nvidia/RFD4-Proteina

The structures behind biomolecular design.
Indexed molecular records, sampling metadata, and traceable source inventories.

Project page · Model · Inventory · Training guide

RFD4-Proteína training datasets

The structure datasets used to train RFD4-Proteína. This repository holds every training dataset used by the current release recipes: the flow model (hero_pretrain, 2026-10-19) and the unified biomolecular autoencoder (2026-10-15). It covers experimental PDB structures and interfaces, protein–ligand systems, glycans and modified residues, AlphaFold DB monomers, predicted complexes used for distillation, RNA, cyclic peptides and small molecules. Each component keeps its original structure shards, byte offsets, indices, sampling tables and provenance.

Looking to generate designs? Start with the model card. You do not need this dataset to run the model.

Snapshot at a glance

	Details
Scope	All training sources of the release flow and autoencoder recipes (code commit 68058693c)
Configurations	19 named metadata views over 17 component directories
Training data	33,527 objects · 2.13 TB (1.94 TiB)
Layout	Same paths as the training data store, so a download can be used directly as DATASET_STORE_ROOT
Storage	Parquet sampling tables and indices, plus independently compressed annotated mmCIF records in binary shards
PDB snapshot	October 4, 2026 PDB snapshot (frozen 2026-01-06 RCSB mirror), annotated with CATH and PLIP
License	CC BY 4.0 for the added annotations, sampling tables and indices; upstream terms apply to the structures (see Source terms and attribution)

Choose a component

Each configuration matches a source config in the code repository (configs/datasets/sources/). Rows is the size of the sampling table. Passing source filters applies the source config's filters at crop size 256, the setting both release recipes use. Release recipes can override them: for example, the flow recipe caps AFDB monomers at 256 standard residues and the AE recipe at fewer than 256. Flow and AE are the mixture probabilities in the two release recipes; a dash means the recipe does not use that dataset. The exact filter expressions are in dataset-inventory.json.

Configuration	Contents	Rows	Passing source filters	Flow	AE	Size
afdb_monomer	AlphaFold DB protein monomers (Genie3 subset)	1,186,995	1,186,995	71.5%	50%	56.3 GB
pdb_interfaces	PDB molecular interfaces, Oct 4 2026 snapshot, CATH + PLIP annotated	13,938,465	8,431,727	14.5%	9%	182.3 GB
antibody_interfaces_pdb	PDB antibody interfaces (ig_type == 'antibody')	81,895	72,562	0.5%	0.5%	shared
tcr_interfaces_pdb	PDB TCR interfaces (ig_type == 'tcr')	9,526	8,608	0.5%	0.5%	shared
pdb_dna_interface	PDB protein–DNA interfaces (reads the PDB snapshot blobs)	12,812	10,951	0.2%	—	12 MB
pdb_ptm	PDB post-translational-modification environment crops	304,918	135,766	—	4%	1.8 GB
pdb_glycan	PDB glycans with attached protein context	520,043	271,694	—	4%	19.1 GB
plinder	PLINDER protein–ligand systems	275,123	275,123	3.5%	—	34.4 GB
cpsea_pdb	CPSea cyclic peptide–receptor complexes rebuilt on PDB parents, minimized	4,677	4,677	0.5%	1%	275 MB
cpsea_afdb	CPSea cyclic peptide–receptor complexes rebuilt on AFDB parents, minimized	81,114	81,114	0.5%	1%	3.9 GB
ted_dimers	TED predicted protein dimers	587,597	510,213	5%	5%	36.1 GB
omol	OMol molecular structures, annotated	53,485,498	43,686,234	1%	—	156.1 GB
rna_monomer	Predicted RNA monomers	371,017	85,221	1%	6%	29.7 GB
na_complex	ESMFold2-predicted protein–DNA complexes (distillation)	20,261	4,308	0.3%	6%	1.4 GB
antibody_interfaces_distill	ESMFold2-predicted antibody–antigen complexes (distillation)	75,576	6,189	0.5%	0.5%	9.7 GB
tcr_interfaces_distill	AF2-predicted TCR–peptide–MHC complexes (distillation)	7,882	7,882	0.5%	0.5%	301 MB
synthetic_ncaa	ESMFold2-predicted PTM/NCAA peptides	1,463,222	367,680	—	4%	8.3 GB
ncaa_scaffolds	NCAA-modified protein scaffold conformers on AFDB/MGnify predicted monomers	288,390	288,390	—	2%	14.4 GB
reglyco	Re-Glyco: GlycoShape Level-1 glycans rebuilt on PDB glycosylation sites	2,992,448	2,307,945	—	6%	1,574.6 GB

Rows across configurations are overlapping sampling views, not counts of unique structures. pdb_interfaces, antibody_interfaces_pdb, tcr_interfaces_pdb and pdb_dna_interface all read structures from the same PDB snapshot blobs, so their size column counts that 182 GB store once, under pdb_interfaces. The antibody and TCR views share immune_interfaces.parquet and are separated by ig_type.

