Celcomen: spatial causal disentanglement for single-cell and tissue perturbation modeling

Area of Science:

• Computational biology
• Systems biology
• Genomics

Background:

• Spatial transcriptomics enables cell-level gene expression analysis within tissue context.
• Understanding gene regulatory networks is crucial for deciphering cellular functions and disease mechanisms.
• Current methods struggle to disentangle complex intra- and inter-cellular regulatory interactions.

Purpose of the Study:

• To develop a computational framework for dissecting gene regulatory programs in spatial transcriptomics data.
• To create a generative graph neural network model for predicting post-perturbation effects in tissues.
• To enable the creation of virtual tissues for studying experimentally inaccessible biological states.

Main Methods:

• Leveraged a mathematical causality framework.
• Employed a generative graph neural network architecture.
• Validated through simulations and analysis of human glioblastoma, human fetal spleen, and mouse lung cancer samples.

Main Results:

• Successfully disentangled intra- and inter-cellular gene regulation programs.
• Demonstrated identifiability of causal structures within the data.
• Generated accurate counterfactual spatial transcriptomics predictions.
• Validated model performance across diverse biological samples.

Conclusions:

• Celcomen provides a novel approach to model gene regulation in spatial transcriptomics.
• Enables the prediction of virtual tissue responses to perturbations.
• Offers new avenues for understanding disease and therapy effects at a single-cell, spatially resolved level.