Application of Deep Learning and Machine Learning Algorithms in Predicting Drug Stability and Bioavailability

Abstract

Predicting the shelf life and oral absorption of drug candidates during the early phases of pharmaceutical development remains a formidable challenge, frequently contributing to expensive late-stage failures. In this investigation, deep learning (DL) and machine learning (ML) techniques were employed to forecast two pivotal pharmaceutical parameters chemical stability and oral bioavailability using a carefully assembled dataset of 450 small-molecule compounds with experimentally validated profiles. Five distinct computational architectures were constructed and systematically benchmarked: Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), Artificial Neural Network (ANN), and Convolutional Neural Network (CNN). Molecular descriptors and extended-connectivity fingerprints served as input representations, and model performance was assessed through root mean square error (RMSE), coefficient of determination (R²), and mean absolute error (MAE). For bioavailability estimation, the CNN architecture produced the strongest predictive capability (R² = 0.91, RMSE = 8.34%), whereas the GBM model outperformed all others in stability forecasting (R² = 0.89, RMSE = 4.21 months). Interpretability analysis using SHapley Additive exPlanations (SHAP) identified lipophilicity, molecular weight, hydrogen bond donor count, and topological polar surface area as the descriptors with the greatest influence on prediction outcomes. These results illustrate the capacity of DL and ML methodologies to meaningfully accelerate pharmaceutical screening by providing early-stage estimates of stability and bioavailability, thereby supporting more informed decision-making throughout the drug development pipeline.