Counterfactual Diffusion Modeling Enables Spatially Targeted Reprogramming of Tissue Microenvironments
Wenhui Ding1, Zhenhua Luo1, Yuanyan Xiong2
1Institute of Precision Medicine, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou 510080, China.
Biology
|July 28, 2026
Summary
This study introduces SPAD-CFR, a novel computational framework for spatial single-cell analysis. It enables counterfactual simulations to understand how tissue structures and cellular behaviors change after specific interventions.
Area of Science:
- Computational Biology
- Systems Biology
- Bioinformatics
Background:
- Spatially resolved single-cell technologies offer insights into tissue heterogeneity but lack mechanistic understanding of cellular responses to perturbations.
- Existing computational models often neglect spatial information or tissue geometry, limiting the integration of molecular data with topological remodeling.
- Understanding tissue responses requires models that can simulate dynamic spatial and molecular changes.
Purpose of the Study:
- To develop a computational framework, SPAD-CFR (Spatial Point-cloud Attention-based Diffusion for CounterFactual Reprogramming), for spatial single-cell analysis.
- To enable the generation of counterfactual tissues by simulating simultaneous changes in cellular molecular profiles and spatial coordinates.
- To provide a tool for validating hypotheses regarding microenvironment reprogramming through biologically informed spatial simulations.
Main Methods:
- Treated tissues as spatial point clouds integrating cellular molecular profiles and physical coordinates.
- Implemented Pearl's three-step workflow for causal inference using deterministic diffusion inversion and sampling.
- Applied interventions to individual cells to generate counterfactual tissue states.
Main Results:
- SPAD-CFR successfully reproduced the hierarchical structure of the mouse cerebral cortex.
- The model simulated phenotypic distribution differences across various histological grades of breast cancer.
- It reconstructed hypoxia-associated mesenchymal phenotypes in triple-negative tumors and demonstrated distance-dependent cytotoxic effects in melanoma.
- Simulated activation interventions on PD-1+ CD8+ T cells in melanoma, showing spatially confined bystander effects.
Conclusions:
- SPAD-CFR is a biologically informed generative framework for spatial simulations.
- The model facilitates counterfactual analysis of microenvironment reprogramming and validation of biological hypotheses.
- It integrates molecular and spatial information for a more comprehensive understanding of tissue dynamics.
