scHashFormer: a hash-driven graph transformer for scalable scRNA-seq clustering

Single-cell RNA sequencing (scRNA-seq) has emerged as a transformative technology for decoding cellular heterogeneity and state diversity within complex tissues through high-throughput transcriptomic profiling of individual cells. scRNA-seq clustering is a critical task for analyzing scRNA-seq data, which resolves high-dimensional expression profiles into interpretable cellular types and states. Although the Transformer, as a powerful foundation model, offers strong representation learning capabilities, its use in single-cell analysis is limited by the absence of a biologically meaningful and computationally scalable tokenization mechanism. Existing methods typically construct tokens through similarity-based neighbor selection, a process that is highly sensitive to metric quality and incurs substantial computational overhead, limiting applicability to large-scale datasets. Here, we introduce a hash-driven tokenization mechanism, scHashFormer, in which we design a novel hash encoder with a learnable hash window size and train it using self-supervised learning to realize similar cells with the same hash codes. Identical hash codes define hash buckets from which token sequences are constructed. By aggregating information from the constructed sequence, similar cells are brought closer in the embedding space, ensuring more effective clustering. Extensive experiments on multiple scRNA-seq datasets demonstrate that scHashFormer achieves competitive clustering effectiveness and scalability. The resulting embeddings enhance performance in downstream tasks, including trajectory preservation and gene differential expression analysis.