A meta-interaction basis for cell-cell communication in tissues

Tissue function depends on signals exchanged between cells and the responses they elicit, but it has remained unclear whether diverse cell-cell interactions in situ form recurrent sender-receiver programs. SpiderNet is presented as an interpretable representation-learning framework that discovers such directed programs as a compact basis of cell-cell meta-interactions (MIs) from spatial transcriptomics.

The method jointly learns which sender regulators, ligand-receptor pairs, and receiver targets define each MI, and where each program is active across neighboring cell pairs. The resulting representation traces multicellular relays and links communication to cell states, perturbation responses, and phenotypes.

SpiderNet recovers ground-truth MIs and their molecular components in simulations and, in real tissues, shows stronger direction-specific agreement with independently curated regulatory programs in senders and receivers than alternative methods. Across more than 5.8 million spatially profiled cells, it resolves an SPP1-THBS relay linking monocytes, fibroblasts, and tumor cells within an immune-suppressive ovarian cancer niche, predicts T-cell responses to held-out melanoma-cell perturbations, and identifies a T-cell-associated brain-aging program and age-predictive signals that transfer across regions and platforms.

It further reveals a recurrent pan-cancer COLLAGEN-linked fibroblast-tumor program whose projected abundance in independent cohorts is associated with poorer survival and non-response to immunotherapy. SpiderNet thus establishes MIs as a reusable organizational layer between molecular interactions and tissue phenotypes, providing a framework to resolve, compare, trace, and perturb multicellular regulation in situ.