ADMET-AI

ADMET-AI online overview

Run aspirin through ADMET-AI and it comes back with a 0.96 probability of human intestinal absorption, which sounds excellent until the percentile column puts it at only the 32nd percentile of approved drugs. That pairing of an absolute prediction with DrugBank context is what ADMET-AI is built for: ranking hundreds of candidate molecules on absorption, distribution, metabolism, excretion and toxicity before any of them is synthesized.

This page runs ADMET-AI 2.0.1 online for ADMET prediction from SMILES. Each molecule gets 41 model predictions trained on Therapeutics Data Commons (TDC) datasets, eight RDKit physicochemical properties, three structural alert counts, and a DrugBank percentile for every one of those 52 values. Results include a sortable table, a DrugBank comparison plot, per-molecule radial plots and structure drawings, and the native CSV files.

Pricing

A one-molecule job costs 4 credits, and the quote is calculated before submission from the number of molecules. CSV headers and rows with no SMILES do not count.

Molecules in one job	Credits
1	4
10	7
25	10
100	17
250	26
500	36
1,000	50

These cells come from the current calculator with one SMILES per line. Model and reference choice, physicochemical properties, percentiles and plot settings do not change the quote.

Inputs

Input	Accepted formats	Notes
SMILES	Pasted text, or .csv, .tsv, .smi, .smiles, .txt	Plain SMILES (CCO) or name-tab-SMILES (aspirin<TAB>CC(=O)Oc1ccccc1C(=O)O) per line. .smi rows follow the conventional SMILES name order.
PubChem fetch	Compound name or CID	Retrieves SMILES from PubChem into the same input.
DrugBank reference CSV	.csv	Custom reference modes only. Uses the native ADMET-AI DrugBank schema, including atc for class filtering.
Chemprop model ensembles (ZIP)	.zip	Custom model modes only. Chemprop ensemble directories with .pt checkpoints.

A job accepts up to 1,000 molecules; Free plans accept up to 10 per job. Each file can be up to 50 MiB. CSV structure and molecule counts are checked before submission, and RDKit validates the chemistry during the run: invalid molecules are reported and removed, and a run with no valid molecules fails.

Checkpoint archives must also expand to at most 50 MiB and use stock Chemprop classes; archives that need executable custom classes are rejected.

Settings

Models and outputs

Parameter	Type	Default	Description
Model and reference files	enum	Bundled models and reference	Bundled models and reference, Custom DrugBank reference, Custom Chemprop models or Custom models and reference. Custom modes require the matching file input.
Physicochemical properties	boolean	true	Includes the RDKit physicochemical properties and structural alert counts.
DrugBank percentiles	boolean	true	Adds DrugBank percentiles, the comparison plot and radial plots. Must stay on in custom reference modes.
ATC filter	string	optional	Therapeutic class name from the ADMET-AI reference data. Blank compares against all approved drugs in the reference.
SMILES column	string	smiles	Exact, case-sensitive SMILES column in CSV or TSV uploads. A missing column is rejected before the run.

Plots

Parameter	Type	Default	Description
Comparison plot format	enum	SVG	SVG, PNG, PDF, JPEG, TIFF, WEBP, EPS or PS for the comparison and radial plots. Structure drawings stay SVG.
X-axis property	enum	Human intestinal absorption	Property on the comparison plot x-axis. Any of the 52 properties can be chosen.
Y-axis property	enum	Clinical toxicity probability	Property on the comparison plot y-axis.
Label plotted molecules	boolean	false	Numbers the submitted molecules on the comparison plot.
Max labels	integer	25	Molecules labeled when labeling is on, 1 to 25.
Job name	string	ADMET analysis	Label shown in the job history.

Outputs

The Results tab is a spreadsheet with one row per valid molecule. Column labels carry units, for example Half-life (hr) or hERG blocking (probability), and each property has a matching DrugBank percentile column.

Group	Columns
Compound identity	ids, smiles
Physicochemical properties	molecular_weight, logp, hba, hbd, lipinski_rules_passed, lipinski_violations, qed_score, stereo_centers, tpsa
Structural alerts	pains_alert, brenk_alert, nih_alert
Absorption	hia_absorption, bioavailability, solubility_logS, lipophilicity, hydration_free_energy, caco2_permeability, pampa_permeability, pgp_inhibitor
Distribution	bbb_penetration, plasma_protein_binding, volume_distribution
Excretion	half_life, clearance_hepatocyte, clearance_microsome
Metabolism	cyp1a2_inhibitor, cyp2c19_inhibitor, cyp2c9_inhibitor, cyp2d6_inhibitor, cyp3a4_inhibitor, cyp2c9_substrate, cyp2d6_substrate, cyp3a4_substrate
Toxicity	herg_blocker, clinical_toxicity, ames_mutagenicity, dili_hepatotoxicity, carcinogenicity, ld50_toxicity, skin_reaction, and the twelve Tox21 nr_* and sr_* assays
Percentile context	*_percentile for every column above except identity and lipinski_violations

The Files tab holds the native admet_predictions.csv with source property names and unrounded values, admet_predictions_with_inputs.csv with submitted IDs and metadata (colliding metadata headers get an input_ prefix), the DrugBank comparison plot, radial plots and SVG structure drawings for the first 25 valid molecules, and asset, runtime and diagnostic records.

lipinski_violations is the only derived column: four minus the native lipinski_rules_passed. Custom model ensembles may return different endpoints, which keep their native names. Version 1 jobs keep admet_predictions.csv for the ID and metadata export and save native predictions as admet_native_predictions.csv.

