suite · audited · verified 2026-07-22

MoleculeNet

The original molecular-machine-learning benchmark of 17 dataset collections and more than 800 prediction endpoints spanning quantum, physicochemical, biophysical, and physiological properties.

+1 more

Benchmark definition

What is counted

Version: original-2017
Total: 17 (original paper dataset collections)
Task formats: molecular property regression; molecular property classification; multitask endpoint prediction
Capabilities: PredictionClassificationRegression
Modalities: Small-molecule structure3D structure

Version history

Version	Status	Release / as-of	Total	Formal tracks
original-2017 `moleculenet-original-2017`	current	2017-03-02	17 (original paper dataset collections)	None registered

Tracks and subsets

ID	Count	Basis	Partition?	Notes
Quantum mechanics collections `moleculenet-quantum`	4	original paper dataset collections	Exclusive & exhaustive	QM7, QM7b, QM8, and QM9.
Physical chemistry collections `moleculenet-physical`	3	original paper dataset collections	Exclusive & exhaustive	ESOL, FreeSolv, and Lipophilicity.
Biophysics collections `moleculenet-biophysics`	5	original paper dataset collections	Exclusive & exhaustive	PCBA, MUV, HIV, BACE, and PDBbind.
Physiology collections `moleculenet-physiology`	5	original paper dataset collections	Exclusive & exhaustive	BBBP, Tox21, ToxCast, SIDER, and ClinTox.

Scientific Task Atlas

Scientific task classification

partial for original-2017. The fixed original suite is mapped at dataset-collection level; heterogeneous biochemical endpoint collections are not atomized into unsupported task counts.

Scientific task	Coverage	Count	Mapping	Evidence
Small-molecule property prediction	explicitly-in-scope	17 other Original paper dataset collections.	official-taxonomy high confidence	`moleculenet-paper-definition-evidence` Umbrella mapping for all 17 original dataset collections, which span quantum, physical, biophysical, and physiological properties.
ADMET and toxicity prediction	explicitly-in-scope	5 other Original paper dataset collections.	official-taxonomy high confidence	`moleculenet-paper-definition-evidence` The five physiology collections are ESOL-independent ADMET or toxicity datasets; no endpoint-level total is asserted here.
Protein-ligand binding affinity	explicitly-in-scope	1 other Original paper dataset collections.	official-taxonomy high confidence	`moleculenet-paper-definition-evidence` PDBbind is the explicitly structure-based protein-ligand affinity collection.

Scientific coverage notes

Domain	Coverage	Count	Interpretation
Medicinal chemistry	explicitly-in-scope	17	Dataset collections cover molecular properties and drug-discovery endpoints; 17 is not the number of individual prediction endpoints.
Protein-ligand binding	explicitly-in-scope	1	PDBbind is the explicit binding-affinity collection; other biophysical activity collections are not relabeled as affinity tasks.

Relationship registry

How works use this benchmark

Partial claims, non-evaluation uses, and third-party summaries stay visible without entering model comparisons.

Partial evaluation claims

evaluation

enhancing-molecular-property-prediction-with-auxiliary-moleculenet-2-use

Work: Enhancing molecular property prediction with auxiliary learning and task-specific adaptation · source version enhancing-molecular-property-prediction-with-auxiliary-pmc-version-1

Selection: filtered · Eight MoleculeNet classification datasets: BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIV, and BACE.
Models: supervised_contextpred (Sup-CP)
Metrics: Test ROC-AUC (Sup-CP; AM,CP,EP,IG,MP)
Linked runs: None

Not reported / unresolved: MoleculeNet version is not reported.; Realized test-set sample sizes are not reported.; Exact checkpoint revision and release date are not reported.; Scaffold splitting is the within-dataset split protocol, not the method used to select the eight benchmark datasets.; Table rows vary by adaptation strategy; those strategies are not normalized here as standalone model identities.; Conflicted result claim omitted after independent verification; the evaluation relationship is published conservatively.; benchmark version; numeric result

AI-assisted double-pass extraction; values are limited to independently supported claims.

