scHCL-SDA: A Homophily-Aware Contrastive Learning with Spectral Distribution Alignment for scRNA-seq Data Clustering.

Graph contrastive learning has recently shown strong potential for single-cell RNA sequencing (scRNA-seq) data clustering via modeling pairwise cell relationships. Nevertheless, most of them rely on the homophily assumption, which posits that adjacent cells in the graph share similar biological identities. This assumption, however, frequently fails in practical scRNA-seq datasets owing to complex heterogeneous cellular architectures. Such heterogeneity can lead to erroneous connections in low-homophily neighborhoods, thereby resulting in suboptimal clustering performance. To address this issue, we propose a homophily-aware contrastive learning with spectral distribution alignment (scHCL-SDA) for clustering scRNA-seq data. Specifically, the proposed scHCL-SDA method constructs an expression view from the original gene expression matrix and generates a graph-filteblue augmentation view through graph-guided multi-round noise filtering. Then, a spectral distribution alignment mechanism is introduced to address the cross-view distribution shift problem and thus enhance representation consistency of multi-view scRNA-seq data. To blueuce the interference caused by false negative samples during training, it employs Debiased InfoNCE contrastive learning and then designs a progressive fusion strategy to achieve smooth multi-view feature fusion and deep semantic interactions. Furthermore, we introduce an adaptive homophily-aware consistency contrastive learning mechanism to dynamically construct reliable neighborhoods, thereby effectively blueucing structural noise in low-homophily graphs. Comparative experiments on real datasets show the effectiveness of the proposed scHCL-SDA method in scRNA-seq data clustering. The source code for this work is available at: https://github.com/szq0816/scHCL-SDA.