OpenADMET

Most compounds that fail in drug development fail on ADMET — absorption, distribution, metabolism, excretion, and toxicity — rather than on target potency. The predictive models that pharmaceutical companies use to triage those liabilities are trained on proprietary assay data and stay behind company walls, which means academic groups and small biotechs have no comparable tool and no way to audit the ones that exist. OpenADMET is an open-science consortium built to close that gap by generating open ADMET data and releasing open models trained on it.

The result is not a single network but a growing family of endpoint-specialized models that all fine-tune one shared backbone, CheMeleon. The suite currently spans cytochrome P450 inhibition (a multitask model over CYP1A2, CYP2D6, CYP3A4, and CYP2C9), pregnane X receptor activation, microsomal intrinsic clearance in human, rat, and mouse liver microsomes, a multitask permeability model covering Caco-2 apparent permeability in both directions along with LogD and human and mouse plasma protein binding, and hERG channel inhibition. The inaugural CYP and PXR models were released in December 2025, with the clearance model following in March 2026.

The consortium is hosted by the Open Molecular Software Foundation and combines a software team there with high-throughput cell-based assay work at Octant and structural and pharmacology groups at UCSF, under funding from ARPA-H, the Gates Foundation, and the Astera Institute. Its framing of the problem — treating the set of proteins that cause ADMET failures as an "avoid-ome" to be systematically characterized — was published in Nature Communications in 2026.

#Key Features

• One backbone, many endpoints: Each released model fine-tunes the same pretrained molecular encoder on a different assay, so the suite grows by adding endpoints rather than by redesigning architectures.
• Multitask heads where assays correlate: The CYP model predicts four isoforms jointly and the clearance model predicts three species jointly, letting sparse endpoints borrow signal from better-populated ones.
• Fully open training data: Every model is trained on public ChEMBL records plus curated data released from the consortium's own blind challenges, so training sets can be inspected, re-split, and rebuilt.
• Reproducible training recipes: Weights ship on HuggingFace under Apache 2.0 alongside the openadmet-models framework and conda and Docker environments, with inference driven by a single openadmet predict command over a SMILES column.
• Honest applicability-domain reporting: Model cards and release notes publish cluster-split as well as random-split performance, making the drop-off outside familiar chemistry explicit rather than hidden.

#Technical Details

CheMeleon is a roughly 12.9M-parameter directed message-passing neural network, built on Chemprop and pretrained on one million PubChem molecules to regress computed Mordred descriptors; OpenADMET fine-tunes it per endpoint through its Anvil training framework. Training sets are small by foundation-model standards and reflect what public databases actually contain: about 4,800 curated datapoints for CYP3A4 and roughly 7,600 across the four CYP isoforms, about 600 for PXR after pChEMBL filtering, and ChEMBL 37 extractions for hERG. The clearance model supplements ChEMBL with data from the ASAP-Polaris antiviral and ExpansionRx blind challenges and predicts log₁₀ CLint, which must be back-transformed for use.

Performance is reported candidly as a baseline. The CYP and PXR models reach roughly 0.6 R² under random splits but fall to around 0.1 under cluster splits, and external benchmark sets share Tanimoto similarities below 0.4 with the training chemistry — the applicability-domain problem, not architecture, is the binding constraint. A systematic ablation across featurizations, data prescaling, pretraining corpora, and message-passing schemes found no change that improved results significantly, and those negative results were published in full.

#Applications

The suite supports early triage in medicinal chemistry: ranking synthesized or virtual compounds for drug–drug interaction risk through CYP inhibition and PXR induction, cardiac liability through hERG, metabolic stability through microsomal clearance, and exposure through permeability, LogD, and plasma protein binding — all before committing compounds to DMPK assays. Because weights and recipes are open, the encoders also serve as starting points for fine-tuning on proprietary in-house measurements, and as reproducible reference baselines for groups benchmarking new ADMET methods.

#Impact

OpenADMET's contribution is as much infrastructural as it is architectural: open weights, open training data, open training code, and openly published negative results in a corner of drug discovery where none of those were previously available. Its blind challenges, run with Polaris and industry partners on antiviral, ExpansionRx, PXR, and CYP datasets, have drawn dozens of teams and produced benchmark sets that outlive the competitions themselves. The models should be read as what their authors call them — baselines rather than production-grade predictors — and no peer-reviewed publication yet describes the models specifically, but they establish a public floor that subsequent work, inside and outside the consortium, can be measured against.