ADMET and Toxicity Prediction

ADMET is a collection of pharmacokinetic and safety properties: absorption, distribution through the body, metabolism, excretion, and toxicity. Different endpoints measure different biological processes, such as solubility, permeability, plasma protein binding, metabolic stability, clearance, or a particular toxicity assay. There is no single ADMET score that proves a compound is suitable for humans. Machine-learning models learn from experimental datasets or curated databases using molecular fingerprints, descriptors, graphs, or learned representations. Labels can be noisy because assays vary in protocol, cell line, organism, concentration, and measurement units. Some datasets combine values from different sources or convert continuous measurements into categories. Endpoint definitions and unit conversions should be checked before training. Evaluation must reflect the intended use. Random splits can place close analogs in training and test sets, making prediction easier than for new chemical scaffolds. Scaffold-based, temporal, or external splits can better probe generalization, depending on the question. Class imbalance, censored values, repeated compounds, and assay leakage also need attention. Report uncertainty and the model's applicability domain: whether a candidate resembles the chemistry and conditions used for training. A positive model prediction is a hypothesis for follow-up, not experimental evidence. Use models to prioritize tests, compare with baselines, and identify compounds where predictions are uncertain or endpoints disagree. High-stakes toxicology decisions require appropriate assays and expert review. A model trained on one endpoint cannot establish broader safety or human outcomes. OECD guidance on QSAR model validation emphasizes a defined endpoint, unambiguous algorithm, defined applicability domain, appropriate measures of goodness-of-fit and predictivity, and mechanistic interpretation where possible. These principles help make structure-activity predictions more transparent. Maintain data provenance and versioning so predictions can be traced to assays, structures, and model configurations.