What is ADMET, and how is it predicted?

Five questions about a molecule, and what two models say when asked the same ones

11 min

Line drawing of fluconazole at the centre of a five-axis chart, one axis per ADMET letter, with two model profiles that nearly coincide on four axes and split widely on the excretion axis.

What is ADMET?

ADMET stands for absorption, distribution, metabolism, excretion and toxicity: the five things that happen to a molecule between swallowing it and clearing it from the body, and the five reasons a potent molecule still fails as a drug. ADMET prediction estimates all five from the structure alone, so that a chemist can drop weak candidates before synthesizing them.

The estimates look more certain than they are. We ran the same 12 approved drugs through two published models, ADMET-AI and Admetica, and the half-life predictions for fluconazole came back as 0.6 hours from one and 30.4 hours from the other. The FDA label gives about 30 hours.[9] Caffeine drew a hERG probability of 0.05 from one model and 0.99 from the other.

Both numbers come out of a neural network that has never seen the molecule in a body. What each model learned, how big its training set was and what its probability really means decide whether a prediction should move a project. The sections below take the five letters in turn, explain how the models are built, and then return to those 12 drugs.

Why the five letters matter

Drugs fail for reasons that have little to do with whether they bind the target. Kola and Landis tracked candidates from 10 large pharmaceutical companies and found that poor pharmacokinetics and bioavailability caused about 40% of failures in 1991 and about 10% by 2000.[1] The failures did not disappear. They moved to efficacy and safety.

Safety is the harder half. Waring and colleagues pooled attrition data from AstraZeneca, Eli Lilly, GlaxoSmithKline and Pfizer and found a link between physicochemical properties and clinical failure for safety reasons, but concluded that further tightening is unlikely to change attrition much.[2]

Simple rules such as Lipinski's rule of five screen for absorption problems.[7] They say nothing about whether a liver enzyme will destroy the compound or a cardiac channel will block it.

Table 1 lists the five questions, the experiment that normally answers each and the quantities a predictor returns.

LetterQuestionTypical assayTypical predicted endpoints
AbsorptionHow much of an oral dose reaches the blood?Caco-2 cell permeability, PAMPA, kinetic solubilityIntestinal absorption, Caco-2 permeability, solubility, lipophilicity, P-glycoprotein inhibition
DistributionWhere does it go, and how much stays free?Plasma protein binding, tissue distribution in animalsBlood-brain barrier penetration, plasma protein binding (%), volume of distribution (L/kg)
MetabolismWhich enzymes change it, and does it block them?Liver microsome and recombinant CYP assaysCYP1A2, 2C9, 2C19, 2D6 and 3A4 inhibition and substrate probabilities
ExcretionHow fast is it cleared?Hepatocyte and microsome clearance, animal and human pharmacokineticsHalf-life (hours), hepatocyte and microsome clearance
ToxicityDoes it harm cells, the heart, the liver or DNA?hERG patch clamp, Ames test, cytotoxicity, animal LD50hERG blocking, mutagenicity, liver injury, clinical toxicity, rat LD50
Table 1. The five ADMET questions, the assays behind them and what a predictor reports. Endpoint names follow the Therapeutics Data Commons datasets that ADMET-AI and Admetica draw on.[4]

How ADMET is predicted

A modern ADMET predictor is a set of separate models, one per endpoint. Each is trained on a public table of molecules with a measured value, for example a few hundred compounds with a hERG result or about 12,000 with a CYP3A4 inhibition result. The input is a SMILES string, which the software turns into a graph of atoms and bonds.

ADMET-AI uses Chemprop message-passing neural networks, which pass information between neighbouring atoms several times to build a fingerprint of the molecule that the network learns itself.[10] Swanson and colleagues trained them on 41 datasets from the Therapeutics Data Commons and report the highest average rank on its ADMET leaderboard.[3] Admetica also uses Chemprop but trains its own 22 models.[5]

Classification endpoints return a probability that the molecule belongs to the positive class, such as blocking hERG or inhibiting CYP3A4. Regression endpoints return a number in the assay's unit, such as log mol/L for solubility. Neither is a measurement. Both are the output of a function fitted to someone else's assay, which fixes what they can and cannot say.

