Authors : Rohith Reddy G K, Sheik Arshad Ibrahim, Sayed Jahangir Ali, Thirumurugan M
DOI : 10.1109/icauc68182.2026.11441036
Volume : 1
Issue : 1
Year : 2026
Page No : 1-8
High attrition rates in drug discovery are frequently driven by unfavourable absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties, leading to late-stage failures despite promising efficacy. To address this challenge, we present ADMET-X, a biologically grounded and interpretable machine learning framework for comprehensive in silico prediction of pharmacokinetic and toxicological endpoints. ADMET-X integrates chemically meaningful physicochemical descriptors and structural fingerprints with ensemble learning models, including Random Forest and XGBoost, to capture complex structure–property relationships while preserving interpretability. The framework was evaluated on 31 benchmark datasets obtained from the Therapeutics Data Commons, spanning absorption, distribution, metabolism, excretion, and toxicity endpoints, using standardized training, validation, and test splits. Model performance was assessed using ROC–AUC and accuracy for classification tasks and RMSE and coefficient of determination (R2) for regression tasks. ADMET-X achieved strong predictive performance across multiple endpoints, with ROC–AUC values exceeding 0.90 for key absorption, metabolism, and toxicity predictions, and competitive regression accuracy for physicochemical properties. Unlike black-box deep learning approaches, ADMET-X provides mechanistic insight through feature importance analysis, enabling transparent interpretation of predictions. Overall, the proposed framework offers a scalable, reproducible, and interpretable platform for early-stage ADMET screening and rational drug candidate prioritization.