Evidence

table: Table 1
Supports: /relation_type
section: Results and discussion > Experimental setup
Supports: /benchmark_id
section: Results and discussion > Experimental setup
Supports: /scope
section: Results and discussion > Experimental setup
Supports: /scope
table: Table 1, dataset columns
Supports: /scope
table: Table 1 footnote
Supports: /scope
other: Experimental setup; reference 17
Supports: /model_ids
table: Table 1 caption and footnote
Supports: /metric_labels

evaluation

enhancing-molecular-property-prediction-with-auxiliary-moleculenet-3-use

Work: Enhancing molecular property prediction with auxiliary learning and task-specific adaptation · source version enhancing-molecular-property-prediction-with-auxiliary-pmc-version-1

Selection: filtered · Eight MoleculeNet classification datasets: BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIV, and BACE.
Models: supervised_contextpred (Sup-CP)
Metrics: Test ROC-AUC (Sup-CP; AM,IG,MP)
Linked runs: None

Not reported / unresolved: MoleculeNet version is not reported.; Realized test-set sample sizes are not reported.; Repeat count is not stated for Table 2.; Exact checkpoint revision and release date are not reported.; Scaffold splitting is the within-dataset split protocol, not the method used to select the eight benchmark datasets.; Table rows vary by adaptation strategy; those strategies are not normalized here as standalone model identities.; benchmark version; numeric result

AI-assisted double-pass extraction; values are limited to independently supported claims.

Evidence

table: Table 2
Supports: /relation_type
section: Results and discussion > Experimental setup
Supports: /benchmark_id
section: Results and discussion > Experimental setup
Supports: /scope
section: Results and discussion > Experimental setup
Supports: /scope
table: Table 2, dataset columns
Supports: /scope
section: Results and discussion > Reproducibility and implementation details
Supports: /scope
other: Experimental setup; reference 17
Supports: /model_ids
table: Table 2 caption and footnote
Supports: /metric_labels

evaluation

enhancing-molecular-property-prediction-with-auxiliary-moleculenet-4-use

Work: Enhancing molecular property prediction with auxiliary learning and task-specific adaptation · source version enhancing-molecular-property-prediction-with-auxiliary-pmc-version-1

Selection: filtered · Eight MoleculeNet classification datasets: BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIV, and BACE.
Models: supervised (Sup)
Metrics: Test ROC-AUC (Sup; AM,CP,EP,IG,MP)
Linked runs: None

Not reported / unresolved: MoleculeNet version is not reported.; Realized test-set sample sizes are not reported.; Repeat count is not stated for Table 3.; Exact checkpoint revision and release date are not reported.; Scaffold splitting is the within-dataset split protocol, not the method used to select the eight benchmark datasets.; Table rows vary by adaptation strategy; those strategies are not normalized here as standalone model identities.; benchmark version; numeric result

AI-assisted double-pass extraction; values are limited to independently supported claims.

Evidence

table: Table 3
Supports: /relation_type
section: Results and discussion > Experimental setup
Supports: /benchmark_id
section: Results and discussion > Experimental setup
Supports: /scope
section: Results and discussion > Experimental setup
Supports: /scope
table: Table 3, dataset columns
Supports: /scope
section: Results and discussion > Reproducibility and implementation details
Supports: /scope
other: Experimental setup; reference 17
Supports: /model_ids
table: Table 3 caption and footnote
Supports: /metric_labels

Creation, training, validation, or model-selection uses

fine tuning

enhancing-molecular-property-prediction-with-auxiliary-moleculenet-1-use

Work: Enhancing molecular property prediction with auxiliary learning and task-specific adaptation · source version enhancing-molecular-property-prediction-with-auxiliary-pmc-version-1

Selection: not applicable
Models: supervised_contextpred (Sup-CP), supervised (Sup)
Metrics: Not reported / not applicable
Linked runs: None

Not reported / unresolved: MoleculeNet version is not reported.; Table 4 prints the Tox21 molecule count ambiguously as 7.831.