ADMET-AI adds one more thing: it ranks each prediction against the model's own predictions for approved drugs in DrugBank. A raw 0.96 absorption probability for aspirin sits at only the 32nd percentile of that reference, which tells you more than the 0.96 does.

Where predicted ADMET goes wrong

The training sets differ enormously in size. The CYP inhibition datasets hold 12,000 to 13,000 compounds each, while the liver injury, carcinogenicity and hERG datasets hold 278 to 648. A model trained on 648 hERG results has seen very little of chemical space, and it still returns a confident-looking probability for a macrocycle, a peptide or a PROTAC that resembles nothing in it.

The label also changes meaning across endpoints. A hERG probability reports whether a compound blocked a channel in an assay.

Whether that matters clinically depends on exposure. Redfern and colleagues compared preclinical hERG data with clinical QT prolongation and torsades de pointes for 100 drugs and proposed a margin of about 30-fold between the free plasma concentration and the hERG IC50, because the relationship between channel block and risk was weak outside the antiarrhythmics.[6] A model of the first cannot deliver the second.

Half-life is the extreme case. It is a property of a whole animal, set by clearance, distribution volume and protein binding together, yet the models predict it from a few hundred or thousand molecules with one number. The next section shows what that does.

Two models, twelve drugs

We ran twelve approved drugs through ADMET-AI 2.0.1 and Admetica 1.4.1 on a CPU, with the bundled models and default settings, from the SMILES in the input file. The set spans hERG-active drugs (terfenadine and cisapride, both withdrawn over QT prolongation, and the antiarrhythmic dofetilide),[6] a strong CYP3A4 inhibitor (ketoconazole),[8] small polar drugs (metformin, aspirin, caffeine) and three common medicines (atorvastatin, imatinib, warfarin).

The raw outputs are in the ADMET-AI and Admetica files. Table 2 shows three endpoints.

DrughERG blockerCYP3A4 inhibitorHalf-life (h)
Terfenadine0.97 / 1.000.05 / 0.0234.6 / 13.7
Cisapride0.97 / 1.000.80 / 0.3344.5 / 8.6
Dofetilide0.92 / 1.000.77 / 0.6815.2 / 6.9
Verapamil0.97 / 1.000.91 / 0.9927.7 / 2.0
Ketoconazole0.96 / 0.770.99 / 1.0050.0 / 33.2
Fluconazole0.36 / 0.230.92 / 1.000.6 / 30.4
Aspirin0.02 / 0.000.00 / 0.00−19.7 / 2.9
Caffeine0.05 / 0.990.00 / 0.005.2 / 14.7
Metformin0.05 / 0.290.00 / 0.00−6.3 / 6.0
Atorvastatin0.72 / 0.560.37 / 0.0949.9 / 5.6
Imatinib0.98 / 1.000.79 / 0.9066.3 / 28.5
Warfarin0.28 / 0.300.19 / 0.0311.5 / 29.2
Table 2. ADMET-AI 2.0.1 and Admetica 1.4.1 predictions for twelve approved drugs. Each cell gives ADMET-AI first and Admetica second. hERG and CYP3A4 are probabilities of the positive class; half-life is in hours. We ran both packages locally with default settings; the drugs are not a validation set and several probably appear in the training data.

The two models agree where the data are plentiful. Their CYP3A4 inhibition probabilities have a Spearman correlation of 0.97 across the twelve drugs and differ on one side of the 0.5 line only for cisapride. Lipophilicity (0.91) and solubility (0.87) also track each other closely. Both give ketoconazole 0.99 or higher, as expected for a drug the FDA lists as a strong CYP3A4 inhibitor.[8]

They agree less on hERG (0.80), with one flip: caffeine. A single drug moving from 0.05 to 0.99 between two models built the same way is a reminder that a small training set does not give a stable decision boundary. The hERG column also shows that a high value is not discriminating, because ADMET-AI calls seven of the twelve drugs blockers (probability above 0.5) and Admetica eight, including widely used medicines such as imatinib and atorvastatin.

Half-life is where they come apart. The correlation is 0.24 and the mean absolute difference 23.5 hours. ADMET-AI predicts negative half-lives for aspirin and metformin, which is not a physical value. Neither model is simply better. Admetica lands near the label value for fluconazole, but one drug cannot show which model is more reliable. The practical reading is that a single model's half-life predictions cannot rank a series.

