GRAIL-Heart

Existing tools for inferring cell-cell communication from spatial transcriptomics identify ligand-receptor (L-R) co-expression patterns but cannot determine which interactions causally drove cell fate decisions. Here, we present an inverse modelling framework that reverses the traditional inference direction; a graph neural network predicts cell fates from L-R interaction signals and assigns per-interaction causal scores via a bilinear score head, while a novel expression-gated 90th-percentile aggregation strategy resolves scores collapse inherent in naïve aggregation. Applied to the Heart Cell atlas v2 (42,654 cells, six cardiac regions), the framework identifies YAP/TAZ (p = 6.7 × 10-20), BMP (p = 2.2 × 10-23), and TGF-β (p = 3.3 × 10-11) as dominant mechanosensitive causal axes in the left ventricle, with TIMP1 → MMP2 emerging as a universal extracellular matrix remodelling hub and complement cascade signalling as an underappreciated causal axis in atrial tissue. External validation on eight independent Visium datasets spanning two species and three disease conditions confirms robust generalisation, including disease-specific causal rewiring in myocardial infarction and Kawasaki vasculitis detected without disease supervision. These results demonstrate that inverse modelling distinguishes causal from correlative cell-cell communication, providing a foundation for identifying actionable signalling targets in cardiac biology. Users can upload their datasets and interact with the GRAIL-Heart portal via https://huggingface.co/spaces/Tumo505/grail-heart, a network explorer with preloaded Heart Cell Atlass v2 dataset has been availed via https://tumo505.github.io/GRAIL-Heart/outputs/cytoscape/index.html.