scComm: a contrastive learning framework for deciphering cell–cell communications at single-cell resolution

Cell-cell communication regulates complex biological processes in multicellular systems. Existing scRNA-seq-based methods typically aggregate gene expression by clusters, overlooking within-cluster heterogeneity. We present scComm, a computational framework that infers cell-cell communications between individual cells using supervised contrastive learning. In simulations, scComm outperforms other methods and achieves up to 95% accuracy. Applied to colorectal cancer, it reveals cell-cell communications linked to PD-1 blockade response and tertiary lymphoid structures. In liver cancer, it identifies three novel tumor subtypes and angiogenesis-promoting neutrophil subtypes that have unique tumor microenvironments. scComm enables high-resolution cell-cell communication analysis, uncovering biological insights missed by existing approaches.