A Counterfactual Framework for Directional Cell–Cell Interaction Analysis in Spatial Transcriptomics

Humaira Anzum, Veena Kochat, Suresh Satpati, Md Ishtyaq Mahmud, Jagan Mohan Reddy Dwarampudi, Pooja Shukla, Milind Javle, Lawrence Kwong, Kunal Rai, Tania Banerjee

Abstract: p>Understanding how neighboring cells influence cellular states in spatial transcriptomics is central to characterizing tissue microenvironments, yet most existing approaches rely on correlation or predefined ligand receptor pairs, limiting directionality and robustness. We propose a counterfactual, intervention-based framework for inferring directional cell cell influence from spatial transcriptomics data that is ligand receptor agnostic and explicitly tests sender specificity. Our method trains a neighborhood conditioned predictive model to estimate receiver cell state from spatial neighbors, with architectural constraints that prevent trivial self reconstruction. Directional influence is quantified by selectively replacing neighbors of a candidate sender cell type with spatially matched non-sender cells and measuring the resulting displacement in predicted receiver state. We define a Counterfactual Displacement Score (CDS) at the single-cell level and aggregate it to obtain a Counterfactual Transfer Score (CTS) for each ordered sender receiver cell type pair. We evaluate the framework on Xenium spatial transcriptomics tissue microarray from human cholangiocarcinoma, revealing reproducible, directionally asymmetric interactions between tumor, immune, and stromal compartments. Rigorous falsification tests, including label permutation, distance preserving neighborhood shuffling, sender agnostic replacement, and spatial coordinate permutation, demonstrate that observed CTS values exceed a majority of the null distributions. Stability is further confirmed via block bootstrap within tissue sample, indicating robustness to biological heterogeneity. Overall, our approach provides a statistically grounded and scalable framework for inferring directional cell cell influence in spatial transcriptomics, suitable for large imaging-based datasets and clinical tissue studies.</p>

External IDs:doi:10.64898/2026.04.05.716510