Spatial Instrumental Variables for Causal Gene Regulatory Network Discovery from Spatial Transcriptomics

Inferring causal gene regulatory networks (GRNs) from observational transcriptomic data remains fundamentally limited: without interventions, one can at best identify Markov equivalence classes, leaving edge orientations ambiguous. Interventional approaches such as Perturb-seq resolve directionality but destroy spatial tissue context. We introduce SpaCI (Spatial Causal Instruments), a framework that exploits naturally occurring morphogen gradients in spatial transcriptomics data as instrumental variables for causal GRN discovery. Our key insight is that spatially structured signaling molecules, such as Wnt, BMP, and Hedgehog ligands, induce continuous variation in downstream regulatory activity through known receptor-mediated pathways, satisfying the exclusion restriction required for instrumental variable identification. We formalize this as a nonparametric spatial IV framework, prove identifiability of causal edge directions under stated assumptions, and develop a scalable three-stage algorithm that combines spatial kernel regression with constraint-based DAG learning. Critically, SpaCI identifies causal directions, not just edges, enabling full directed acyclic graph (DAG) recovery. On synthetic spatial GRN benchmarks with known ground truth, SpaCI recovers causal edge orientations with significantly higher accuracy than existing methods (AUROC 0.87 vs. 0.71 for the best baseline). On Drosophila embryo spatial transcriptomics data, SpaCI recovers known anterior-posterior patterning regulatory relationships and identifies novel spatially mediated regulatory interactions supported by independent chromatin accessibility (ATAC-seq) data. Our results establish spatial tissue architecture as a previously unexploited source of causal identification for gene regulatory network inference.