scHiCGNN: Per-Chromosome Graph Attention for Single-Cell Hi-C Clustering

Motivation Single-cell Hi-C measures three-dimensional chromatin organization in individual nuclei, revealing cell-to-cell variability in genome folding that bulk Hi-C cannot resolve. The central computational task is to cluster cells by chromatin state, but each cell yields a sparse, high-dimensional contact matrix poorly suited to linear principal component analysis pipelines or to hand-designed multi-stage neural pipelines. Results We present scHiCGNN, a graph attention framework that encodes each chromosome's contact matrix as a weighted graph, applies multi-head graph attention to learn node-level features, and fuses per-chromosome embeddings into a compact cell-level representation through a lightweight self-attention block. The resulting embedding is consumed by standard unsupervised clustering. We evaluate scHiCGNN on six public datasets: Flyamer, Collombet, Nagano (mouse) and Ramani, Lee, 4DN (human), spanning 171 to 19,388 cells. scHiCGNN consistently improves over principal component analysis baselines on all six datasets and matches or exceeds the published Higashi and MRscHiC numbers on the three larger human datasets, reaching Adjusted Rand Index values of 0.944 (Ramani), 0.674 (Lee), and 0.875 (4DN); on the smaller mouse datasets, MRscHiC's engineered pipeline retains an advantage. Direct graph representation learning is a practical and conceptually clean alternative to flattened-vector approaches for single-cell Hi-C clustering. Availability and implementation https://github.com/changlabtw/schicgnn