ADMET prediction is now routine in early drug discovery, and the tools are easier to access than they were ten years ago. Submit a SMILES string and receive predicted absorption, distribution, metabolism, excretion, and toxicity properties within minutes. The risk is that convenient outputs appear more authoritative than they are. These predictions can inform decisions, but confidence intervals may be wide. Use them constructively without treating them as definitive.
What ADMET models are based on
Most ADMET models use machine learning trained on databases of measured results. PAMPA data commonly informs intestinal permeability models. hERG patch clamp data supports cardiotoxicity models, while fluorescent substrate assays supply CYP inhibition data for metabolic stability predictions. The models learn structure-property relationships from these datasets and apply them to new compounds.
Prediction quality depends mainly on query similarity to the training data and on that property's dataset size and diversity. For properties with large public datasets, including logP, hERG, and Caco-2 permeability, predictions for drug-like compounds with standard scaffolds are reasonably reliable. They are directionally correct and within about a half-log unit of the true value for most compounds. Smaller or poorly curated datasets, such as those covering strain-specific mutagenicity, reactive metabolite formation, or CYP induction versus inhibition, produce less reliable results. Use these outputs chiefly as alerts, not quantitative estimates.
Absorption: where predictions are strongest
Absorption is among the most reliable ADMET areas. Calculated logP, molecular weight, hydrogen bond donor and acceptor counts, topological polar surface area (TPSA), and rotatable bond count come directly from structure. They underpin Lipinski's rules and the Veber criteria for oral bioavailability. These are calculations, not probabilistic predictions. A compound with TPSA above 140 square angstroms is unlikely to cross the intestinal epithelium passively, based on a strong physical rationale.
Predicted Caco-2 and PAMPA permeability are more genuinely predictive because they use ML models. They work best within the training set's applicability domain. Results are less dependable for macrocycles, highly charged compounds at physiological pH, or compounds with reactive electrophiles. Flag such compounds for confirmatory assay earlier than otherwise planned.
Metabolism: where caution is needed
CYP inhibition predictions, especially for CYP3A4, CYP2D6, and CYP2C9, are useful early filters. Polypharmacy makes CYP interactions a clinical liability even outside drug-drug-interaction indications. ML models perform acceptably for strong inhibitors, with IC50 below 1 uM, and weak non-inhibitors, but less well in between. A compound predicted to show moderate CYP3A4 inhibition, IC50 3 to 10 uM, may in fact show weak, moderate, or no clinically meaningful inhibition. Do not let that prediction alone determine whether to synthesize it.
Metabolic stability predictions, often reported as predicted human liver microsomal half-life or intrinsic clearance, can flag rapid-metabolism alerts: benzylic positions beside electron-withdrawing groups, unprotected para-positions on phenyl rings, N-methyl groups in some amine environments, and others. These rules are expert-coded heuristics, not ML predictions, and have established experimental precedent. When one appears, the alert is often more trustworthy than an ML-derived number.
Toxicity: the least certain area
The ADMET toxicity panel includes Ames mutagenicity, hERG cardiotoxicity, hepatotoxicity, and sometimes skin sensitization or reproductive toxicity. These are the results most prone to misuse.
Larger training datasets have improved hERG predictions, and current models reasonably flag strong hERG binders, IC50 below 1 uM, among drug-like compounds. False positives remain common for some scaffolds, especially those combining basic nitrogen atoms with aromatic rings. That pattern correlates with hERG binding in training data but also occurs in many safe, approved drugs. Do not treat every hERG flag as disqualifying without reviewing the confidence interval and comparing examples from the same scaffold class.
Hepatotoxicity is among the least reliable standard ADMET predictions. Drug-induced liver injury (DILI) can result from direct cytotoxicity, mitochondrial dysfunction, immune-mediated reactions, or reactive metabolites. These mechanisms look similar clinically but have different structural drivers. Models trained on clinical DILI outcomes must learn a multi-mechanism endpoint from limited labeled data. Results are, at best, directionally informative, and the false positive rate is high. A DILI flag without structural or mechanistic support should not drive action during scaffold exploration.
Using ADMET predictions constructively
In early discovery, treat ADMET predictions as flags, not gates. A hERG, metabolic stability, or absorption flag should prompt two questions: does the structure plausibly explain it, and does it alter synthesis priorities? Do not automatically remove the compound unless the flag reflects a well-established structural alert with strong experimental precedent.
A useful three-tier approach is: 1. Structural alerts, including Michael acceptors, nitroaromatics, certain quinones, and other established covalent reactivity alerts. Consider these seriously before synthesis. 2. ML property flags from known-good models, such as strong hERG signal, very low predicted solubility, or very high predicted CYP inhibition. Consider them with docking results when prioritizing synthesis, but do not automatically eliminate the compound. 3. Less reliable predictions, including DILI, reactive metabolite, and skin sensitization flags. Record them for downstream awareness, but do not let them drive synthesis during scaffold exploration.
The ADMET-first case and its limits
Running ADMET predictions before binding screens can make sense in some workflows. In a chemical series with known ADMET liabilities, filter those liabilities first, then dock the remainder. This is more efficient than docking everything and filtering the top scorers afterward. It matters especially in GPCR programs, where some scaffold classes tend toward hERG liability and CYP inhibition and the filters remove a meaningful share of scaffold space.
For most structure-based programs targeting well-characterized kinases or proteases with drug-like chemotype preferences, docking and ADMET are run in parallel and the results are combined. They assess different aspects of compound quality. A compound performing well on both is a higher-confidence starting point than one performing well on only one.
