Causal Descriptors in QSAR: Deconfounding High-Dimensional Molecular Features via Double Machine Learning and Hypothesis Testing

Quantitative Structure-Activity Relationship (QSAR) modeling is a pillar of computational drug discovery. However, standard machine learning (ML) models are often confounded by the high-dimensional and intensely correlated nature of molecular descriptors. A model may identify a "bulk" property (e.g., molecular weight) as highly predictive, when in fact it is merely a proxy for a true, specific pharmacophore (e.g., a hydrogen bond donor). This correlational insight can misdirect costly synthesis efforts. We propose a statistical framework to move from correlational QSAR to causal QSAR. Our approach uses Double/Debiased Machine Learning (DML) to estimate the unconfounded causal effect of each molecular descriptor on biological activity, treating all other p-1 descriptors as potential confounders. We then apply the Benjamini-Hochberg procedure to these p estimates to perform high-dimensional hypothesis testing and control the False Discovery Rate (FDR). We validate this framework using a simulation study that explicitly models the high-correlation and confounding structures endemic to chemoinformatics. We show that baseline models (Lasso, Random Forest) are easily misled, consistently ranking non-causal but confounded "bulk" descriptors as highly important. In contrast, our DML + FDR framework successfully "sees through" the confounding, correctly identifies the true causal descriptors, and rejects the spurious ones, while maintaining the target FDR. This causal inference framework provides a robust method for "deconfounding" the molecular descriptor space. By identifying features with a statistically significant causal link to activity, it can provide medicinal chemists with more reliable, interpretable, and actionable hypotheses for rational drug design.