Application of Deep Learning to Predict Human Oral Bioavailability of Pharmaceuticals

High failure rates in drug development are predominantly driven by suboptimal ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties, with human oral bioavailability (HOB) serving as a critical determinant of therapeutic efficacy and safety. Traditional HOB assessment methods, reliant on animal models and clinical trials, face inherent limitations in cost, scalability, and reproducibility. To address these challenges, this study proposes a deep learning framework integrating the directed message-passing neural network (D-MPNN) from the Chemprop tool with RDKit-derived molecular descriptors, enhancing predictive accuracy through hybrid representations of atomic/bond-level graph features and global physicochemical properties. Bayesian optimization automated hyperparameter tuning, while ensemble learning (20 models) ensured robustness for model development. The optimized model achieved an AUC of 0.8299 and accuracy of 77.65% on internal validation, outperforming existing tools with 75 % accuracy on external FDA-approved drugs. Interpretability analysis identified critical substructures correlated with high HOB, providing actionable insights for rational drug design. This work establishes a novel method for high-throughput screening of candidates with favorable bioavailability, highlighting the potential of deep learning to decode complex structure-property relationships in pharmaceutical optimization.