstream analysis [5]. A key challenge is the alignment of multimodal omic features during fusion. A commonly adopted solution is adversarial training by implementing a discriminator of different omic features [1]. However, discriminators have several drawbacks affecting real-world performance [8]. In this study, we propose to use contrastive learning for better omic alignment by forcing different clusters of latent features to be separable and compact in the same space. We also aim to incorporate prior knowledge of interactions across genomics entities, specifically the gene regulatory network (GRN) for better clustering. Prior studies have shown GRN's important role in cell type classification [3, 4]. To our best knowledge, no end-to-end clustering method that incorporates GRN exists [1].