Datasets

Dataset configs (D_*.yaml) define how raw data is released into versioned shards. Each config may include description and citation keys for documentation (ignored at runtime). citation may be a string or a list of strings when multiple references apply.

D_AFCATH

CATH domain structures from AlphaFold Swiss-Prot for inverse folding evaluation, with API fallback for missing entries.

Please cite

Orengo et al. “CATH — A Hierarchic Classification of Protein Domain Structures.” Structure 5, 1093–1109 (1997).

Jumper et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

adapter:
  AlphaFoldInvBenchAdapter:
    af_name: swissprot_pdb
    af_version: v4
    use_api_fallback: true
    api_workers: 64

D_AFEC00

AlphaFold structures for E. coli K-12 with sequence-length filtering and fixed train/validation/test splits.

Please cite

Jumper et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

adapter:
  AlphaFoldAdapter:
    name: UP000000625_83333_ECOLI
transforms:
- FilterSequenceLength:
    max_length: 1024
- SceneSplit:
    test_size: 100
    val_size: 100

D_AFFULL

AlphaFold Swiss-Prot structures with Foldseek-based exclusion filtering, length cap, and train/validation/test splits.

Please cite

Jumper et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

adapter:
  AlphaFoldExclusionAdapter:
    name: swissprot_pdb
    version: v4
transforms:
- FilterSequenceLength:
    max_length: 1024
- SceneSplit:
    test_size: 1000
    val_size: 1000

D_AFSP00

AlphaFold Swiss-Prot structures with sequence-length filtering and train/validation/test splits.

Please cite

Jumper et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

adapter:
  AlphaFoldAdapter:
    name: swissprot_pdb
transforms:
- FilterSequenceLength:
    max_length: 1024
- SceneSplit:
    test_size: 1000
    val_size: 1000

D_ARES00

RNA structure models from the ARES puzzle benchmark.

Please cite

Townshend et al. “Geometric Deep Learning of RNA Structure.” Science 373, 1047–1051 (2021). https://doi.org/10.1126/science.abe5650

adapter: AresAdapter

D_INVC42

CATH domain structures and sequences from the ProteinInvBench inverse folding suite.

Please cite

Li et al. “ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Structures.” (ProteinInvBench repository; see A4Bio/ProteinInvBench for the current citation).

adapter: ProteinInvBenchAdapter

D_PROGYM

Deep mutational scanning assays and structures from ProteinGym.

Please cite

Notin et al. “ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design.” NeurIPS Datasets and Benchmarks Track (2023).

adapter: ProteinGymAdapter

D_PROGYM_BLAT_ECOLX

ProteinGym BLAT_ECOLX deep mutational scanning assay with AlphaFold2 structures.

Please cite

Notin et al. “ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design.” NeurIPS Datasets and Benchmarks Track (2023).

Stiffler et al. “Protein Stability Engineering Insights Revealed by High-Throughput Screening.” PNAS 112, E3098–E3106 (2015). https://doi.org/10.1073/pnas.1504567112

adapter:
  ProteinGymAdapter:
    assays:
    - BLAT_ECOLX_Stiffler_2015

D_PSAFSP

ProteinShake AlphaFold Swiss-Prot structures with deduplication and scene splits.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: AlphaFoldDataset_swissprot
transforms:
- DeduplicateAtoms
- SceneSplit

D_PSEC00

ProteinShake enzyme commission (EC) classification structures.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: EnzymeCommissionDataset
transforms:
- DeduplicateAtoms

D_PSGO00

ProteinShake gene ontology molecular function labels.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: GeneOntologyDataset
transforms:
- DeduplicateAtoms

D_PSLDEC

ProteinShake protein–ligand decoy structures for virtual screening.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: ProteinLigandDecoysDataset
transforms:
- DeduplicateAtoms
- SceneSplit

D_PSLINT

ProteinShake protein–ligand interface residues and structures.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: ProteinLigandInterfaceDataset
transforms:
- DeduplicateAtoms

D_PSPFAM

ProteinShake Pfam family classification structures.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: ProteinFamilyDataset
transforms:
- DeduplicateAtoms

D_PSPPI0

ProteinShake protein–protein interface structures.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: ProteinProteinInterfaceDataset
transforms:
- DeduplicateAtoms

D_PSRCSB

ProteinShake RCSB PDB structures with deduplication and scene splits.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: RCSBDataset
transforms:
- DeduplicateAtoms
- SceneSplit

D_PSSCOP

ProteinShake SCOP fold classification structures.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: SCOPDataset
transforms:
- DeduplicateAtoms

D_PSTMAL

ProteinShake TM-align structural similarity pairs.

Please cite

Günther et al. “ProteinShake: Building Blocks and Benchmarks for Data-Driven Protein Modeling.” NeurIPS Datasets and Benchmarks Track (2023).

adapter:
  ProteinShakeAdapter:
    dataset: TMAlignDataset
transforms:
- DeduplicateAtoms

D_QNTMA9

QM9 small-molecule quantum chemistry properties with random train/validation/test splits.

Please cite

Ramakrishnan et al. “Quantum Chemistry Structures and Properties of 134 Kilo Molecules.” Scientific Data 1, 140022 (2014). https://doi.org/10.1038/sdata.2014.22

adapter: QuantumMachinesAdapter
transforms:
- SceneSplit:
    train_size: 0.8
    test_size: 0.1
    val_size: 0.1

D_RMDASP

Aspirin molecular dynamics trajectories from the revised MD benchmark.

Please cite

Chmiela et al. “Machine Learning Accurate Exchange and Correlation Functionals of the Electronic Density.” Nature Communications 10, 3887 (2019). https://doi.org/10.1038/s41467-019-12827-2

adapter:
  RevisedMolecularDynamicsAdapter:
    name: aspirin

D_SRFC42

Surface-residue-filtered CATH domains from ProteinInvBench for inverse folding.

Please cite

Li et al. “ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Structures.” (ProteinInvBench repository; see A4Bio/ProteinInvBench for the current citation).

adapter: ProteinInvBenchAdapter
transforms:
- ResiduePositions: null
- IsSurfaceResidue: null
- FilterResiduesByValue:
    attribute: residue_is_surface
    value: 1

D_TINY00

Small AlphaFold subset (M. jannaschii) for fast local testing and CI.

Please cite

Jumper et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2

adapter:
  AlphaFoldAdapter:
    name: UP000000805_243232_METJA
transforms:
- FilterSequenceLength:
    max_length: 1024
- SceneSplit:
    test_size: 100
    val_size: 100

D_WNGDK0

Protein–ligand docking structures from the Weng lab benchmark (version 5.5).

Please cite

Weng et al. “Docking Benchmark Version 5.5.” (see Weng lab docking benchmark for the current citation).

adapter:
  WengDockingAdapter:
    version: '5.5'
transforms:
- DeduplicateAtoms
- SceneSplit