suite · audited · verified 2026-07-22

ATOM3D

A living collection of eight curated 3D molecular-learning tasks spanning small molecules, protein interactions and mutations, ligand binding, and protein/RNA structure ranking.

+6 more

Benchmark definition

What is counted

Version: v0.2.6
Total: 8 (curated 3D benchmark datasets)
Task formats: 3D molecular regression; 3D molecular classification; residue classification; structure ranking
Capabilities: PredictionClassificationRegression
Modalities: 3D structure

Version history

Version	Status	Release / as-of	Total	Formal tracks
v0.2.6 `atom3d-v0-2-6`	rolling	2020-12-07	8 (curated 3D benchmark datasets)	None registered

Scientific Task Atlas

Scientific task classification

complete for v0.2.6 · as of 2026-07-22. The eight official task abbreviations are mapped one-to-one to their documented scientific prediction problems.

Scientific task	Coverage	Count	Mapping	Evidence
Small-molecule property prediction	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` SMP, small-molecule property prediction from molecular structure.
Protein-protein interface prediction	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` PIP predicts whether residue pairs from two proteins contact when the proteins bind; it is narrower than binary interaction detection.
Protein residue-identity prediction	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` RES, residue identity prediction from the local structural environment.
Protein-complex mutation stability prediction	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` MSP predicts whether a single mutation stabilizes a protein complex; it is not a monomer thermostability task.
Protein-ligand binding affinity	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` LBA, ligand binding-affinity prediction.
Protein-ligand functional efficacy prediction	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` LEP classifies whether a bound molecule activates protein function from active and inactive target conformations; it does not predict binding presence or affinity.
Protein model quality assessment	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` PSR, protein structure ranking.
RNA structure quality assessment	explicitly-in-scope	1 tasks Curated 3D benchmark datasets.	official-track high confidence	`atom3d-paper-definition-evidence` RSR ranks candidate RNA structures and is therefore not relabeled as de novo RNA folding.

Scientific coverage notes

Domain	Coverage	Count	Interpretation
Protein-protein binding	explicitly-in-scope	1	Protein Interface Prediction is one of the eight official datasets.
Protein-ligand binding	explicitly-in-scope	2	Ligand Binding Affinity and Ligand Efficacy Prediction are distinct official datasets.
Protein structure	explicitly-in-scope	3	Residue Identity, Mutation Stability Prediction, and Protein Structure Ranking operate on protein structures; their task semantics remain distinct.

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

atom3d-creator-full

From ATOM3D: Tasks On Molecules in Three Dimensions

Scopefull · n=8

ShotsNot applicable

TurnsNot applicable

System prompt publicNot applicable

Reasoning / effortNot applicable

BrowserNot applicable

InternetNot applicable

DatabasesNot applicable

Code executionNot applicable

ContainerNot reported

External toolstask-specific 3D CNN, graph neural network, and equivariant neural network pipelines

Token budgetNot applicable

Time / cost budgetNot reported

TemperatureNot applicable

SeedNot reported

Repeats3

Graderdeterministic task-specific scorer · human review: no

StatisticsNative per-task metrics with standard deviations over three replicates; structure-ranking correlations are computed per target before summary.

ContaminationTask-specific sequence-identity, time, target, or scaffold splits.

Metrics, results, and full protocol

Metrics

Metric	Kind / baseline	Unit	Aggregation	Threshold / tolerance
MAE	absolute	task-specific	held-out examples	Not reported
RMSE	absolute	task-specific	held-out examples	Not reported
AUROC	absolute	area	held-out examples	Not reported
Accuracy	absolute	proportion	held-out examples	Not reported
Pearson correlation	absolute	correlation	task-specific	Not reported
Spearman correlation	absolute	correlation	target-level structure ranking	Not reported

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

Evidence

table: Sections 2-5 and Tables 1-8; Appendix D (Defines all eight datasets, official splits, model classes, native metrics, three-replicate aggregation, and creator baseline results.) — supports /scope, /protocol, /metrics

Evidence and change history

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

ATOM3D: Tasks On Molecules in Three Dimensions · table: Sections 2-4 and Tables 1-2 (Defines SMP, PIP, RES, MSP, LBA, LEP, PSR, and RSR and their task-specific evaluation.) · Supports 30 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
/scientific_task_classification/entries/3
/scientific_task_classification/entries/4
/scientific_task_classification/entries/5
/scientific_task_classification/entries/6
/scientific_task_classification/entries/7

atom3d-project-resource · web-anchor: Datasets; SMP, PIP, RES, MSP, LBA, LEP, PSR, and RSR detail pages (Confirms the eight-task living catalog and gives exact task, split, download, and license statements for each dataset.) · Supports 11 fields

Open source →

/summary
/domains
/modalities
/task_counts/total
/coverage_notes
/access/level
/access/tasks
/access/grader
/access/license
/resources
/versions/0/as_of

atom3d-repository-resource · repository-path: README.md; atom3d/__init__.py; examples/{smp,pip,res,msp,lba,lep,psr,rsr} at 4c2f3b7e9efe128791b83f03b2e8cae91e78b018 (Confirms package version, public loaders, eight task implementations, and MIT license.) · Supports 8 fields

Open source →

/latest_version
/access/tasks
/access/artifacts
/access/grader
/resources
/implementations
/versions/0/as_of
/versions/0/formal_tracks

View source-level modification history on GitHub →

Access & reproducibility

Access: Fully open
License: MIT for code; datasets retain their original source licenses and citation requirements.
Grader: Deterministic MAE/RMSE, AUROC, accuracy, and rank-correlation evaluation depending on task.
Implementations: 1 recorded · official