AI for Bio · Benchmark 04

TDC (ADMET + DTI)

De facto ADMET leaderboard; useful for method screening, not a basis for development decisions.

bench-04Type BEvidence L2–L3Small molecules● Active Board中文

Overview

NameTDC (ADMET + DTI) · aliases: TDC, TDC-2, PyTDC, TDC ADMET Benchmark Group, Therapeutics Data Commons
Full nameTherapeutics Data Commons
Ambiguous names (not auto-linked)None
FamilyTDC, TDC-2 and PyTDC are software and paper iterations of one platform; this entry evaluates only the ADMET and DTI-DG groups, not TDC-2 multimodal data.
  • PyTDC 0.0.1 (PyPI) ·
  • TDC paper ·
  • TDC-2 · †
  • PyTDC paper ·
DomainSmall molecules
InclusionOriginal 30
TypeB · Static dataset suite with submission leaderboards
Evidence levelL2–L3 · Scaffold split; DTI-DG reaches L3
PublisherHarvard Medical School (Department of Biomedical Informatics) (Academic lab)
First release · code release
Basis: Earliest public evidence: first PyPI upload PyTDC 0.0.1 (2020-11-13); TDC paper arXiv v1 2021-02-18; TDC-2 bioRxiv v1 2024-06-14; PyTDC paper arXiv v1 2025-05-08. source
Other milestones
  • Preprint v1: · TDC paper arXiv 2102.09548 v1 source
  • Code release: · PyTDC 0.0.1 uploaded to PyPI source
Latest update · paper revision
Basis: Revision of the PyTDC paper (arXiv 2505.05577)
PyTDC paper (arXiv 2505.05577, v1 2025-05-08) revised; latest PyPI release 1.1.15 on 2025-03-31, last repo commit 2025-04-18. source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe maintainers have announced no further release or ADMET group revision.
Status● Active · Active; ADMET group data and rules not revised since 2021
ScaleADMET group: 22 datasets (475–13,130 records); full library 66 datasets across 22 tasks in the original paper.
  • ADMET group datasets: 22 datasets (ADMET Benchmark Group)
  • Records per dataset: 475–13,130 records (ADMET Benchmark Group)
  • DTI-DG group datasets: 1 datasets (BindingDB_Patent)
  • Datasets in original library: 66 datasets (TDC paper (2021))
  • Official benchmark groups: 8 groups
Wet-lab refreshStatic data · None for the ADMET group; data are reused from ChEMBL and public deposits.
Own wet lab needed to evaluate a new modelYes, to claim real generalization
Models reported4 counted (board 3, self 1, unconfirmed 2)
Unverified fields† family.members[2], ranking.entries[0], ranking.entries[1], ranking.entries[2], ranking.entries[3], ranking.entries[4], models_reported[2], models_reported[4], models_reported[5]
As of2026-11-16

What it measures

  • Task: ADMET endpoint prediction under a scaffold split, plus one drug-target interaction domain-generalization task.
  • Truth: Labels compiled from literature and databases, not measured under uniform conditions.
  • Primary metric: Per task: MAE, AUROC, AUPRC or Spearman
  • Secondary metrics: Mean and SD over 5 seeds; DTI-DG: Pearson correlation
  • Version traps: TDC datasets differ from official groups: Davis and KIBA are not official groups; boards lack versions and dates, and loaders may yield different molecule counts.

Current ranking

Board

Metricper-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
Eval setTDC ADMET Benchmark Group, 22 datasets
Snapshot
Boardhttps://tdcommons.ai/benchmark/admet_group/overview/
Retrieved2026-11-16

Caveat. Self-reported, no independent re-run or sanity check, no versions or dates; top ranks often within 1-2 SD; boosting-plus-fingerprint baselines beat most pretrained models.

No headline entry: None of the 22 official boards dates its entries, so every score is withheld and no entry can serve as headline.

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
—MiniMolWithheld: Board entries carry no date, so under the honesty rule no score or rank is shown.per-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
TDC ADMET Benchmark Group, 22 datasets
Board
No sourced date†
source Ties with MapLight + GNN for the most first places across the 22 task boards.
—MapLight + GNNWithheld: Board entries carry no date, so under the honesty rule no score or rank is shown.per-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
TDC ADMET Benchmark Group, 22 datasets
Board
No sourced date†
source Not a site model; ties with MiniMol for most first places; remaining firsts go to several boosting and ensemble methods.
—admet_ai_v2Withheld: Board entries carry no date, so under the honesty rule no score or rank is shown.per-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
TDC ADMET Benchmark Group, 22 datasets
Board
No sourced date†
source The only ADMET-AI-labelled submission on the 22 boards, Caco2 only and last; value looks like a unit error; the paper's best-average-rank claim has no live counterpart.
—Chemprop-RDKit / ChempropWithheld: Board entries carry no date, so under the honesty rule no score or rank is shown.per-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
TDC ADMET Benchmark Group, 22 datasets
Board
No sourced date†
source Board rows are original Chemprop / Chemprop-RDKit submissions, not v2; no first place on any task; pairing to the v2 page is a proxy.
—MolEWithheld: No row for this model on the live boards and no entry date, so no score is shown.per-task: MAE / AUROC / AUPRC / Spearman (5-seed mean ± sd), scaffold split
TDC ADMET Benchmark Group, 22 datasets
Board
No sourced date†
source No MolE row on any of the 22 official boards in a 2026-11-16 scan; the "10 of 22 first places" claim belongs to its 2024 paper only.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 4. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Official leaderboard or results (3)

