What ADMET values should flag a compound for follow-up?

Flag Caco-2 permeability below -5.15 log cm/s, which indicates poor oral absorption; hERG inhibition with a probability above 0.5, a cardiac safety risk; positive CYP3A4 inhibition, a drug-drug interaction liability; positive blood-brain barrier penetration for a non-CNS drug, which implies off-target CNS effects; positive AMES mutagenicity, a carcinogenicity concern; positive hepatotoxicity, a liver safety risk; and plasma protein binding above 95%, which leaves a low free drug fraction. These are triage thresholds for deciding what to test, not measurements.

Why run ADMET prediction before committing to synthesis?

Because design changes are cheap before synthesis and the failures are expensive afterwards. Roughly 90% of drug candidates fail in clinical trials, and poor pharmacokinetic properties are a leading cause; the average cost of bringing a single drug to market is estimated at $2.6 billion. Predicting absorption, distribution, metabolism, excretion and toxicity computationally lets you screen every compound while it is still only a structure. The predictions are probabilistic estimates rather than experimental measurements, so use them to prioritize which compounds to test, not to replace testing.

What input do ADMET prediction tools need?

A molecular structure, given as a SMILES string; some tools also accept SDF. Prediction runs from the structure alone, which is why a screen can cover compounds that have not been synthesized. If you are working from a protein-ligand docking result, extract the ligand SMILES. If you are designing molecules with a generative tool such as GenMol or REINVENT4, the output already includes SMILES directly. Every ADMET prediction starts at that structure, so structure generation or extraction is the first step of the workflow.

How do I redesign a compound flagged for an ADMET liability?

Change the structural feature that drives the flag, then re-screen. For high lipophilicity causing poor solubility, introduce polar groups or reduce aromatic rings. For CYP inhibition, modify the metabolically labile sites identified by CYP substrate prediction. For hERG liability, reduce basicity and lipophilicity, which are common hERG pharmacophore features. For poor blood-brain barrier penetration in a CNS program, reduce molecular weight, increase lipophilicity and reduce hydrogen bond donors. GenMol or REINVENT4 can generate analogs with ADMET-guided scoring functions, and the top candidates are then re-screened.

Which ADMET endpoints are most likely to kill a drug program?

hERG inhibition, AMES mutagenicity and hepatotoxicity, together with CYP450 interactions. hERG inhibition is a cardiac toxicity risk through ion channel blockade causing QT prolongation; AMES mutagenicity indicates DNA mutation potential and is a carcinogenicity indicator; hepatotoxicity is drug-induced liver injury risk; and inhibition of the major metabolic enzymes CYP1A2, 2C9, 2C19, 2D6 and 3A4 creates drug-drug interaction liability. LD50 acute oral toxicity and skin sensitization are further safety endpoints that can terminate a program.

ADMET Prediction Guide for Drug Candidates

ADMET — absorption, distribution, metabolism, excretion, and toxicity — determines whether a promising compound can become a viable drug. Roughly 90% of drug candidates fail in clinical trials (Sun et al., Acta Pharmaceutica Sinica B 2022), and poor pharmacokinetic properties are a leading cause. The average cost of bringing a single drug to market is estimated at $2.6 billion (DiMasi et al., J. Health Economics 2016). Predicting ADMET properties computationally, before committing to synthesis, is one of the most impactful applications of AI in drug discovery.

What ADMET Covers#

Absorption#

How well a compound enters the bloodstream after oral administration. Key endpoints include:

Caco-2 permeability: Intestinal absorption proxy (>-5.15 log cm/s is "high")
Human intestinal absorption (HIA): Percentage absorbed from the GI tract
Oral bioavailability: Fraction reaching systemic circulation unchanged
P-glycoprotein substrate/inhibitor: Whether efflux pumps reduce absorption

Distribution#

How the compound spreads through body compartments:

Plasma protein binding (PPB): Fraction bound to serum albumin — high binding reduces free drug
Volume of distribution (VDss): Extent of tissue distribution
Blood-brain barrier (BBB) penetration: Critical for CNS drugs, a liability for others