Table 2 does not measure accuracy. We compared the models with each other, not with experiment, and only the fluconazole half-life was checked against a label. The agreement is probably flattering, since well-known drugs are the compounds most likely to sit in both training sets. Novel chemistry is the harder case.

Using predictions without over-reading them

The way to use ADMET predictions follows from the pattern in Table 2. Rank within a series and look at the endpoints with large training sets. Distrust single-model calls on small-data endpoints. Look for disagreement between models as a signal that the molecule is far from what they learned. Table 3 turns that into a routine.

OutputWhat it cannot tell youSign of troubleCheap check
CYP inhibition probabilityClinical interaction risk, which depends on dose and exposureProbability near 0.5 or a flip between modelsRun a second model and confirm with a CYP panel for the leading series
hERG probabilityWhether QT prolongation will occur at the doseMost of a series scored above 0.9Rank within the series; confirm by patch clamp and compare with free exposure
Half-life or clearanceAnything about a whole animalNegative or implausible values, models differing several-foldUse the DrugBank percentile, then measure microsomal stability
Solubility, lipophilicityFormulation behaviourPrediction outside the training rangeCompare with a calculated LogP and a kinetic solubility assay
Toxicity score or alert countMechanism or dose dependenceAn alert on a drug-like scaffoldInspect the matched substructure before discarding the compound
Table 3. What to do with each kind of ADMET output. The checks are the cheapest ones that would catch the failure shown in Table 2.

Predictions decide which compounds deserve an assay, not whether the assay is needed. They fit best after a virtual screening run, as a filter on the shortlist, and before any compound is ordered.

ADMET tools on ProteinIQ

Start with a broad profile, add a second model on the endpoints that decide the project, then use rule-based filters for the substructure questions that models handle poorly. The ADMET prediction use case walks through the sequence.

TaskTool
Profile a molecule on 41 ADMET endpoints with DrugBank percentilesADMET-AI
Get a second model for metabolism, clearance, half-life and hERGAdmetica
Predict lipophilicity, permeability and a solubility flagBoltz ADME
Score general toxicity probability and synthetic accessibilityeToxPred
Flag problem substructures with PAINS, Brenk and NIH alertsStructural alert screening
Count rule-of-five violationsLipinski's rule of 5
Table 4. ADMET and toxicity prediction tools on ProteinIQ.

Frequently asked questions

What does ADMET stand for?

ADMET stands for absorption, distribution, metabolism, excretion and toxicity. It covers how a drug enters the body, where it goes, how enzymes change it, how it leaves and whether it causes harm. Poor pharmacokinetics caused about 40% of clinical failures in 1991 and about 10% by 2000.[1]

How is ADMET predicted?

Models learn the relationship between molecular structure and a measured property from public assay data. ADMET-AI trains Chemprop graph neural networks on 41 Therapeutics Data Commons datasets and reports the highest average rank on that leaderboard.[3] The input is a SMILES string and the output is a probability or a value for each endpoint.

How accurate are ADMET predictions?

It depends on the endpoint. Datasets range from 278 compounds for some toxicity endpoints to about 13,000 for CYP inhibition. In our comparison of two published models on 12 approved drugs, CYP3A4 inhibition probabilities correlated at 0.97 while half-life predictions correlated at 0.24 with a mean difference of 23.5 hours.

Can ADMET prediction replace lab assays?

No. A prediction ranks compounds and flags risks so fewer assays are wasted on poor candidates. A hERG probability does not give a QT prediction, because clinical risk depends on exposure, and Redfern and colleagues proposed a margin of about 30-fold between free plasma concentration and the hERG IC50.[6]

What is the difference between ADMET and drug-likeness?

Drug-likeness rules such as Lipinski's rule of five use a few calculated properties to judge the odds of oral absorption.[7] ADMET models estimate specific outcomes such as CYP inhibition or hERG blocking, so they cover metabolism and toxicity that the rules never see.

Is there a free online ADMET predictor?

The ADMET-AI authors run a free web server and release the Python package as open source.[3] ProteinIQ runs the same package online with saved jobs, and eToxPred adds a general toxicity score.[11]