AI for Bio · Benchmark 03

MoleculeNet (+OGB)

Historical default suite for molecular representation learning: 17 public datasets, for literature alignment, not drug-discovery claims.

bench-03Type BEvidence L2–L3Small molecules○ Frozen Not comparable中文

Overview

NameMoleculeNet (+OGB) · aliases: MoleculeNet, MolNet, molnet, ogbg-molhiv, ogbg-molpcba
Full nameMoleculeNet: a benchmark suite for molecular machine learning (with OGB-derived molecular sets)
Ambiguous names (not auto-linked)OGB
FamilyNot a family entry. Single dataset suite; OGB molecular sets are derived splits recorded as a version trap, not a family.
DomainSmall molecules
InclusionOriginal 30
TypeB · Static dataset suite with no official leaderboard
Evidence levelL2–L3 · Random split L2; scaffold split (weak OOD) L3
PublisherStanford University (academic group), DeepChem open-source project (Academic lab)
First release · paper v1
Basis: arXiv 1703.00564 v1 source
Other milestones
  • Preprint v1: · arXiv 1703.00564 v1 source
  • Journal online: † · Chem. Sci. journal version (year only; exact online date not retrieved) source
Latest update · paper revision
Basis: Last-updated date of v3 in the arXiv API
arXiv v3 posted (final paper revision); Chem. Sci. version published 2018; no versioned data snapshot since.
Note: DeepChem loaders are still maintained; library upkeep, not a benchmark change source
The most recent sourced event, not a claim that nothing changed since.
Next expectedThe datasets are frozen and no new version has been announced.
Status○ Frozen · Effectively frozen since 2018; still widely reused for literature alignment
Scale17 datasets in the paper, from FreeSolv (642) to PCBA (437,929 compounds); QM9 has 133,885 molecules.
  • Datasets: 17 (paper (Chem. Sci. 2018))
  • Smallest dataset (FreeSolv): 642 molecules (paper)
  • Largest dataset (PCBA): 437,929 compounds (paper)
  • QM9 molecules: 133,885 molecules (paper)
  • DeepChem molnet loaders: 44 (DeepChem master, checked 2026-11-16 (includes non-paper datasets))
Wet-lab refreshStatic data · No new experiments; public databases are reused and QM9 is a computed DFT oracle.
Own wet lab needed to evaluate a new modelNo, run it as a download
Models reported2 counted (self 1, modified protocol 1, unconfirmed 1)
Unverified fields† release_milestones.journal_online, ranking.entries[1], ranking.entries[2], models_reported[2]
As of2026-11-16

What it measures

  • Task: Small-molecule property prediction across 17 public datasets covering quantum, physical chemistry, biophysics and toxicity endpoints.
  • Truth: DFT values for QM sets; others from literature compilations, PubChem screens and Tox21 cell assays.
  • Primary metric: ROC-AUC (classification); RMSE/MAE (regression)
  • Secondary metrics: PRC-AUC (PCBA, MUV)
  • Version traps: Same-name splits differ by implementation; OGB sets use fixed scaffold splits, not comparable; Tox21 here differs from the original challenge.

Current ranking

Not comparable

MetricROC-AUC (classification), MAE/RMSE (regression)
Eval setMoleculeNet 17 datasets; splits (scaffold/random) and seeds differ per paper
Snapshot
BoardNo official board
Retrieved2026-11-16

Caveat. No official leaderboard; data frozen since 2018 and each paper re-splits, so cross-paper numbers are no ranking; OGB-derived sets are not comparable.

No headline entry: Not comparable: no official board and each paper re-splits, so no headline number.

The entries below use different protocols, so they are listed, not ranked, and the overview table shows no score.

RankModelScore (verbatim)Metric · eval set / protocolSource type · dateSource and note
—MoLFormer-XLWithheld: not-comparable: protocols/subsets differ, so no score is displayed in the ranking row; the per-protocol values live on the entity page.ROC-AUC (classification), MAE/RMSE (regression)
MoleculeNet 17 datasets; splits (scaffold/random) and seeds differ per paper
Self-reported
source Self-reported (Table 1, scaffold splits): authors claim first on 3 of 6 sets and close second on the rest; no board, self-reported positions only.
—Uni-Mol2Withheld: not-comparable: protocols/subsets differ, so no score is displayed in the ranking row; the per-protocol values live on the entity page.ROC-AUC (classification), MAE/RMSE (regression)
MoleculeNet 17 datasets; splits (scaffold/random) and seeds differ per paper
Self-reported
source second-hand Reports only 8 QM9 properties with a scaffold rather than the recommended random split; per-property values not re-verified this round.
—Chemprop v2Withheld: Incomplete provenance (missing url); the honesty rule forbids displaying a number.ROC-AUC (classification), MAE/RMSE (regression)
MoleculeNet 17 datasets; splits (scaffold/random) and seeds differ per paper
Third-party
No sourced date†
no source link No v2-specific MoleculeNet numbers exist; circulating numbers belong to the 2019 D-MPNN original; pairing unverified.

Models reported

Grouped by relation type; each line shows its source type. Counted relations add up to 2. A pairing whose evidence rests only on this survey's own page is shown as unverified.

Developer report, official protocol (1)

  • MoLFormer-XL Self-reported · MoleculeNet scaffold/random splits · evidence: link

Developer report, modified protocol (1)

  • Uni-Mol2 Self-reported · QM9 only, scaffold split †partly verified · evidence: link

Unconfirmed (1, not counted)

  • Chemprop v2 Unconfirmed source · Page marks it unconfirmed; numbers belong to the 2019 original †disputed · evidence: link

Caveats

  1. Known data errors are not corrected, and simple fingerprint baselines often match pretrained models; split implementations differ, so numbers are often not comparable.
  2. Third-party analyses report unparsable SMILES and duplicates in BBBP, stereoisomer activity gaps in BACE, and an ESOL range far beyond industrial assays.
  3. A 2025 preprint reports that MoleculeNet's Tox21 training and test sets differ from the original challenge.
  4. A 2025 preprint found that of 25 pretrained molecular embedding models, all but one fail to beat an ECFP baseline.
  5. No blind or prospective test; no study links its scores to prospective hit rates.

How to use it

Run statically via DeepChem loaders or OGB fixed splits; no wet lab, retrospective evidence only. Use fixed scaffold or cluster splits, report an ECFP plus tree-model baseline, at least 10 seeds with SD, cleaned BBBP/BACE, and deduplicate pretraining data. Pair with TDC, Polaris and FS-Mol.

Pairs well with: TDC (ADMET + DTI) · Polaris / ASAP / OpenADMET · FS-Mol · CASF-2016

Sources

  1. MoleculeNet paper (arXiv 1703.00564) (preprint)
  2. MoleculeNet journal version (Chem. Sci. 2018) (journal)
  3. DeepChem molnet loader source (repository)
  4. Cheminformatics blog: benchmark data-quality analysis (2023-08) (second-hand) second-hand
  5. Pretrained molecular embeddings vs ECFP baseline (arXiv 2508.06199) (third-party paper)
  6. Tox21 reproduction study (arXiv 2511.14744) (third-party paper)
  7. OGB graph property prediction docs (official site)
  8. MoLFormer-XL paper (arXiv 2106.09553) (preprint)
  9. Uni-Mol2 paper (arXiv 2406.14969) (preprint)

Corrections

1 correction
FieldWasNowBasisFound
ranking.entries[2] (Chemprop v2)MoleculeNet × Chemprop v2 listed as a pairing with numberswithheld: incomplete provenance (missing url); numbers belong to the 2019 D-MPNN originalsource2026-11-16

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