Understanding results

Probabilities and regression values

Each column is one of two kinds. Classification endpoints, labeled (probability), give the probability that the molecule has the property, such as crossing the blood-brain barrier or blocking hERG. Regression endpoints predict a value in the unit shown: log mol/L for solubility, % for plasma protein binding, L/kg for volume of distribution, or log 1/(mol/kg) for rat LD50, where higher means more toxic.

Regression values are reported exactly as the model returns them, so they can fall outside a physical range. In the public aspirin example, half_life is -19.7 hours. That is not a measurement; it says the half-life model has little to offer for this molecule, and the 2nd-percentile rank is the more usable reading.

DrugBank percentiles

A percentile places the molecule's prediction among predictions for the DrugBank approved-drug reference: 2,845 drugs in the bundled ADMET-AI 2.0.1 data, or the subset selected by ATC filter. A value of 32 means roughly a third of reference drugs got a lower prediction. Percentiles compare predictions with predictions, not with measured drug properties.

Direction depends on the property. Higher is better for absorption, bioavailability and solubility, and worse for hERG blocking, clinical toxicity or DILI. Aspirin's hERG probability of 0.02 sits at the 13th percentile, a reassuring pair.

Aspirin's DILI probability of 0.67 sits at the 64th, so about a third of approved drugs score higher on the same model, and the raw 0.67 is not a liver-injury verdict. Narrowing the reference with ATC filter to the class a compound is meant to join often makes the comparison more meaningful.

Radial plot

Each radial plot shows five percentiles on axes where outward is better. Bioavailable and Soluble are the oral bioavailability and aqueous solubility percentiles; Blood-Brain Barrier Safe, hERG Safe and Non-Toxic are 100 minus the BBB penetration, hERG and clinical toxicity percentiles. For aspirin the three safety axes read 56, 87 and 47, and the other two 71 and 75. A peripheral drug wants the BBB axis high; a CNS drug wants it low.

Structural alerts and physicochemical properties

pains_alert, brenk_alert and nih_alert count substructure matches against the PAINS, Brenk and NIH filter sets. They are counts, not verdicts: aspirin scores one Brenk alert, for its phenol ester, and is still an approved drug. The PAINS filter and Brenk filter show which substructure matched, and Lipinski's rule of five explains the lipinski_rules_passed count. These columns are computed with RDKit, not predicted.

What not to over-read

Every ADMET column is a Chemprop model trained on one TDC dataset, and those datasets differ a lot. CYP inhibition models learned from 12,000 to 13,000 compounds each, while DILI, carcinogenicity and hERG come from 278 to 648 compounds.

A molecule unlike anything in the training data, such as a macrocycle, peptide, PROTAC or organometallic, still receives a confident-looking number, and the tool gives no applicability domain warning. Use the scores to rank and flag within a series, then confirm the endpoints that matter with assays.

ProteinIQ and the ADMET-AI web server

The ADMET-AI authors run the ADMET-AI web server, a collaboration between Greenstone Biosciences and James Zou's lab at Stanford University. ProteinIQ is not affiliated with that server. It runs the published ADMET-AI package on its own infrastructure.

The two give different numbers. According to the ADMET-AI repository, the authors' web server is based on ADMET-AI v1 with Chemprop-RDKit models, while v2 uses retrained Chemprop v2 models whose predictions do not exactly match v1.

The authors' server compares molecules with 2,579 DrugBank approved drugs, shows predictions for the first 25 molecules on screen with the full set as a CSV download, clips displayed regression values to a valid range, and does not store submitted molecules.

ProteinIQ runs v2.0.1 with its updated reference and the PAINS, Brenk and NIH alert counts. The results table covers every molecule in the job without clipping, plots come in eight formats, and custom DrugBank references or Chemprop ensembles can replace the bundled assets.

Jobs are saved with their inputs and results, which suits repeated triage but not molecules that must never be stored. Both cap a job at 1,000 molecules, and Free plans here are limited to 10. Paid plans can also submit jobs through the ProteinIQ API.

How ADMET-AI works

ADMET-AI v2 represents each molecule as a graph of atoms and bonds and passes it through an ensemble of Chemprop v2 message-passing neural networks trained on 41 TDC ADMET datasets. RDKit computes the physicochemical properties and alert counts directly. The DrugBank reference holds the same models' predictions for approved drugs, and each new prediction is ranked against that distribution to give the percentiles.

For a second opinion from different models, Admetica predicts 22 ADMET endpoints with its own Chemprop models, and eToxPred scores toxicity and synthetic accessibility. In a virtual screening campaign, ADMET-AI fits after docking as a filter on the shortlisted hits.

Limitations

Predictions are only as transferable as the TDC assays behind them. Endpoints differ in noise, species (rat LD50, rodent carcinogenicity, mouse skin reaction) and assay definitions, so a probability from one model is not comparable in confidence with another. Small training sets make some toxicity calls fragile, and molecules outside drug-like chemical space are extrapolations.

Use the output for prioritization and hypothesis generation, not standalone go or no-go decisions. Version 2 predictions are not numerically identical to version 1 because the models and reference data were retrained, so compare scores only within one version.