The Dose Makes the Poison: Exposure-Aware Evaluation of Generative Drug Design

Generative models are increasingly used to propose molecules for drug discovery, producing far more candidates than can be experimentally characterized. Computational screening and selection are therefore essential, and commonly rely on static predictions of potency, toxicity, and drug-like properties. Drugs, however, are administered at doses, and both efficacy and toxicity depend on the resulting exposure. We ask whether candidates that look promising under static molecular scores remain promising once pharmacokinetics and dose are taken into account. We examine three PARP1-targeted generative pipelines and re-evaluate shortlisted molecules using physiologically based pharmacokinetic simulations. Seven of 60 shortlisted molecules cannot achieve the required target exposure under our dosing constraints despite acceptable individual property predictions. Among the remaining candidates, 28\% are classified as risky based on their predicted exposure, and 85\% are outperformed on dose requirement and both safety margins by at least one candidate that the static ranking left behind. Our results show that molecular property scores alone can mis-rank generated candidates, and motivate bringing dose and exposure earlier into computational drug design.