CauSTI: causal spatial transcriptomic interaction inference for cell-cell communication

Abstract

Introduction Cell-cell communication plays a critical role in tissue organization, development, immune regulation, and disease progression. Spatial transcriptomics enables the inference of such communication within the native tissue context. However, existing methods often rely on ligand–receptor co-expression or spatial proximity alone, without explicitly modeling the directional nature of intercellular signaling. Methods We propose CauSTI, a causal spatial transcriptomic interaction inference framework for identifying direction-specific cell-cell communication. CauSTI integrates spatial neighborhood structure with ligand–receptor-derived causal priors through a causal-prior-guided variational graph autoencoder. Spatial proximity defines the tissue graph, while ligand–receptor evidence provides biologically informed directional constraints. An asymmetric decoder models sender–receiver relationships between ordered cell pairs. The optimization objective jointly incorporates spatial reconstruction, causal-prior consistency, variational regularization, and directional regularization. Results Experiments on three spatial transcriptomic datasets—seqFISH+, MERSCOPE, and Xenium—demonstrate that CauSTI consistently outperforms representative baseline methods in AUROC, ACC, and AP. Ablation studies further confirm the contributions of causal priors and directional regularization. Discussion These findings suggest that integrating spatial context with causal priors and explicit directional modeling provides an effective framework for inferring biologically plausible, direction-specific cell-cell communication landscapes from spatial transcriptomic data.