Active Causal Hypothesis Testing for AI-Guided Drug Target Discovery

The identification of causal mechanisms underlying disease pathology is the cornerstone of effective drug discovery. Traditional methods rely on slow, iterative experimental cycles, while modern computational approaches often prioritize correlation over causation. This paper introduces the Active Causal Hypothesis Testing (ACHT) framework, a novel AI-guided methodology designed to function as a "virtual experimenter" to automate and accelerate the discovery of therapeutic drug targets. ACHT integrates Graph Neural Networks (GNNs) for representation learning of biological networks, differentiable causal discovery for generating mechanistic hypotheses, and Bayesian active learning to prioritize the most informative hypotheses for validation. We apply ACHT to analyze 100,000 protein-protein interactions from the STRING v12.0 database. The framework autonomously identifies and prioritizes 3,529 high-confidence drug target candidates, including known drivers like TP53 and TNF. We validate these prioritized hypotheses using a novel Monte Carlo Wavelet Coherence analysis, demonstrating that the identified targets exhibit highly significant (p < 0.001) spectral signatures indicative of genuine therapeutic mechanisms, with a very large effect size (Cohen’s d=17.36). ACHT provides a scalable, interpretable, and automated pipeline for transforming large-scale biological data into actionable therapeutic hypotheses.