Robust causal gene network estimation for large-scale single-cell perturbation screens using reduced control function

Single-cell CRISPR perturbation screens offer a powerful framework for causal discovery in gene regulatory networks, but existing methods struggle with high-dimensional count data, unmeasured confounding, and the increasing prevalence of high-multiplicity-of-infection (MOI) designs. We introduce RICE, a scalable framework for causal gene network estimation that integrates a reduced control function to address latent confounding with a constrained generalized linear model accommodating both hard and soft interventions. By enforcing differentiable acyclicity constraints, RICE enables efficient GPU-based optimization for large-scale data. Across synthetic benchmarks, RICE achieves higher accuracy and robustness than existing methods and remains stable under strong confounding and high-MOI settings. Applied to multiple single-cell perturbation datasets, including CRISPRi screens in K562 and RPE1 cells and a Perturb-CITE-seq data set with CRISPR-Cas9 knockout (KO), RICE recovers biologically coherent networks with edge weights consistent with perturbation effects and enriched for known regulatory interactions. These results establish RICE as a flexible and scalable approach for causal discovery in modern single-cell perturbation studies.