Component directories

Directory	Objects	Size
train_datasets/streaming-latest/synthetic/glycans/2026_09_27_reglyco/	4,221	1,574.6 GB
train_datasets/streaming-latest/pdb/2026_10_04_tmol_plip_cath/	3,341	182.3 GB
train_datasets/streaming-latest/synthetic/omol_annotated/	3,327	156.1 GB
train_datasets/streaming-latest/synthetic/afdb_genie3_subset/	907	56.3 GB
train_datasets/streaming-latest/synthetic/ted_dimers/	2,332	36.1 GB
train_datasets/streaming-latest/pdb/2026_05_20_plinder/	148	34.4 GB
train_datasets/streaming-latest/synthetic/2026_09_21_rna_monomers/	6,929	29.7 GB
train_datasets/streaming-latest/pdb/2026_09_19_glycan/	2,234	19.1 GB
train_datasets/streaming-latest/synthetic/ncaa/2026_09_26_ncaa_esmfold2_scaffolds/	962	14.4 GB
train_datasets/streaming-latest/synthetic/2026_09_18_antibody_antigen_esmfold2/	3,601	9.7 GB
train_datasets/streaming-latest/synthetic/ncaa/2026_09_21_ptm_esmfold2/	132	8.3 GB
train_datasets/streaming-latest/synthetic/2026_10_04_cpsea_afdb/	4,637	3.9 GB
train_datasets/streaming-latest/pdb/2026_09_19_ptm_crop/	41	1.8 GB
train_datasets/streaming-latest/synthetic/2026_09_22_protein_dna_esmfold2/	133	1.4 GB
train_datasets/streaming-latest/synthetic/2026_09_09_tcr_af2/	236	301 MB
train_datasets/streaming-latest/pdb/2026_09_27_cpsea_pdb/	323	275 MB
train_datasets/streaming-latest/pdb/pdb_dna_interface_2026_10_13/	23	12 MB

Load metadata first

To inspect a component without downloading structure shards:

pip install datasets huggingface_hub

from datasets import load_dataset

rows = load_dataset("nvidia/RFD4-Proteina", "pdb_glycan", split="train", streaming=True)
print(next(iter(rows)))

The rows contain metadata, not decoded structures. load_dataset does not apply the source filters, cluster weighting, or mixture probabilities. The train split name describes the sampling table; it is not a leakage-audited train/test split for your task.

Download and read molecular records

The format separates which example to sample from where its structure is stored:

component/
├── examples.parquet / metadata.parquet / sampling/*.parquet   # sampling and quality metadata
├── README.md                                                   # methods, filters, counts and caveats
├── provenance/                                                 # audit evidence for each processing step
└── blob/
    ├── index.parquet                                           # record ID → shard, offset, length
    └── data/*.bin                                              # independently compressed records

Download one component with allow_patterns, then read one record. This needs the RFD4/AtomWorks dependencies:

from pathlib import Path
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download
from atomworks.ml.utils.blob_store import BlobStore

prefix = "train_datasets/streaming-latest/pdb/2026_09_27_cpsea_pdb/"   # ~0.3 GB
local = Path(snapshot_download("nvidia/RFD4-Proteina", repo_type="dataset",
                               allow_patterns=[prefix + "*"], local_dir="./rfd4-data"))
blob = local / prefix / "blob"
row = pq.read_table(blob / "index.parquet").slice(0, 1).to_pylist()[0]
store = BlobStore(str(blob / "data"))
Path("example.cif").write_bytes(store.get_bytes(row["shard"], int(row["offset"]), int(row["length"])))
print(row["example_id"], "saved to example.cif")

This reads one indexed structure, not a training-sampler draw. BlobStore handles per-record decompression.

Train from a local copy

The repository mirrors the training data store's layout. Download the components you need into one directory, then point the training configs at it:

hf download nvidia/RFD4-Proteina --repo-type dataset --local-dir ./rfd4-data \
  --include "train_datasets/streaming-latest/synthetic/afdb_genie3_subset/*" \
            "train_datasets/streaming-latest/pdb/2026_10_04_tmol_plip_cath/*"
export DATASET_STORE_ROOT=$PWD/rfd4-data

The recipes, mixture weights and loader settings are in the training guide and in configs/releases/ of the code repository. AtomWorks' blob loader reads local storage or S3 ranged reads; Hugging Face ranged reads are not a drop-in backend, so download components locally. Keep an index and its shards from the same revision.

