GRAIL-heart: A graph attention network for inferring ligand-receptor interactions in spatial transcriptomics

Abstract

Cell-cell communication through ligand-receptor (L-R) interactions orchestrates cardiac development, homeostasis, and disease progression, yet existing computational methods cannot infer directional signalling networks or distinguish causal from correlational interactions in spatial transcriptomics data. We present GRAIL-Heart, a graph-attention-based framework that integrates spatial tissue topology with multi-task learning to predict context-dependent L-R interactions, reconstruct gene expression, and infer causal signalling pathways. Validated on 42,654 cells across six cardiac regions from the Human Heart Cell Atlas, GRAIL-Heart achieves 94.3% AUROC for L-R prediction, successfully recovers known cardiac signalling pathways, and reveals region-specific complement system involvement in cardiac homeostasis. The method outperforms existing approaches by 56%, is uniquely capable of gene expression reconstruction (R 2 = 0.996), and provides an interpretable, generalisable framework for prioritising high-confidence ligand-receptor hypotheses for downstream experimental validation across diverse tissues. Key innovations:•Spatial graph integration: Dual-edge architecture encoding both spatial proximity and ligand-receptor-specific relationships for context-aware interaction prediction.•Multi-task learning with causal inference: Simultaneous optimisation of L-R prediction, expression reconstruction, and inverse modelling to distinguish functionally important from correlational interactions.•Open-source implementation: Fully reproducible code, pretrained cardiac model, and interactive web explorer enabling broad adoption across spatial transcriptomics applications.

Abstract of "GRAIL-heart: A graph attention network for inferring ligand-receptor interactions in spatial transcriptomics" by Kgabeng T, Wang L, Ngwangwa H et al., https://doi.org/10.1016/j.mex.2026.104093, used under its Creative Commons Attribution license (CC-BY); markup removed, otherwise unmodified.