Abstract

Cell-cell communication (CCC) analyses of spatial transcriptomics commonly prioritize ligand--receptor expression or spatial proximity, whereas the effect of ligand exposure on a receiver-cell response remains confounded by tissue microenvironment. We introduce CASPER, a Bayesian structural equation model for the effect of kernel-weighted ligand exposure on pathway activity, conditional on receptor expression. A shared Gaussian Markov random field represents unmeasured spatial confounding. We establish identification under a known spatial covariance, homoskedastic independent errors, and constant linear coefficients; heterogeneous marginal latent variances provide one sufficient, but not necessary, source of identification. A spike-and-slab prior yields posterior inclusion probabilities, supplemented by model-averaged effects, credible intervals, and practical-effect probabilities. Simulations compare CASPER with ordinary, spatial outcome-only, and oracle regressions and assess residual diagnostics and sensitivity to the fitted spatial covariance. Four public Xenium datasets support comparison with native CellPhoneDB on a fixed candidate set. An uncapped renal analysis estimates a negative myeloid CD86-exposure effect on T-cell inflammatory-response activity, conditional on CTLA4 expression and the shared-field model. Its interpretation requires care because of strong posterior dependence between direct and latent effects. CASPER distinguishes a model-based pathway-response contrast from interaction-expression enrichment without establishing receptor-mediated signaling experimentally.