Counterfactual Spatial Biology: A Causal Generative AI Framework for Explainable Cell–Cell Communication and Therapeutic Intervention Prediction

Mahule Roy, Subhas Roy

Keywords: Spatial biology, causal AI, spatial transcriptomics, multi-omics integration, counterfactual modeling, therapeutic prediction, drug discovery

TL;DR: SCARF predicts mechanistic drivers of disease and simulates therapeutic interventions in spatial omics data using biologically-informed causal AI.

Abstract: Spatial biology technologies have transformed our understanding of tissue organization, but current computational methods remain largely associative and lack causal interpretability for therapeutic decision-making. We introduce SCARF (Spatial Causal AI for Regulatory Forecasting), a framework that integrates causal inference with generative AI to enable counterfactual reasoning in spatial biology. SCARF learns a structured causal model of cell–cell communication from spatial omics data and simulates targeted interventions at single-cell resolution. Our approach addresses key limitations of existing methods by: (1) incorporating biological priors through a causal graph encoding ligand–receptor interactions and signaling pathways, (2) employing an intervention calculus respecting hierarchical cellular organization, and (3) generating biologically plausible counterfactual tissue states under therapeutic perturbations. Evaluation across multiple datasets (10x Visium breast cancer, IMC pancreatic cancer, MERFISH mouse brain) and LINCS L1000 drug perturbations demonstrates SCARF’s ability to predict mechanistic drivers of disease with in silico and experimental validation and to forecast intervention outcomes. SCARF enables unprecedented “what-if” analyses for drug discovery, representing a shift from pattern recognition to mechanistic reasoning in spatial biology.

Submission Number: 2