Latent causal diffusions for single-cell perturbation modeling.

Latent causal diffusion (LCD) predicts transcriptome responses to unseen perturbations and uncovers gene regulatory causal structures in single-cell RNA-sequencing data.

• Why it matters: Understanding cellular responses to perturbations is crucial for mapping regulatory mechanisms, but current models struggle with prediction accuracy and causal interpretation, limiting biological insights.
• What they did: The authors developed LCD, a generative model treating gene expression as a stationary diffusion process with measurement noise, and CLIPR, a method to infer causal gene effects, tested on simulated data and genome-wide screens.
• The result: LCD outperforms existing approaches in predicting perturbation effects, while CLIPR identifies causal gene relationships and functional modules, enabling mechanistic insights beyond standard differential expression analysis.