Learning Stable Causal Structures from Perturbed Genomic Data: Robust GRN Inference Under Adversarial Interventions

Causal discovery from observational data is fundamentally challenged by distribution shifts, which are ubiquitous in biological systems. In gene regulatory networks (GRNs), such shifts often arise from adversarial interventions—either naturally occurring (e.g., pathogenic perturbations, cellular stress) or experimentally engineered (e.g., CRISPR knockout screens). While standard causal discovery methods assume identical data distributions across environments, adversarial interventions deliberately alter network topology or regulatory dynamics, violating this assumption and leading to spurious causal inferences. This paper introduces a novel framework for robust causal discovery in GRNs under adversarial distribution shifts. We model adversarial interventions as unknown, targeted perturbations that can modify both marginal and conditional dependencies among genes. Leveraging a combination of invariant causal prediction and moment-matching adversarial training, our method learns causal structures that remain stable across heterogeneous environments without prior knowledge of intervention targets. Through extensive simulations of synthetic GRNs and real-world single-cell RNA-seq data with latent adversarial perturbations, we demonstrate that our approach significantly outperforms existing causal discovery algorithms (e.g., PC, LiNGAM, DAG-GNN) in terms of structural accuracy and edge orientation reliability. Notably, we recover known regulatory interactions in the E. coli SOS response network under chemically-induced stress—a setting where standard methods fail due to distribution shift. Our results establish a principled pathway for causal discovery in non-stationary biological systems and highlight the necessity of modeling adversarial interventions for accurate GRN inference.