AI and machine learning innovations for ADMET and toxicity prediction in modern drug discovery

Absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties are fundamental determinants of drug efficacy and safety, yet they remain a leading source of attrition across the drug development pipeline. Traditional experimental methods for evaluating these endpoints, both in vitro and in vivo, are costly, slow, ethically constrained, and often limited in cross-species predictability. Computational ADMET modeling has therefore become an indispensable component of modern drug discovery, transitioning from early quantitative structure-activity relationship (QSAR) models to advanced machine learning frameworks capable of learning directly from molecular graphs, 3D geometries, and protein sequence data. Advances in artificial intelligence and machine learning have transformed this field. Graph neural networks, transformer-based chemical language models, and self-supervised pretraining on massive unlabeled chemical datasets now deliver state-of-the-art benchmarks on established public challenges such as Therapeutics Data Commons, MoleculeNet, and Tox21. Emerging multi-task, transfer, and federated learning paradigms are expanding the chemical and biological diversity available for model training, while uncertainty quantification, conformal prediction, and applicability domain estimation are increasing the interpretability and trustworthiness of predictions. Nevertheless, challenges persist: data imbalance for rare toxicities, limited generalization to novel scaffolds, opaque model decision-making, and limited regulatory acceptance of AI-derived endpoints all hinder translation into pharmaceutical environments. Addressing these barriers requires methodological transparency, open evaluation standards, and collaborative integration among computational scientists, medicinal chemists, toxicologists, and regulators.This Research Topic aims to consolidate frontier research and practical developments in AI- and ML-driven ADMET and toxicity prediction, emphasizing algorithmic innovation, reproducible validation, interpretability, and integration into decision-making and regulatory contexts for safer, more efficient drug discovery. To advance understanding and application in this area, we welcome articles encompassing theoretical developments, dataset resources, and applied studies. Submissions emphasizing reproducibility, open data, and seamless integration between computational and experimental pipelines are particularly encouraged. Topics of interest include, but are not limited to: • Deep learning architectures for ADMET prediction: graph neural networks, transformer models, 3D-aware and equivariant models, and chemical language frameworks • Toxicity endpoint modeling, including hepatotoxicity, cardiotoxicity (hERG, QT prolongation), nephrotoxicity, genotoxicity, carcinogenicity, and organ- or pathway-specific toxicities • Data curation, augmentation, and benchmarking: harmonized datasets, noise reduction, small-data learning, and community-wide performance evaluation • Interpretability, explainability, and uncertainty quantification: feature attribution, mechanistic insights, conformal prediction, and applicability domains • Multi-task, transfer, and federated learning for cross-endpoint and privacy-preserving modeling • Integration of ADMET predictors into generative chemistry and lead optimization pipelines using reinforcement or active learning • Physiologically based pharmacokinetic (PBPK) and hybrid ML approaches for systemic drug behavior and interaction prediction • Translational and regulatory perspectives for model validation, deployment, and alignment with emerging new approach methodologies (NAMs) and regulatory standards Topic editors: Rahul Singh, Balaji Seetharaman Source: https://www.frontiersin.org/research-topics/84942/ai-and-machine-learning-innovations-for-admet-and-toxicity-prediction-in-modern-drug-discovery