What is and isn't included

• Selection rule: every dataset that a release recipe (configs/releases/flow/2026_10_19_hero_pretrain.yaml, configs/releases/ae/2026_10_15_unified_biomolecule.yaml) samples from. Datasets that only appear in non-release configs, such as MGnify AF2 monomers, are not included.
• Live shards only: after in-place reannotation campaigns, some components kept superseded shards so that running jobs could finish. Only shards referenced by the current blob/index.parquet are published. Large prior-version table snapshots under provenance/ (>256 MiB each) are also left out. Together these come to 237.1 GB, listed under excluded_from_hub in the inventory.
• Component READMEs are published unchanged. They mention s3:// paths in the original data store; inside this repository the same files sit under train_datasets/streaming-latest/.

Filters, provenance, and reproducibility

Each component's README.md documents its construction, filters, counts and known issues. provenance/ folders hold the audit evidence: hashes, reversible prior indices, and code revisions. dataset-inventory.json records, for every configuration, its sampling table, blob directory, source filters, row counts and recipe weights, plus the path and size of every published object.

For a reproducible run, record the dataset commit, code and checkpoint revisions, component selection, runtime filters, mixture weights and sampler settings. Directory names containing latest do not pin data; the Hub commit does. Do not mix metadata from one revision with shards from another.

Important component qualifications

• PDB interfaces: source filters include deposition before 2024-01-01, resolution below 9 Å, and bond, ligand-identity, interface-cluster and size gates. Interface definitions and geometry-dependent flags were inherited, not recertified on new coordinates.
• PDB PTM and glycan crops, Re-Glyco: training views use the train experiment split, deposition before 2024-01-01, and in-vocabulary PTM/glycan filters. Re-Glyco has a held-out split (holdout) that the recipes don't sample.
• OMol: the component README records an unresolved identity mismatch in the source metadata. In a 128-record audit, 126 path values did not match the CIF block retrieved by lmdb_idx. Structures and their computed counts are consistent; the associated identifiers, categories, charges and energies are not reconciled. See synthetic/omol_annotated/README.md.
• Distilled sets (na_complex, antibody_interfaces_distill, tcr_interfaces_distill, synthetic_ncaa, ncaa_scaffolds, ted_dimers, rna_monomer) are predicted structures, not experimental measurements. Their passing counts depend on confidence filters recorded in each source config.
• CPSea: partial regenerations of CPSea; only accepted, minimized complexes are released. They have no train/test split and no experimental validation.
• Missing counts: null means unknown or unaudited, not zero. Read the component README for exact atom/residue-count conventions before filtering.

Intended use and limitations

Use these datasets for training, fine-tuning, data inspection and reproducibility studies with compatible molecular loaders. Predicted and distilled structures are not experimental measurements, and coverage and errors inherit the biases of the upstream experimental and predicted sources. Large row counts do not imply equivalent diversity or independent samples.

Before you build an evaluation, audit overlap with your targets at the sequence, structure, complex and source-date levels. Neither the Hub configurations nor the recipe mixtures perform that audit for you.

Benchmarks and project media

Besides the training data, this repository hosts files that the code repository downloads instead of storing:

• benchmarks/ae_pdb_temporal_holdout_2026_10_04/: the autoencoder validation set. It has 640 annotated PDB structures (128 each of protein–ligand, modified-residue, glycan, DNA and RNA examples, all deposited on or after 2024-01-01), stored as per-file Zstandard-compressed mmCIF, with a README describing the selection.
• benchmarks/nanobody/: 160 nanobody design validation inputs, 40 target structures crossed with 4 VHH frameworks.
• metrics/cdr_torsion_data/: reference data for the cdr_torsion_naturalness antibody metric.
• media/: images and video for the project page and README. This is not training data.

make setup in the code repository downloads the two benchmark directories into data/benchmarks/. They are not part of the load_dataset configurations above.

Source terms and attribution

The structures keep the terms of their upstream sources. Check each component README, and the original source, before reuse:

• PDB entries (all pdb/ components, Re-Glyco context, CPSea-PDB parents) come from the wwPDB archive, whose data are available under the CC0 1.0 waiver.
• AlphaFold DB structures (EMBL-EBI / Google DeepMind; afdb_monomer, CPSea-AFDB parents, TED, NCAA scaffolds) are available under CC BY 4.0, which requires attribution.
• GlycoShape glycan conformers (Re-Glyco) are Level-1 entries under CC BY 4.0; a snapshot of the licensing text is in that component's provenance/glycoshape_terms_snapshot/.
• PLINDER, TED, OMol25, CPSea, RNAcentral and MGnify inputs are credited in their component READMEs. Cite those resources when you use the corresponding components.
• Predicted complexes (ESMFold2 and AlphaFold2 distillations) were generated for RFD4-Proteína training from the sources their READMEs list.
• Derived annotations such as CATH domain labels (CC BY 4.0) and PLIP interaction annotations come from their own sources; cite them when you use them.

The sampling tables, indices, annotations and packaging added for RFD4-Proteína are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0). The model's BSD-3-Clause license is separate and does not govern these data.

When reporting work, cite the RFD4-Proteína dataset repository at the exact commit, credit each upstream source used, and describe any filtering or resampling.

Model and examples · Dataset history · Code and issues