Identifying Short Half-Life Compounds from Structure Information Using Multi-Task Graph Neural Networks

Abstract Accurate prediction of terminal half-life (t1/2) is paramount for prioritizing drug candidates during the lead optimization stage, as it minimizes the risk of clinical trial failure while optimizing resource allocation. Drugs characterized by suboptimal t1/2 profileseither excessively short or prolongedpresent significant challenges in dosage design and are frequently associated with increased risks of adverse effects. Traditionally, the determination of t1/2 has necessitated in vivo experiments to derive total clearance (CLT) and volume of distribution (Vss). In this study, we curated a comprehensive dataset of over 1400 compounds and developed deep learning-based regression models to predict t1/2 directly from molecular structures using Graph Convolutional Networks (GCNs). We implemented both single-task and multi-task learning architectures; the latter integrated CLT, Vss, and other related pharmacokinetic parameters to leverage their inherent physiological correlations. Our results indicate that a multi-task model achieved higher predictive performance than the single-task counterpart, with notably high precision in identifying compounds with short t1/2. By enabling the structural-based identification of unfavorable t1/2 profiles without the need for experimental data, this approach provides a robust and efficient tool for streamlining candidate selection in the early stages of drug development. This methodology is expected to accelerate the discovery of compounds with favorable pharmacokinetic properties, reducing the reliance on late-stage experimental validation.