Leveraging machine learning models in evaluating ADMET properties for drug discovery and development

Abstract

Background and purpose: The evaluation of ADMET properties remains a critical bottleneck in drug discovery and development, contributing significantly to the high attrition rate of drug candidates. Traditional experimental approaches are often time-consuming, cost-intensive, and limited in scalability. This review aims to investigate how recent advances in machine learning (ML) models are revolutionizing ADMET prediction by enhancing accuracy, reducing experimental burden, and accelerating decision-making during early-stage drug development.

Experimental approach: This article systematically examines the current landscape of ML applications in ADMET prediction, including the types of algorithms employed, common molecular descriptors and datasets used, and model development workflows. It also explores public databases, model evaluation metrics, and regulatory considerations relevant to computational toxicology. Emphasis is placed on supervised and deep learning techniques, model validation strategies, and the challenges of data imbalance and model interpretability.

Key results: ML-based models have demonstrated significant promise in predicting key ADMET endpoints, outperforming some traditional quantitative structure - activity relationship (QSAR) models. These approaches provide rapid, cost-effective, and reproducible alternatives that integrate seamlessly with existing drug discovery pipelines. Case studies discussed in this review illustrate the successful deployment of ML models for solubility, permeability, metabolism, and toxicity predictions.

Conclusion: Machine learning has emerged as a transformative tool in ADMET prediction, offering new opportunities for early risk assessment and compound prioritization. While challenges such as data quality, algorithm transparency, and regulatory acceptance persist, continued integration of ML with experimental pharmacology holds the potential to substantially improve drug development efficiency and reduce late-stage failures.

Keywords: ADMET prediction; AI/ML; computational toxicology; molecular descriptors; pharmacokinetics.

Figure 1.

This figure depicts the absorption, distribution, metabolism, and excretion (ADME) process for drugs administered orally. When a tablet or capsule is ingested, it disintegrates in the gastrointestinal (GI) tract, releasing drug molecules. These molecules either dissolve for absorption or remain in a precipitated state, eventually being excreted. Absorbed drug molecules must cross the gut wall, where they may be transported back into the intestinal lumen or metabolized by enzymes. Those that successfully traverse the gut barrier enter the liver via the portal circulation. The liver plays a crucial role in drug metabolism, utilizing phase I (modification) and phase II (conjugation) enzymatic reactions to increase the hydrophilicity of xenobiotics, facilitating their elimination through the kidneys. Drugs that escape metabolism enter systemic circulation, though a portion binds to plasma proteins, limiting their bioavailability. Only the free, unbound drug and its metabolites can reach target cells and interact with biomolecules to exert therapeutic effects. Meanwhile, some drug molecules are rapidly cleared by the kidneys. The drug’s efficacy is determined by its ability to reach and maintain an optimal concentration at the site of action while navigating these physiological processes (Courtesy:NIH Bioart; Bioicons).

Figure 2.

Commonly used AI/ML algorithms for developing ADMET prediction models

Figure 3.

The diagram illustrates the stepwise workflow for generating a machine learning model, starting from raw data collection to model evaluation. Initially, raw data, which can include both labeled and unlabeled datasets, undergoes preprocessing and structuring. This iterative step ensures data quality before proceeding further. Once the data is refined, it is split into training and test sets, commonly in an 80:20 ratio of the total volume of the data. The 80 % training data is then subjected to various machine learning algorithms, including supervised, unsupervised, reinforcement learning, and deep learning approaches. Through an iterative process, the best-performing candidate model is selected. The chosen candidate model is further refined through hyperparameter optimization and feature selection, leading to the final optimized model. This model is then validated using cross-validation techniques such as 5-fold cross-validation to ensure robustness. Following validation, the model is tested using the 20 % test dataset. The predicted outcomes (y-values) are analysed, and the model's performance is evaluated using classification and regression metrics, ensuring its reliability and accuracy for real-world applications (Courtesy: Bioicons; Flaticon).