Overview
| Name | ADMET-AI · aliases: ADMET-AI, admet_ai |
|---|---|
| Full name | The name is itself an acronym compound; the paper gives no expanded form |
| Ambiguous names (not auto-linked) | None |
| Family | Not a family entry. A single platform; v1 and v2 are version axes inside one record |
| Organization | Stanford University (computer science), Greenstone Biosciences |
| Organization type | Mixed |
| Country | United States |
| Model class | Property prediction (multitask ADMET ensemble) |
| Domain | Small molecules |
| First release | · paper revision Basis: Preprint posted (registry record) Note: The site recorded 2024, the journal year source |
| Other milestones | |
| Latest update | · new version Basis: v2.0.1 patch (latest tag) v2.0.0 (2026-02-17) retrained all models from scratch so predictions differ from the v1 web service; v2.0.1 fixed MPS/CPU handling and dropped a dataset-library dependency. Per the package README the web service still runs v1 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Major version and patch released within 12 months |
| Openness | Code: Open Weights: Open Code on GitHub with data and models archived in a public repository plus a free web service; licence text not retrieved this round |
| Benchmarks reported | 1 counted (board 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † release_milestones.code_release, open_status.code_license, open_status.weights_license, background.funding |
| As of | 2026-11-16 |
What it does
Returns predictions for 41 absorption, distribution, metabolism, excretion and toxicity endpoints per molecule and compares them with the distribution of approved drugs, aimed at large-scale library screening (the paper reports a few hours for a million molecules locally). The model itself is not new: it is a descriptor-augmented multitask ensemble of the previous Chemprop generation, and its value is engineering and coverage; unlike pretrained foundation models such as MolE and MiniMol it does no large-scale pretraining.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 04 TDC (ADMET + DTI) | Official leaderboard or results | — | Board | Submitted to the TDC ADMET group board (only one Caco2 row on the live board) · evidence |
Other evaluations (outside the 44)
- Own throughput measurement (Own task set) Self-reported: The paper reports the time to predict a million molecules locally; an engineering figure, not an accuracy evaluation
- Comparison with approved-drug distributions (Own task set) Self-reported: Places predictions on the distribution of approved drugs as percentiles; an interpretive output
Developers report: a free web service plus open local inference covering 41 ADMET endpoints at once, with the best average rank on the public ADMET group board as it stood at publication.
Independent evidence: No independent evidence found. No third-party evaluation of the platform was found, and the current board holds no entry supporting the best-average-rank claim
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | 41 ADMET datasets (10 regression, 31 classification) from a public therapeutics dataset library, 22 of which form its official ADMET leaderboard; five models per dataset as an ensemble source |
| b. Private or licensed data | No | source |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The authors come from a US university computer science department and also carry the affiliation of a biotechnology company in the university's research park; financial-data archives record the company as founded in 2021, working on induced pluripotent stem cells and organoids with machine learning for discovery and safety prediction, and claiming a large stem-cell biobank.
- Funding: Public archives list three venture investors in the company; academic funding is not itemised in the paper.
- Products and partnerships: The product is a free web service plus open local inference; on 2026-06-19 the company announced a collaboration with a semiconductor company to process its stem-cell biobank data on that company's compute (press release wording).
- Related models: Chemprop v2, MiniMol, MolE
Reading notes
- "Best average rank" aggregates 22 datasets as of publication and does not mean first on each endpoint; no matching entry exists on the current board, so any citation needs a date and source qualifier.
- The only submission on the current board labelled with this model covers one task, sits last and carries a suspected unit error; it cannot be read as its present standing.
- It shows a well-tuned graph-network ensemble with descriptors remains strong on ADMET, so pretrained foundation models do not hold a stable edge.
- From v2.0.0 all models were retrained while the web service still runs v1, so the two give different predictions and comparisons must state the version.
Caveats
- The paper's board standing does not match the board as of 2026-11-16; the published statement is corrected.
- The public ADMET board uses public test sets, so overfitting and differing submission protocols matter.
- No wet lab and no private data; all training data is public.
- No release-notes text was retrieved for v2.0.0; the version description comes from the package documentation.
Sources
- Journal full text (public archive) (journal)
- Journal paper 2024 (journal)
- Release list (repository)
- Package registry page (package registry)
- Public ADMET group board overview (checked 2026-11-16) (leaderboard)
Corrections
3 corrections
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| benchmarks_reported | Best average rank across 22 datasets | The claim holds only as of the paper; the only submission labelled with this model on the current board is on a single task and in last place (suspected unit error); no matching entry | source | 2026-11-16 |
| first_release | 2024 | 2023-12-28 (preprint posted) | source | 2026-11-16 |
| latest_update | 2026-02-17(v2.0.0) | 2026-02-22 (v2.0.1, latest tag), with the full v2 retraining also recorded | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.