Predict ADMET properties to prioritize drug candidates early

What it does

This OpenClaw skill predicts key pharmacokinetic and safety properties for drug-like compounds. It covers Absorption (Caco-2 permeability, HIA), Distribution (volume of distribution, plasma protein binding, blood-brain barrier penetration), Metabolism (cytochrome P450 interactions), Excretion (clearance, half-life) and Toxicity alerts (hERG inhibition, DILI risk, Ames mutagenicity). The tool outputs structured profiles with confidence scores and actionable recommendations, helping researchers filter large compound libraries against Lipinski, Veber and other drug-likeness rules.

Who should use it

Medicinal chemists, computational biologists and drug discovery teams who need rapid in-silico triage of candidate molecules. It is particularly useful for prioritizing synthesis efforts by flagging structural liabilities—such as potential hepatotoxicity or poor oral bioavailability—before committing resources to wet-lab assays.

Setup surface

The skill requires Python and the RDKit library (pip install rdkit). Optional advanced modeling capabilities depend on deepchem and admet-x. Users interact via command-line interfaces provided by the skill scripts, such as /admet "SMILES_string" --full for single-molecule analysis or batch processing using .sdf files. Documentation references specific model types (ML/QSAR) but does not specify which underlying algorithms are active without further inspection.

Runner test plan

Future validation will include static checks for dependency integrity and injection vulnerabilities. Testing will verify that script arguments parse correctly and that output formats match documented Markdown structures. Sandbox execution will confirm that predictions generate without runtime errors for known SMILES strings. No human review or live experimental comparison has occurred; all claims rely solely on the provided technical documentation.

Risk notes

Predictions are approximate and carry inherent uncertainty. Novel chemical scaffolds often fall outside model applicability domains, leading to unreliable results. The tool explicitly warns against over-reliance: experimental validation remains mandatory for critical decisions. In vitro-in vivo gaps mean computer-predicted clearance or toxicity may not translate directly to living systems. Species differences also affect accuracy, as many models train on animal data while targeting human outcomes. Low-confidence flags require immediate manual investigation rather than automated rejection.

Source links

• Original skill documentation
• Skill listing