Stochastic hypergraph co-contrastive learning for single-cell RNA-seq data imputation

Single-cell RNA sequencing (scRNA-seq) enables large-scale characterization of cellular heterogeneity but suffers from severe data sparsity caused by dropout events. Recent graph-based imputation methods have achieved progress by modeling cell–cell relationships, yet their fixed neighborhood structures limit the ability to capture high-order and uncertain interactions within such sparse data. In this paper, we propose Stochastic Hypergraph Co-contrastive Learning (SH-CoCL), a framework that aligns embeddings from two stochastic hypergraph views, enabling uncertainty-aware representation learning and accurate imputation for scRNA-seq data. Specifically, SH-CoCL introduces a probabilistic hyperedge generator that models cell–gene associations as learnable random variables, rather than relying on fixed neighborhood definitions constructed from similarity measures. By applying Gumbel–Sigmoid relaxation, it generates differentiable hypergraph structures that adapt to the data distribution and encode structural uncertainty. Then, it utilizes two probabilistic hypergraph views optimized with a contrastive objective, enabling the encoder to learn invariant and robust cell embeddings. In this process, the encoder employs hypergraph convolution to capture high-order interactions among cells, while a channel-wise self-attention mechanism enhances intra-cell feature dependencies. Finally, the learned representations are applied to neighborhood-based non-parametric imputation, reconstructing biologically consistent expression profiles. Comprehensive experiments on six real and five simulated datasets demonstrate that SH-CoCL consistently outperforms state-of-the-art baselines in recovering dropout expressions and enhancing downstream analyses, highlighting its potential for probabilistic representation learning in single-cell modeling.