dotears: Scalable and consistent directed acyclic graph estimation using observational and interventional data

Abstract

New assays such as Perturb-seq link parallel CRISPR interventions to transcriptomic readouts, providing insight into gene regulatory networks. Causal regulatory networks can be represented by directed acyclic graphs (DAGs), but lack of identifiability and a combinatorial solution space complicate learning DAGs from observational data. Score-based methods have improved the practical scalability of inferring DAGs, but are sensitive to error variance structure. Furthermore, correction for error variance is difficult without prior knowledge of structure. We present dotears [doo-tairs], a continuous optimization framework leveraging observational and interventional data to infer causal structure, assuming a linear Structural Equation Model. dotears exploits structural consequences of hard interventions to estimate and correct for error variance structure. dotears is a provably consistent estimator of the true DAG under mild assumptions and outperforms other state-of-the-art methods in varied simulations. In real data, differential expression tests and high-confidence protein-protein interactions validate dotears-inferred edges with higher precision and recall than others.