Metabolism#

How the body chemically transforms the drug:

CYP450 inhibition: Whether the compound blocks major metabolic enzymes (CYP1A2, 2C9, 2C19, 2D6, 3A4)
CYP450 substrate: Which enzymes metabolize the drug — affects drug-drug interactions
Metabolic stability: Half-life in liver microsomes

Excretion#

How the body eliminates the drug:

Total clearance: Rate of elimination from plasma
Half-life: Time for plasma concentration to decrease by half
Renal OCT2 substrate: Organic cation transporter involvement in kidney excretion

Toxicity#

Safety endpoints that can terminate a drug program:

hERG inhibition: Cardiac toxicity risk (ion channel blockade causing QT prolongation)
AMES mutagenicity: DNA mutation potential (carcinogenicity indicator)
Hepatotoxicity: Drug-induced liver injury risk
LD50: Acute oral toxicity classification
Skin sensitization: Allergic contact dermatitis potential

AI Models for ADMET Prediction#

Several AI tools predict ADMET properties from molecular structure alone (SMILES or SDF input):

ADMET-AI#

ADMET-AI uses Chemprop-RDKit graph neural networks trained on the Therapeutics Data Commons benchmark datasets, from authors at Stanford and Greenstone Biosciences (Swanson et al., Bioinformatics 2024). It predicts 41 ADMET endpoints and computes 8 RDKit physicochemical properties.

Input: SMILES string
Output: 41 ADMET endpoints plus 8 physicochemical properties, each reported alongside its percentile against 2,579 approved DrugBank drugs
Strength: Highest average rank on the TDC ADMET leaderboard, and fast enough for library-scale screening
Not provided: no per-prediction uncertainty and no applicability-domain score. Both fields are deliberately left absent rather than defaulted, so a novel scaffold returns values with no warning that it sits outside the training distribution. Judge domain fit yourself from the DrugBank percentiles and your own series

ADMETLab3 — retired, not available#

ADMETLab3 predicts 80+ ADMET properties (Fu et al., Nucleic Acids Research 2024), and it is worth knowing as an external reference point. It is no longer available here. The tool was retired on 2026-10-04 because its upstream service changed what it returned without changing its provenance, so a job that reported success could carry a different model's output. Every new dispatch is refused, and it is deliberately not re-routed to ADMET-AI, since silently serving one model's numbers under another's name is the same defect in a new place. Existing results stay viewable.

ProTox3 — retired, not available#

ProTox-3 classifies acute oral toxicity (LD50) into GHS categories and predicts 33 toxicity endpoints including organ-specific toxicity, immunotoxicity and endocrine disruption (Banerjee et al., Nucleic Acids Research 2024). Its Modal app was stopped, so it no longer runs here.

What to use instead depends on the endpoint, and it is worth being precise because the replacement is partial:

Overlapping endpoints — DILI, hERG, AMES, ClinTox, Tox21 and CYP inhibition — are covered by ADMET-AI.
Structural alerts are covered by a safety-alert screen.
The acute-toxicity GHS class and the organ-specific toxicity endpoints have no replacement. If your program turns on an LD50 class or an organ-level call, that is a gap to fill outside this platform rather than a number to substitute.

pkCSM#

A graph-based signature approach predicting pharmacokinetic properties (Pires et al., J. Med. Chem. 2015). Useful as an orthogonal method alongside neural network-based tools.

Input: SMILES
Output: 30 pharmacokinetic and toxicity parameters
Strength: Fast, interpretable, good for quick assessments

Practical Workflow#

Step 1: Generate or Obtain SMILES#

Every ADMET prediction starts with a molecular structure represented as a SMILES string. If you're working from a protein-ligand docking result, extract the ligand SMILES. If designing molecules with GenMol or REINVENT4, the output includes SMILES directly.

Step 2: Run Multi-Model ADMET Screen#

On Ivy Biosciences, submit the SMILES to ADMET-AI for a 41-endpoint screen. For toxicity-critical programs, add a structural-alert screen; note that organ-specific toxicity and an acute-toxicity GHS class are not available here since ProTox-3 was retired.

