Free ADMET prediction in the browser: rules + ML, no setup

Docking tells you what might bind. ADMET tells you whether a hit is worth pursuing — is it soluble, permeable, metabolically stable, and not obviously toxic? Catching a doomed compound early saves real wet-lab money. Here's how to get a fast, free ADMET read on any molecule in the browser, combining classic medicinal-chemistry rules with a machine-learning model.

Two layers: rules and ML

MolHub shows both on every molecule's detail page, because they answer different questions.

Rule-based filters (instant, deterministic)

• Lipinski's rule of five — MW ≤ 500, LogP ≤ 5, H-bond donors ≤ 5, acceptors ≤ 10. Two or more violations flags likely poor oral absorption.
• Veber — rotatable bonds ≤ 10 and TPSA ≤ 140 for oral bioavailability.
• Lead-likeness and crude GI-absorption / blood-brain-barrier heuristics.
• Structural alerts — PAINS (pan-assay interference compounds that produce false hits) and Brenk (reactive / unstable groups). A PAINS hit is a strong "be careful" signal.

ADMET-AI predictions (machine learning)

On top of the rules, MolHub runs ADMET-AI — an open graph neural network trained on 41 Therapeutics Data Commons datasets — to predict ~100 endpoints, surfaced as readable highlights:

• Aqueous solubility, lipophilicity, Caco-2 permeability, human intestinal absorption
• Blood-brain-barrier penetration, oral bioavailability, half-life
• CYP3A4 / CYP2D6 inhibition (metabolism / drug-drug interactions)
• hERG blockade (cardiotoxicity), Ames mutagenicity, DILI (liver injury), LD50

How to use it

Open any molecule from the library — the ADMET panel sits beside the physicochemical properties, with a single verdict badge (drug-like, moderate, or caution) so you can rank a hit list at a glance. After a batch dock, this is how you decide which top-scoring poses are actually worth ordering.

Don't over-trust it

ADMET predictions — rules and ML alike — are estimates for triage, not measurements. A "drug-like" verdict doesn't guarantee a clean assay, and a flag doesn't doom a molecule. Use them to prioritize: spend your experiments on compounds that pass both docking enrichment and an ADMET sanity check, and confirm everything in the lab.

Why it's free here

The rule filters are RDKit; ADMET-AI is open-source and runs CPU-only as an isolated service, so there's no GPU bill and no per-seat licence. That's the same reason the rest of MolHub — docking, AlphaFold targets, search — is free for academics.