admet-property-prediction

1. When to use, and when not

Use this skill when you need predicted ADMET endpoint classifications for a compound. Entry points include:

• Screening a compound's predicted ADMET profile to identify which in vitro studies to prioritize.
• Flagging structural features associated with hERG channel blockade risk.
• Assessing predicted metabolic stability, CYP inhibition risk, permeability, and solubility class for a compound.
• Re-analyzing stored predictions against updated program thresholds without recomputing.

Do not use when:

• You need measured ADMET data or in vitro assay results -> bioactivity-landscape (when it covers in vitro data).
• You need in vivo pharmacokinetics, clearance, or half-life -> untooled.
• You need PBPK modelling -> untooled.
• You need drug-likeness rules (Lipinski, Veber) or structural alerts (PAINS, Brenk, aggregator filters) -> compound-property-profile.
• You need binding affinity or docking scores -> binding-mode-analysis.
• You need genetic constraint or loss-of-function intolerance -> target-genetic-evidence.

2. Preconditions

• Input: a SMILES string. Multi-fragment SMILES (salts, mixtures) are accepted: counterions are stripped, the parent molecule is selected by heavy atom count, and the compound.fragment_stripped relay fires. Unparseable SMILES exits 9 (Refusal) — a different input string is the remedy.
• RDKit: must be installed. Run oase doctor before first use. It ends with a verdict line: STOP means fix or report before running anything; PROCEED means work. Do not judge by the warning count; the verdict line grades them for you.
• No authentication needed — all operations are local.
• No network — all subcommands are offline.

3. Tool invocations

Run predict before analyze. predict computes molecular descriptors and generates endpoint predictions. analyze reads from disk, applies the admet-endpoints threshold set, and produces a verdict with per-endpoint statuses.

When phase-1 output lives in a non-default directory, pass --from to analyze. --out overrides the output directory independently. The <slug> is derived from the canonical SMILES — deterministic and filesystem-safe. Pass the canonical SMILES printed by predict to analyze.

4. Interpretation contract

Verdicts

analyze applies the admet-endpoints@1.0 threshold set and emits one of:

• developable — all five endpoints in acceptable range.
• flagged — no critical liabilities but one or more endpoints in marginal range.
• liabilities-identified — one or more endpoints in unacceptable range.

Per-endpoint statuses

Each of the five endpoints receives a status:

• acceptable — within the configured threshold range.
• marginal — borderline; warrants attention but not a liability.
• unacceptable — outside the configured threshold range; a liability.

The five endpoints: metabolic_stability, cyp_inhibition, permeability, herg_liability, solubility.

What the predictions mean

All five endpoints are rule-based predictions from molecular descriptors and SMARTS pharmacophore patterns — not trained ML models and not measurements. Each prediction record names its source publication and the rules applied.

• Metabolic stability (Gleeson 2008): liability class from LogP and aromatic ring count. Not clearance or half-life.
• CYP inhibition: risk flags per isoform (CYP2D6, CYP3A4, CYP2C9) from structural and lipophilicity rules. Not measured inhibition.
• Permeability (Egan 2000): passive permeability class from TPSA, LogP, and MW. Not Caco-2 or PAMPA measurements.
• hERG liability (Aronov 2005): pharmacophore-based flag from basic nitrogen, LogP, and aromatic ring count. Not measured IC50 or patch-clamp.
• Solubility (Delaney 2004): predicted LogS for the free form. Not salt form solubility.

Thresholds are cited by name. The values in force are in the artifact: every .analysis.json carries threshold_set, thresholds_applied, threshold_sources, and threshold_provenance. Quote values from the analysis being cited, never from prose.

Model applicability domains

Each prediction is derived from published rules calibrated on a specific chemical space. Predictions for compounds outside that space are extrapolations and should be flagged, not trusted at face value.

When a compound falls outside an endpoint's domain:

• Flag the prediction as out-of-domain in the finding. State which endpoint is affected and why (e.g. "MW 780 Da; ESOL was trained on molecules predominantly <500 Da").
• Do not treat the prediction as reliable — it is an extrapolation, not a calibrated estimate.
• If the question being answered depends on the out-of-domain endpoint, report the endpoint as untooled for that compound class, not as a confident prediction with a caveat.

Mandatory relays

Upstream relays. The tool propagates upstream compound.fragment_stripped warnings from multi-fragment SMILES handling via the phase-1 sidecar. Any upstream relay present in the analysis record must also be satisfied:

Check mandatory_relays in the .analysis.json and in the phase-1 .predict.meta.json. Every relay code present must be satisfied in the finding.

Consequence rules

• Developable: state the verdict, threshold set, and canonical SMILES. The admet.prediction_not_measurement relay fires — satisfy it before proceeding. A clean predicted profile is a reason to prioritize confirmatory studies, not to skip them.
• Flagged: state which endpoints are marginal and the threshold set. Marginal endpoints warrant attention but are not liabilities.
• Liabilities-identified: state which endpoints are unacceptable (from the analysis, not from recomputation) and the threshold set. These predict where experimental confirmation is most urgent.

Cross-disciplinary consequences

• A clean predicted ADMET profile does not substitute for measured ADMET data — every endpoint is a rule-based prediction, not an in vitro measurement.
• Predicted permeability is descriptor-based (Egan egg model), not cell-based. It is not a Caco-2 or PAMPA result.
• Predicted metabolic stability class does not predict in vivo clearance or half-life.
• Predicted solubility is for the free form. Salt form solubility is a formulation property not addressed by this prediction.
• The hERG structural flag has a documented false-negative rate. Absence of the flag is not evidence of hERG safety.
• A compound at exactly a threshold boundary is classified by the tool. Do not reclassify — the verdict under the configured threshold set is the finding.

What this section produces

Following this contract produces a Layer 1 finding in findings/ that cites the .analysis.json, satisfies every mandatory relay that fired, names the canonical SMILES and the threshold set, and never substitutes a clean prediction for a measured value. Everything the tools emitted stays under raw/admet/.

5. Failure modes and anti-fabrication guard

Do not compute these values yourself. The value is whatever the tool emitted. If the tool did not run, there is no value. If a required tool is unavailable, report the task as blocked. Do not estimate, and do not proceed on an assumed result.

Named pathologies

• Treating "no hERG flag" as hERG safety clearance. The flag has a documented false-negative rate. Absence means the compound did not match one pharmacophore pattern — not that it is safe from hERG liability. Many hERG blockers do not match this pattern.
• Conflating predicted permeability with Caco-2. The tool uses descriptor-based rules (Egan egg model), not cell-based data. A "high permeability" prediction cannot substitute for a Caco-2 result.
• Conflating stability class with in vivo half-life. Predicted metabolic stability (Gleeson 2008) does not predict clearance or half-life. "Low metabolic liability" identifies structural features, not pharmacokinetics.
• Over-reading solubility for salt forms. Predicted solubility (ESOL model) is for the free form. Salt form solubility is a formulation property not predicted by this tool.
• Treating a "developable" verdict as regulatory-ready. Every endpoint is predicted, not measured. The admet.prediction_not_measurement relay exists for this.
• Computing ADMET predictions yourself instead of using the tool. The tool pins RDKit version, descriptor set, and published rule sets. The finding must cite the tool's output, not a recalculation.