Hypergraph-Attention-Guided Graph Refinement and Feature Decoupling for Single-Cell RNA-seq Clustering.

Advances in single-cell RNA sequencing (scRNA-seq) enable the exploration of cellular heterogeneity at single-cell resolution. Accurate cell clustering is a prerequisite for delineating these identities; however, the inherent sparsity and prevalent dropout noise of scRNA-seq data pose major algorithmic challenges. Existing graph-based deep learning methods often struggle in this regime, suffering from the structural fragility of static neighbor graphs and the dimensional redundancy of latent embeddings. Thus, we introduce a novel method named Hypergraph-Attention-Guided Graph Refinement and Feature Decoupling for Single-Cell RNA-seq Clustering (scHAG), a deep graph clustering framework that synergizes generative pre-training, dynamic topological refinement, and orthogonal feature decoupling. To establish a robust semantic foundation, scHAG combines a Masked Autoencoder (MAE) with Zero-Inflated Negative Binomial (ZINB) modeling to mitigate dropout effects. A Hypergraph Attention-guided Structural Refinement (HAG-SR) module then captures high-order cellular relationships and reweights potentially spurious connections to obtain a more reliable graph topology. To prevent dimensional collapse, the framework incorporates a Barlow Twins objective that enforces cross-correlation orthogonalization, yielding highly discriminative embeddings. An adaptive gating module subsequently balances global topological priors with local neighborhood information. Experimental results on fourteen real datasets demonstrate that scHAG achieves highly competitive clustering accuracy and robust performance across diverse noise conditions. By circumventing the quadratic memory bottlenecks of standard graph networks, scHAG provides a memory-efficient and scalable framework for analysis of large-scale single-cell transcriptomic data.