henhao
Sign in
BioinformaticsBioinformatics
Peer-reviewed journal article

ADMET-AI: A machine learning ADMET platform for evaluation of large-scale chemical libraries

Kyle SwansoniD, Parker Walther, Jeremy LeitziD, Souhrid MukherjeeiD, Joseph C. WuiD, Rabindra V. ShivnaraineiD, James ZouiD

Abstract

SummaryThe emergence of large chemical repositories and combinatorial chemical spaces, coupled with high-throughput docking and generative AI, have greatly expanded the chemical diversity of small molecules for drug discovery. Selecting compounds for experimental validation requires filtering these molecules based on favourable druglike properties, such as Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET). We developed ADMET-AI, a machine learning platform that provides fast and accurate ADMET predictions both as a website and as a Python package. ADMET-AI has the highest average rank on the TDC ADMET Benchmark Group leaderboard, and it is currently the fastest web-based ADMET predictor, with a 45% reduction in time compared to the next fastest ADMET web server. ADMET-AI can also be run locally with predictions for one million molecules taking just 3.1 hours. Availability and ImplementationThe ADMET-AI platform is freely available both as a web server at admet.ai.greenstonebio.com and as an open-source Python package for local batch prediction at github.com/swansonk14/admet_ai (also archived on Zenodo at doi.org/10.5281/zenodo.10372930). All data and models are archived on Zenodo at doi.org/10.5281/zenodo.10372418.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
393
FWCI
82.83
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (28 Dec 2023) · CC BY 4.0

ADMET-AI, a Chemprop-RDKit graph neural network trained on 41 TDC datasets, has the best average leaderboard rank and scores one million molecules in 3.1 hours.

Docking and generative AI produce very large hit lists, many with poor druglike properties, so fast ADMET filtering is needed. Per the authors, most existing tools lack local batch prediction and context from approved drugs; ADMET-AI adds both and runs faster.

Key results

  1. ShownThe ensemble of Chemprop-RDKit models has the best average rank among all models evaluated on every dataset of the TDC ADMET leaderboard, though other models can be better on individual properties.

    Best average rank on 22 leaderboard datasets (9 regression, 13 classification), leaderboard as of October 4, 2023; scaffold-split test sets, five models trained per dataset with different seeds. · Abstract; Results; TDC Leaderboard

  2. ShownSingle-task models perform well on many of the 41 TDC datasets, and the two multi-task models (regression, classification) reach very similar performance while predicting much faster, so they are the ones deployed.

    R above 0.6 for five of the ten regression datasets; AUROC above 0.85 for 20 of the 31 classification datasets; 80%/10%/10% random splits with five seeds. Per-dataset values are in a supplementary table. · Results; TDC Single-Task

  3. ShownThe ADMET-AI web server was the fastest of the seven web servers compared at every batch size tested, with a 45% time reduction over the next best server at 1,000 molecules.

    45% less time than the next best server for 1,000 molecules; batches of 1, 10, 100 and 1,000 DrugBank molecules; median of three trials; compared with six servers (SwissADME, ADMETLab 2.0, pkCSM, vNN-ADMET, ADMETboost, PreADMET). · Abstract; Comparison to Alternatives: Speed; Figure legends

  4. ShownLocal batch prediction cuts the time for one million molecules sharply versus the web server, and even a modest CPU-only setup finishes in a few hours.

    28.1 hours (web server, two CPU cores, no GPU) vs 3.1 hours (32 CPU cores plus GPU), a further 89% reduction; 8 CPU cores without GPU takes five hours. Web server time is 1,000 molecule time x 1,000. · Abstract; Comparison to Alternatives: Speed; Figure legends

  5. ClaimedEach prediction is returned with its percentile against approved drugs, which the authors say makes values easier to interpret; the reference can be restricted by ATC code. No user study or accuracy test of this benefit is reported.

    Reference set of 2,579 approved drugs from DrugBank v5.1.10; ATC code counts: 14 at level 1, 86 at level 2, 223 at level 3, 625 at level 4. · DrugBank Reference; User Experience; Figure legends

Methods, data and reproducibility

A Chemprop message-passing graph neural network augmented with 200 RDKit descriptors was trained on each TDC dataset using five splits and ensembled. Single-task and two multi-task models were compared, and web and local runtimes were benchmarked.

  • TDC ADMET single-task datasets (10 regression, 31 classification), TDC v0.4.1; hERG_Karim, herg_central and ToxCast excluded · n = 41 datasets
  • TDC ADMET Benchmark Group leaderboard datasets (9 regression, 13 classification), scaffold split with TDC test set · n = 22 datasets
  • TDC multi-task regression dataset merged by SMILES across the 10 regression datasets · n = 22,100 unique compounds
  • TDC multi-task classification dataset merged by SMILES across the 31 classification datasets · n = 35,774 unique compounds
  • DrugBank v5.1.10 approved drugs used as percentile reference set and as speed benchmark input · n = 2,579 drugs
  • Archive of all data and models for ADMET-AI · 10.5281/zenodo.10372418
Methods & tools
Chemprop-RDKit (Chemprop v1.6.1), RDKit v2023.3.3, Therapeutics Data Commons v0.4.1, Flask v2.3.2, DrugBank v5.1.10, SwissADME, ADMETLab 2.0, pkCSM, vNN-ADMET, ADMETboost, PreADMET
Code
http://github.com/swansonk14/admet_aihttp://doi.org/10.5281/zenodo.10372930
Data & accessions
http://doi.org/10.5281/zenodo.10372418https://docs.google.com/spreadsheets/d/1bh9FEHqhbfHKF-Nxjad0Cpy2p5ztH__p0pijB43yc94/edit?usp=sharinghttps://docs.google.com/spreadsheets/d/1uKjqHMrAjnB8X-53cXgclb5jwXTortFeOYvEkCmNsSY/edit?usp=sharinghttps://docs.google.com/spreadsheets/d/1_bvDO73MvYb7wl7_kH6Vagonyi-EJ6H-u8KpwQQ9Srg/edit?usp=sharing
Software & environment
Chemprop v1.6.1, RDKit v2023.3.3, TDC v0.4.1, Flask v2.3.2, DrugBank v5.1.10.
Assay conditions & compute
Local runs used 8 or 32 CPU cores with or without a GPU; the web server uses two CPU cores and no GPU. One million molecules took 3.1 hours with 32 cores plus GPU.
Reproducibility gaps
GPU model not specified; supplementary tables are hosted as Google Sheets links rather than in a permanent archive.
Limitations (3)
  • Random 80/10/10 splits in the single-task and multi-task benchmarks may give overly optimistic test performance, because molecular scaffolds are shared between training and test sets. · stated by the authors
  • Some timing baselines are extrapolated: the web server one million molecule time is 1,000 molecule time x 1,000, and pkCSM and SwissADME 1,000 molecule times are 100 molecule time x 10, so reported reductions partly rest on extrapolation. · noted in this summary
  • Leaderboard rank is a snapshot as of October 4, 2023 on retrospective benchmarks, and no prospective experimental validation of predictions is reported in the text. · noted in this summary

Who should read:Read this if you need a fast, open-source ADMET predictor for screening docking or generative AI libraries, with percentile context against approved drugs.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISNair '26Varghese '26Padi '23Boral '22Stamenov '26Luo '26
76% semantic matchBioinformatics
Divya N. Nair, Rohit Singh Yadav, Prajakta Mate Jondhale+10 more·bioRxiv·Jun 13, 2026