Chemistry and compounds · What you can do here
Predict ADMET properties of a compound
Standardise small-molecule structures and predict 52 ADMET and toxicity endpoints using ADMET-AI and Chemprop D-MPNN ensembles, benchmarked against reference distributions of approved drugs.
What this involves
Predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) endpoints allows you to detect pharmacokinetic liabilities and safety risks before advancing compounds to synthesis or testing. A complete assessment pairs deterministic physicochemical descriptors and rule criteria with neural property predictions across critical parameters, including cardiotoxicity via hERG, blood–brain barrier permeability, clearance, and clinical toxicity.
Reliable computation begins with strict structure curation. Undefined stereochemistry, unstripped salt counterions, and protonation states alter molecular weight, calculated logP, and rule compliance if not corrected. Furthermore, regression outputs and classification probabilities must be read alongside structural alert catalogues and approved drug distributions to interpret liabilities accurately.
The platform research agent runs these standardisation routines, descriptor calculators, and machine-learning ensembles on its own machine and returns structured data and depictions directly, requiring no local software installation.
How it goes
- 1Structure verification and name resolutionResolve chemical names and synonyms against ChEMBL and PubChem using search_compounds, verifying the resulting 27-character InChIKey to confirm connectivity and stereochemical integrity.
- 2Fragment stripping and charge neutralisationStandardise the molecule using RDKit's LargestFragmentChooser to remove counterions and Uncharger to neutralise charges, yielding a canonical SMILES for downstream tools.
- 3Physicochemical descriptor and alert evaluationCalculate Wildman–Crippen cLogP, TPSA, Fsp3, and ring counts with RDKit or the Molecular Property Calculator, then screen for Lipinski Ro5, Veber, PAINS, and Brenk rule violations.
- 4ADMET and toxicity endpoint predictionPredict absorption, metabolism, and safety endpoints using ADMET-AI or the Chemprop D-MPNN toxicity panel, alongside IBM Research MAMMAL for blood–brain barrier permeability.
- 5Reference benchmarkingEvaluate each predicted endpoint against Therapeutics Data Commons training distributions and approved drug benchmarks to contextualise liability severity.
Start it here
Ask in plain language. The agent works on a Linux machine of its own — it installs and runs what the job needs, keeps every file, and reads the results back with the numbers explained. Nothing to set up, and a long run continues after you close the page.
What to watch for
- Retaining salt counterions such as mesylate or bis-TFA adds 36 to 230 Da and skews molecular weight, cLogP, and rule evaluations.
- Canonicalising tautomers to report a structure rather than to compare molecules misrepresents which tautomeric form exists in solution.
- Assuming an unannotated SMILES is racemic or flat overlooks that missing stereochemical tags denote undefined stereochemistry.
- Mistaking calculated Wildman–Crippen MolLogP for experimental logD7.4 creates apparent discrepancies of two to three units for basic compounds.
- Treating PAINS filter alerts as definitive evidence of assay interference overlooks that roughly 5% of approved drugs trigger an alert.
What it runs
Models
- ADMET-AIMITPredicts 52 small-molecule ADMET endpoints simultaneously using Chemprop D-MPNN ensembles.
- Chemprop D-MPNN — toxicity panelMITPredicts small-molecule toxicity endpoints and structural alerts from chemical structures.
- MAMMALApache-2.0A biomedical foundation model predicting binding affinity, blood–brain barrier permeability, and solubility.
- Molecular Property CalculatorMITComputes physicochemical descriptors and druglikeness rule compliance for small molecules from SMILES strings.
Questions
Do I need to install anything to predict admet properties of a compound?
No. The research agent does this work on a Linux machine of its own — it installs what the job needs, keeps the files, and reads the results back to you. Long runs continue if you close the page.
Which predictive models are used for ADMET and toxicity profiling?
The platform runs ADMET-AI, a Chemprop directed message-passing neural network (D-MPNN) ensemble trained on Therapeutics Data Commons datasets across 52 endpoints. It also offers the Chemprop D-MPNN toxicity panel for specific liabilities like Ames mutagenicity and DILI, as well as IBM Research's MAMMAL for blood–brain barrier permeability.
Why does calculated logP differ from experimental literature values?
Descriptors.MolLogP employs the Wildman–Crippen atom-contribution model for neutral compounds, which carries an RMSE of around 1 log unit. Additionally, experimental publications typically report logD7.4, which diverges significantly from logP for ionisable compounds at physiological pH.
Are compounds rejected automatically if they violate Lipinski's Rule of Five?
No. The Rule of Five describes orally bioavailable drugs and allows one violation, but it does not apply to natural products, macrocycles, peptides, or actively transported molecules. Violations are reported without discarding the compound.
How are structural alerts like PAINS interpreted?
PAINS and Brenk alerts are used as triage flags to inspect specific substructures rather than terminal verdicts. When an alert fires, the specific pattern is identified while keeping borderline molecules visible.
Written from the method the research agent follows for this work and the documentation of the tools listed above. Statements citing a figure that does not appear there are removed before publication. Corollary is a non-commercial research demonstration; results are not for clinical use.