Inferring gene regulatory networks from single-cell perturbation data by optimizing causal reach of latent mediator dynamics

CausalODE infers causal gene regulatory graphs from single-cell perturbation data by optimizing causal reach and latent protein dynamics, improving top-edge recovery by ~60% on CaRPool-seq and ~80% on Norman et al.

• Why it matters: Existing causal discovery methods for pooled CRISPR perturbations struggle with nonlinear, dynamic, feedback regulation and with RNA-only measurements that miss protein-mediated control. Evaluation based on expression prediction error may not reflect graph quality, motivating a perturbation-response–based metric.
• What they did: CausalODE is a differentiable framework that learns a graph by maximizing a causal reach score, modeling transient RNA responses and latent protein dynamics. It is benchmarked on simulated Perturb-seq datasets and evaluated on CaRPool-seq knockdowns and the Norman et al. activation dataset using a causal-reach metric.
• The result: On real data, CausalODE outperforms the strongest baseline by ~60% on CaRPool-seq and ~80% on Norman et al., with top-ranked edges enriched for evidence from cell-type-matched ChIP-seq, CollecTRI, and regulatory STRING. It also distinguishes non-additive from approximately additive interactions and tracks simulated protein dynamics with median temporal correlations of 0.79–0.99.

AI summary of the abstract · may contain errors · not medical advice · read the paper.