Auditing correlational gene regulatory networks through perturbation-grounded causal analysis for drug target prioritization

Gene regulatory network (GRN) inference for drug-target prioritization relies overwhelmingly on undirected co-expression, with highly connected “hub” genes routinely advanced as candidate regulators without independent causal verification. Although perturbation-based technologies now make direct causal validation feasible at genome scale, no study has directly tested whether correlational hub status reliably predicts causal regulatory importance or real-world druggability against an independent outcome rather than another network. This study evaluates that question and tests whether a simple, perturbation-free correction can recover any resulting predictive gap. Using the K562 essential-gene Perturb-seq dataset (2,285 pseudobulk CRISPR-interference profiles; 8,563 genes), an undirected correlational network and an independently constructed, perturbation-anchored causal network were built from identical data and compared across 1,868 shared genes via Spearman rank correlation and top-K overlap across three thresholds; both rankings were cross-referenced against Open Targets for documented drug mechanisms, and a logistic regression classifier quantified predictive value for known-drug status, with two specificity-adjustment refinements tested symmetrically on both networks. Rank agreement was weak and threshold-invariant (ρ = 0.091–0.158; zero top-20 overlap), yielding 77 correlational “false hubs” and 92 causal “hidden drivers.” Hidden drivers showed no category-level druggability enrichment (7.6% vs. 10.9% background) but included compelling, independently corroborated exceptions, CDK7 and AURKB, both invisible or near-invisible to correlation alone. Causal position outperformed correlational position for predicting known-drug status (AUC 0.655 vs. 0.625; combined 0.704), and specificity correction applied to the correlational network matched causal performance without perturbation data (AUC 0.655), reaching AUC 0.736 combined. These findings indicate correlational hub centrality is an unreliable proxy for causal importance and druggability, but a simple correction can recover much of the causal network’s advantage when tested empirically in both directions.