AI-assisted double-pass extraction; values are limited to independently supported claims.

Evidence

section: Results and discussion > Experimental setup
Supports: /relation_type
section: Results and discussion > Experimental setup
Supports: /benchmark_id
section: Results and discussion > Experimental setup
Supports: /scope
section: Results and discussion > Experimental setup
Supports: /scope
table: Table 4, dataset columns
Supports: /scope
section: Results and discussion > Reproducibility and implementation details
Supports: /scope
other: Experimental setup; reference 17
Supports: /model_ids
other: Experimental setup; reference 17
Supports: /model_ids

Evaluation registry

Works and run settings

A setting change—scope, prompt, tools, budget, grader, or repeats—creates a separate run. Charts never cross a comparability group.

Evaluation run

moleculenet-creator-full

From MoleculeNet: a benchmark for molecular machine learning

Scopefull · n=17

ShotsNot applicable

TurnsNot applicable

System prompt publicNot applicable

Reasoning / effortNot applicable

BrowserNot applicable

InternetNot applicable

DatabasesNot applicable

Code executionNot applicable

ContainerNot reported

External toolsDeepChem featurizers, splitters, conventional ML and graph-based models

Token budgetNot applicable

Time / cost budgetNot reported

TemperatureNot applicable

SeedNot reported

Repeats3

Graderdeterministic dataset-specific scorer · human review: no

StatisticsMean and standard deviation over three independent runs for each dataset-model setting; no cross-dataset normalized total.

Contamination80/10/10 train-validation-test partitions with dataset-specific random, stratified, scaffold, or time splits.

Metrics, results, and full protocol

Metrics

Metric	Kind / baseline	Unit	Aggregation	Threshold / tolerance
MAE	absolute	dataset-specific property units	held-out examples and endpoints	Not reported
RMSE	absolute	dataset-specific property units	held-out examples	Not reported
ROC-AUC	absolute	area	dataset-specific macro endpoint average where applicable	Not reported
PRC-AUC	absolute	area	dataset-specific endpoint average where applicable	Not reported

No numeric result rows are published yet; the verified protocol remains useful.

Evidence

table: Methods Sections 3.1-3.5; Results and Discussion; Appendix Performances; Tables 1-3 (Defines all datasets, 80/10/10 partitions, recommended split/metric per collection, featurizers, models, three-run mean/standard-deviation aggregation, and creator results.) — supports /scope, /protocol, /metrics

Evidence and change history

Source locators remain visible; expand an item to inspect the exact Registry fields it supports.

MoleculeNet: a benchmark for molecular machine learning · table: Sections 3-5, Figure 2 and Tables 1-3 (Defines 17 collections, four categories, over 800 endpoints, 80/10/10 splits, recommended splitters, metrics, and creator baselines.) · Supports 25 fields

Open source →

/name
/aliases
/summary
/kind
/organizations
/release_date
/latest_version
/domains
/capabilities
/modalities
/task_formats
/task_counts/total
/task_counts/basis
/task_counts/subsets
/coverage_notes
/access/level
/access/license
/resources
/versions/0/release_date
/versions/0/task_counts/total
/versions/0/task_counts/basis
/versions/0/task_counts/subsets
/scientific_task_classification/entries/0
/scientific_task_classification/entries/1
/scientific_task_classification/entries/2

moleculenet-deepchem-resource · repository-path: deepchem/molnet; LICENSE at fbe3b911a94a6eb8c84c1a9e7ac67a472c49ea80 (Pins the maintained loader and evaluation implementation while preserving the original-paper version boundary.) · Supports 7 fields

Open source →

/access/tasks
/access/artifacts
/access/grader
/access/license
/resources
/implementations
/versions/0/formal_tracks

View source-level modification history on GitHub →

Access & reproducibility

Access: Fully open
License: MIT for DeepChem code; the Chemical Science paper is CC BY-NC 3.0 and underlying datasets retain source-specific terms.
Grader: Deterministic regression and classification metrics with dataset-specific recommended splits.
Implementations: 1 recorded · official