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

Abstract

Single-cell CRISPR perturbation screens provide a foundation for causal discovery in gene regulatory networks, but existing methods struggle with latent confounding, count-valued expression data, and high-multiplicity-of-infection (high-MOI) designs. We introduce RICE, a unified framework that addresses all three challenges. At its core is a reformulated control function tailored to the network-estimation setting, which restores robustness to exclusion-restriction violations that are pervasive in real CRISPR screens and that cause standard control function approaches to fail. RICE pairs this estimator with a constrained negative binomial model and a differentiable acyclicity penalty, accommodating hard and soft interventions and natively supporting high-MOI designs within a single, GPU-scalable model. Across extensive synthetic benchmarks, RICE consistently outperforms existing methods and remains stable under strong confounding, exclusion violations, and high-MOI conditions. Applied to CRISPRi screens, RICE achieves stronger causal-discovery performance on held-out data. RICE further reconstructs the canonical interferon (IFN)-{gamma} signaling pathway in melanoma cells upon immune stimulation, and nominates stable regulatory candidates that conventional significance tests may overlook due to limited statistical power. Together, these results establish RICE as a robust and scalable framework for causal discovery in single-cell perturbation genomics.