Proximity-Aware Graph Attention Networks for Spatially Resolved Cell-Cell Inference

Cell-cell inference is a fundamental mechanism driving tissue homeostasis, immune regulation, and disease progression. Existing CCC inference tools operate on dissociated single-cell RNA-seq data, discarding the spatial constraints that govern short-range paracrine and juxtacrine signaling. We present S pa GNN, a proximity-aware graph attention network that jointly encodes physical cell adjacency and ligand-receptor co-expression from spatial transcriptomics data to infer spatially resolved CCC interactions. A heterogeneous cell communication graph is constructed by combining spatial proximity edges with molecular ligand-receptor edges, and a distance-modulated attention mechanism enforces the locality constraint of cell-to-cell signaling. The model is trained with a composite self-supervised objective that integrates spatial co-localization enrichment and downstream target gene activation. We evaluate S pa GNN on four spatial transcriptomics datasets spanning 10X Visium, MERFISH, and Slide-seq platforms. Compared with state-of-the-art methods including COMMOT, SpaOTsc, CellChat, and NicheNet, our method achieves superior interaction AUROC (0.871 vs. 0.831 for COMMOT), a 3.42× spatial enrichment score, and an F1 of 0.681 in communication hotspot detection. Ablation studies confirm that spatial distance modulation is the single most impactful component, contributing a 0.030 AUROC improvement over a non-spatial baseline.