Research work · 2026
Machine Learning in Predicting the Physicochemical Properties of Drug Candidates
Anila Nuthi, Vaibhav A. Dixit
Machine Learning and Deep Learning in Drug Design
Abstract
Physicochemical properties strongly influence the pharmacokinetic and pharmacodynamic behaviour of drugs and drug-like molecules. Experimental measurement of these properties is resource-intensive, requiring significant time, cost, and specialized equipment. As a result, machine learning (ML) and deep learning (DL) methods have emerged as efficient alternatives for predicting key properties that impact absorption, distribution, metabolism, elimination, and toxicity (ADMET). ML models trained on large datasets can provide accurate predictions using 2D and 3D molecular descriptors, while DL approaches further enhance performance by learning directly from molecular representations, capturing complex factors such as electronic effects, structural symmetry, and conformational flexibility. Numerous tools, both commercial and open-source, are now available for predicting physicochemical properties relevant to drug discovery. However, the scarcity of high-quality experimental data and reliance on calculated values for model training remain major challenges for building robust and generalizable models. Moving forward, the integration of expert chemical knowledge with ML/DL frameworks is expected to improve predictive accuracy and reliability, helping to reduce late-stage failures and enabling earlier, more informed decision-making in drug development.