Step 3: Interpret the Results#

Key decision thresholds to watch for:

Property	Flag if...	Why it matters
Caco-2 permeability	< -5.15 log cm/s	Poor oral absorption
hERG inhibition	Probability > 0.5	Cardiac safety risk
CYP3A4 inhibition	Positive	Drug-drug interaction liability
BBB penetration	Positive (for non-CNS drugs)	Off-target CNS effects
AMES mutagenicity	Positive	Carcinogenicity concern
Hepatotoxicity	Positive	Liver safety risk
PPB	> 95%	Low free drug fraction

Step 4: Iterate on Design#

Compounds flagged for ADMET liabilities can be optimized:

High lipophilicity causing poor solubility? Introduce polar groups or reduce aromatic rings
CYP inhibition? Modify metabolically labile sites identified by CYP substrate prediction
hERG liability? Reduce basicity and lipophilicity — common hERG pharmacophore features
Poor BBB penetration (for CNS drugs)? Reduce molecular weight, increase lipophilicity, reduce hydrogen bond donors

Use GenMol or REINVENT4 to generate analogs with ADMET-guided scoring functions, then re-screen the top candidates.

Limitations to Keep in Mind#

AI ADMET predictions are probabilistic estimates, not experimental measurements:

Applicability domain: Models are most reliable for molecules similar to their training data. Novel scaffolds may produce unreliable predictions.
Endpoint correlation: Some predicted endpoints are derived from similar training data and may not be truly independent.
Species differences: Most models are trained on human data. Predictions may not translate to animal models used in preclinical testing.
Formulation effects: ADMET predictions assume standard oral administration. Formulation strategies can significantly alter bioavailability.

Always validate computational ADMET predictions with experimental assays for compounds advancing to synthesis. Use AI predictions to prioritize which compounds to test, not as a replacement for testing.

Key Takeaways#

ADMET failure accounts for a significant proportion of clinical trial failures — predicting early saves time and resources
Run ADMET screens on all compounds before committing to synthesis, when changes are cheap
Cross-check a liability against a second instrument where one exists, and treat an endpoint with only one available predictor as weaker evidence
Pay special attention to hERG inhibition, CYP interactions, and mutagenicity — these are program-killing liabilities
Iterate: use ADMET predictions to guide molecular design, not just as a pass/fail gate

References#

Swanson, K., Walther, P., Leitz, J. et al. "ADMET-AI: A machine learning ADMET platform for evaluation of large-scale chemical libraries." Bioinformatics 40, btae416 (2024). doi:10.1093/bioinformatics/btae416
Banerjee, P., Kemmler, E., Dunkel, M. & Preissner, R. "ProTox-3.0: a webserver for the prediction of toxicity of chemicals." Nucleic Acids Research 52, W513–W520 (2024). doi:10.1093/nar/gkae303
Fu, L., Shi, S., Yi, J. et al. "ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support." Nucleic Acids Research 52, W422–W431 (2024). doi:10.1093/nar/gkae236
Pires, D.E.V., Blundell, T.L. & Ascher, D.B. "pkCSM: Predicting Small-Molecule Pharmacokinetic and Toxicity Properties Using Graph-Based Signatures." Journal of Medicinal Chemistry 58, 4066–4072 (2015). doi:10.1021/acs.jmedchem.5b00104
DiMasi, J.A., Grabowski, H.G. & Hansen, R.W. "Innovation in the pharmaceutical industry: New estimates of R&D costs." Journal of Health Economics 47, 20–33 (2016). doi:10.1016/j.jhealeco.2016.01.012
Sun, D., Gao, W., Hu, H. & Zhou, S. "Why 90% of clinical drug development fails and how to improve it?" Acta Pharmaceutica Sinica B 12, 3049–3062 (2022). doi:10.1016/j.apsb.2021.11.005

A Practical Guide to ADMET Prediction for Drug Candidates