Latent Causal Diffusions for Single-Cell Perturbation Modeling.

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Perturbation screens offer insights into cellular regulation at single-cell resolution.
• Predicting transcriptome-wide responses and inferring causal gene interactions remain significant computational hurdles.
• Current methods struggle with noise, lack causal inference, and underperform baselines.

Purpose of the Study:

• To develop a novel generative model for predicting gene expression dynamics under perturbations.
• To create a method for inferring causal gene regulatory structures from perturbation data.
• To improve the accuracy and interpretability of single-cell RNA sequencing (scRNA-seq) perturbation analyses.

Main Methods:

• Introduced latent causal diffusion (LCD), a generative model treating gene expression as a diffusion process with measurement noise.
• Developed causal linearization via perturbation responses (CLIPR) to approximate direct causal effects from LCD dynamics.
• Validated LCD-CLIPR on simulated data and a genome-wide scRNA-seq perturbation screen.

Main Results:

• LCD accurately predicts distributional shifts in unseen perturbation combinations, outperforming existing methods.
• CLIPR successfully recovers causal gene regulatory structures in simulations and experimental data.
• The framework identifies functional gene modules and resolves causal relationships missed by standard differential expression analysis.

Conclusions:

• The LCD-CLIPR framework effectively integrates generative modeling and causal inference for perturbation prediction.
• This approach provides a powerful tool for mapping complex gene regulatory mechanisms at the transcriptome level.
• It advances the ability to understand and predict cellular responses to genetic or chemical perturbations.