ADMET Prediction
Potency gets a compound into a programme; ADMET decides whether it survives one. ADMET-AI is a
Chemprop-RDKit graph network trained on 41 Therapeutics Data Commons datasets, tops the TDC ADMET
leaderboard, and runs thousands of molecules a minute on a CPU. This skill is about reading its
output as a developability verdict rather than a wall of numbers.
Tool: ADMET-AI 2.0.1, MIT, pip install admet-ai
(requires Python 3.11+). Weights download on first use. No GPU needed.
Checked against: PyPI 2.0.1, February 2026.
Read references/running-admet-ai.md before your first run,
references/endpoints.md to know which endpoints actually stop
programmes, and
references/interpreting-predictions.md before acting on
a number — that one is judgement, not syntax.
The two scripts
| Script | Answers |
|---|
admet_batch.py | How do I feed a library in without wasting the run? |
admet_report.py | Which of these compounds has a liability worth acting on? |
Rank within a series; do not trust absolute values
This is the thing to get right. A public model has systematic offsets against your assay —
different protocol, different lab, different chemistry. Within a congeneric series those offsets
are largely shared, so the ordering survives even where the values do not.
Use predictions to decide which twenty of these hundred to make and assay. Do not use them to
decide whether this compound will pass. A predicted hERG of 0.7 versus 0.3 within a series is a
real signal; 0.7 in absolute terms is not a measurement.
The percentile column is the point
ADMET-AI reports every prediction against the distribution of approved drugs in DrugBank, in
<endpoint>_drugbank_approved_percentile. It is the most useful thing the tool adds over a bare
model and the column most often ignored.
"Predicted clearance 12" is hard to act on. "More extreme than 92% of approved drugs" prompts the
right question: drugs exist out here, but not many — what is the argument that this one works?
Flagging a set
python skills/admet-prediction/scripts/admet_report.py report --csv predictions.csv
smiles liabilities flagged out_of_domain
c1ccccc1CCNC(=O)c1ccc(Cl)cc1 5 hERG|DILI|Solubility_AqSolDB|Lipophilicity|Half_Life
CCO 0 molecular_weight=46.07 outside [150, 700]
Each endpoint is flagged against its own direction — high solubility is good, high clearance is
bad, high hERG is very bad — so a single summed "score" over the columns would be meaningless.
admet_report.py endpoints prints the full registry with thresholds; they are this skill's
conventions and are meant to be argued with.
Note the second row. Ethanol is flagged as out of domain, not clean. ADMET-AI reports no
applicability domain, so a prediction on anything unlike its training data arrives with the same
confident four decimal places as a reliable one.
BBB penetration has no liability direction — essential for a CNS target, a liability
everywhere else. The script leaves it unflagged rather than guessing your programme.
Preparing input
python skills/admet-prediction/scripts/admet_batch.py prepare --smiles library.smi --out-dir admet_in
# 3 input, 2 unique (1 duplicates collapsed), 1 chunk(s)
# warning: 1 SMILES contain `.` -- a salt, mixture, or counterion.
admet_predict --smiles_path admet_in/chunk_0000.csv --save_path admet_in/chunk_0000_pred.csv --smiles_column smiles
Three things this prevents. ADMET-AI needs a CSV with a header — a bare .smi list silently
loses its first molecule. Duplicates cost twice and add nothing, since the model is
deterministic. And a . in a SMILES is a salt or mixture: the model predicts on the string as
given, so the answer describes the wrong species. Desalt with datamol first.
Four ways predictions mislead
- Classification outputs are probabilities, not classes. hERG at 0.55 is a coin flip. Move
the threshold with the cost of being wrong — screen hERG at 0.3, not 0.5.
- Endpoints are not equally trustworthy. Lipophilicity and solubility are well predicted;
DILI, clearance, and Vd are barely better than a coin flip. The leaderboard's average rank
hides that.
- Real liabilities are simply absent. Time-dependent CYP inhibition, reactive metabolites,
transporters beyond Pgp, phospholipidosis, mitochondrial toxicity — none are covered.
- Over-filtering early is the expensive mistake. Most ADMET liabilities are fixable by
medicinal chemistry; poor potency and a wrong target are not. Filtering a primary screen on
predicted DILI discards real chemistry on the basis of noise.
When to stop using this
If your project has more than a few hundred measured compounds for an endpoint, train a Chemprop
model on your own data — the applicability domain finally matches your chemistry, and it will beat
any public model on it. For time-dependent CYP inhibition, transporters, or reactive metabolites,
there is no model; run the assay.
Composing with the rest of the bundle
medchem → before: structural alerts and PAINS cost nothing and catch much of this first.rdkit / datamol → before: desalt and standardise, or you predict on the wrong species.chemical-space → before: this is a good filter stage in an ultra-large cascade.pkpd-translation → after: predicted clearance, half-life, and PPB become dose projections.deepchem / pytdc → instead: when you want to train on your own data rather than use a
ready-made model.
Reporting results honestly
Give the percentile beside the value. Name the thresholds used and say they are conventions. State
whether the molecule sits inside a drug-like property window. Never write "this compound is a hERG
blocker" from a prediction — write "predicted hERG 0.82, above the 90th percentile of approved
drugs; assay before progressing". Say what the predictions decided: they choose what to assay,
they do not replace it.