Interpretable AI model reveals spatial codes of cell recruitment in tissues.

Abstract

Understanding which molecular signals recruit specific cells to precise tissue locations is fundamental to tissue biology, immunity and disease. Existing computational tools for spatially resolved transcriptomics (SRT) describe where cells reside but cannot explain why. Here we introduce SPACER, a fully interpretable multi-instance learning framework that decodes gene-level recruitment rules directly from SRT data. Applied to 37 SRT datasets spanning multiple human tumour types and mouse heart, SPACER achieved 4-21-fold improvement over existing methods in various tasks of recovering biologically meaningful recruitment signals. SPACER revealed (1) that tumour-expressed mucins are associated with physical barriers to T cell engagement, which predicts resistance to immune checkpoint inhibitors and patient prognosis, (2) that local antigen presentation is the dominant correlate of T cell engagement across cancer types and (3) that Cd4⁺ T cells are the dominant functional effectors in severe immune checkpoint inhibitor-induced myocarditis. These findings establish SPACER as a discovery engine for cellular recruitment biology with direct translational applications.