  • ADMET-AI Board · Submitted to the TDC ADMET group board (only one Caco2 row on the live board) · evidence: link
  • MiniMol Board · Self paper plus official board entries · evidence: link
  • MolE Board · Paper compares with a 2023-09 board snapshot; no row on the live boards †partly verified · evidence: link

Developer report, official protocol (1)

  • TxGemma Self-reported · 66 TDC tasks, recommended splits · evidence: link

Unconfirmed (2, not counted)

  • Chemprop v2 Unconfirmed source · Unversioned board entries, linked indirectly via ADMET-AI low confidence †disputed · evidence: link
  • Enchant v1/v2 Unconfirmed source · Obach human PK result not shown to use the TDC split; three site surfaces unified to unconfirmed Resolution: The developer post never names the benchmark, the official split or the protocol; the comparison is a hyperlink to a baseline page. The published surfaces disagree three ways (zh matrix 'self', zh model page 'unconfirmed', en pairs.json 'self'); all three are unified to unconfirmed. Not an established pairing. †disputed · evidence: link

Caveats

  1. Submissions are self-reported with no independent re-run or sanity check; scaffold splits do not stop near-analog leakage, and gradient-boosting baselines are strong.
  2. Labels are compiled across sources; small sets such as hERG and DILI have low power, and binary thresholds come from literature.
  3. A user report notes PPBR_AZ molecule counts differ by loader and test duplicates carry conflicting labels (not an independent audit).
  4. Models pretrained on ChEMBL or ZINC may have seen test molecules with no official deduplication; rules on external data and ensembles are weak.
  5. No verifiable evidence links board gains to experimental hit rates.

How to use it

Install PyTDC, load data with admet_group get and score with evaluate_many over 5 seeds. Report seed SD, SMILES standardisation and deduplication, any external pretraining or precomputed features, and a gradient-boosting baseline; generalisation claims need time-split own data or prospective blind tests.

Pairs well with: Polaris / ASAP / OpenADMET · MoleculeNet (+OGB) · FS-Mol

Sources

  1. TDC paper (arXiv 2102.09548) (preprint)
  2. ADMET group overview and leaderboards (leaderboard)
  3. Official benchmark group overview (official site)
  4. DTI-DG group overview (official site)
  5. PyTDC paper (arXiv 2505.05577) (preprint)
  6. PyTDC on PyPI (package registry)
  7. TDC repository issue #377 (PPBR_AZ data issue) (repository)
  8. MapLight preprint (arXiv 2310.00174) (third-party paper)
  9. ADMET-AI paper (Bioinformatics 2024) (journal)

Corrections

6 corrections
FieldWasNowBasisFound
first_release2021-02-182020-11-13source2026-11-16
first_release_basisarXiv v1 of TDC paper (PyPI PyTDC 0.0.1 appeared earlier, 2020-11-13)first public code release: PyTDC 0.0.1 on PyPI (2020-11-13); TDC paper arXiv v1 2021-02-18; TDC-2 bioRxiv v1 2024-06-14; PyTDC paper arXiv v1 2025-05-08.source2026-11-16
members[{"member": "PyTDC first PyPI", "date": "2020-11-13"}, {"member": "TDC paper (arXiv)", "date": "2021-02-18"}, {"member": "PyTDC (arXiv v1)", "date": "2025-05-08"}][{"member": "PyTDC first PyPI release 0.0.1", "date": "2020-11-13"}, {"member": "TDC paper (arXiv v1)", "date": "2021-02-18"}, {"member": "TDC-2 (bioRxiv v1)", "date": "2024-06-14"}, {"member": "PyTDC paper (arXiv v1)", "date": "2025-05-08"}]source2026-11-16
ranking.entries[model-mole]MolE '10 of 22 first places' presented as a board standingNo MolE row on any of the 22 official boards (2026-11-16 scan); claim attributed to its 2024 paper onlysource2026-11-16
ranking.entries[model-admet-ai]ADMET-AI 'best average rank over the 22 datasets' presented as a board standingOnly one ADMET-AI-labelled board submission (Caco2, last place, value likely a unit error); the average-rank claim has no live counterpartsource2026-11-16
models_reported.model-enchant-v1-v2self-reported (zh matrix, en pairs.json)disputed-or-unconfirmedsource2026-11-16

Part of the AI for Bio benchmarks survey · all 44 benchmarks · open in the explorer. Record as of 2026-11-16.