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

Experimental (PDB) and predicted (AlphaFold DB) molecular structures packaged for RFD4-Proteína training: AFDB protein monomers and PDB interfaces, including protein–DNA, antibody and TCR interfaces. Original structure shards, byte offsets, indices, and sampling tables are preserved.

This release contains only PDB-derived and AlphaFold DB-derived components. Other components of the full training mix (distilled and synthetic sets, small-molecule, RNA and protein–ligand collections) are not part of this release, and neither are components that were derived from them. Dataset availability here does not establish which components a particular checkpoint used.

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

Snapshot at a glance

	Details
Configurations	5 named metadata views over 2 blob directories (AFDB monomers, PDB snapshot)
Payload	4,251 objects · 238.52 GB (222.15 GiB) across the three published component directories
Storage	Parquet sampling tables and indices, plus independently compressed records in binary shards
PDB snapshot	October 4, 2026 PDB snapshot, 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)

The source snapshot date is not a universal model-training cutoff.

Choose a component

Counts and proposed sampling shares come from dataset-inventory.json. “Passing filters” refers to the recorded audit filters, not an independently revalidated quality score.

Configuration	Contents	Published examples	Passing filters	Proposed share in the full mix
afdb_monomer	AFDB protein monomers	1,186,995	914,831	69%
pdb_interfaces	PDB molecular interfaces	13,938,465	8,431,727	14.5%
antibody_interfaces_pdb	PDB antibody interfaces	81,895	72,562	0.5%
tcr_interfaces_pdb	PDB TCR interfaces	9,526	8,608	0.5%
pdb_dna_interface	PDB protein–DNA interfaces	12,812	—	—

The shares are preserved from the source manifest of the full training mix and are not renormalized for this subset; they sum to less than 100% here because the other components are not released. pdb_dna_interface has no recorded share. These are overlapping sampling views, not counts of globally unique structures. The antibody/TCR PDB views share immune_interfaces.parquet and are distinguished by runtime filters; the general PDB view also shares their blob directory, and pdb_dna_interface reads structures from the same PDB blobs.

Load metadata first

For inspection without downloading structure shards:

pip install datasets huggingface_hub

from datasets import load_dataset

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

The rows contain metadata, not decoded molecular structures. load_dataset does not apply the manifest's runtime filters, cluster weighting, or proposed sampling shares. The train label describes the exposed sampling table; it does not imply 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 or metadata.parquet   # sampling and quality metadata
└── blob/
    ├── index.parquet                     # record ID → shard, offset, length
    └── data/*.bin                        # independently compressed records

Shards are at most about 67 MB, so a single record can be read without downloading a component. From an environment with the RFD4/AtomWorks dependencies installed:

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

repo = "nvidia/RFD4-Proteina"
prefix = "train_datasets/streaming-latest/synthetic/afdb_genie3_subset/"
index = hf_hub_download(repo, prefix + "blob/index.parquet", repo_type="dataset")
row = pq.read_table(index).slice(0, 1).to_pylist()[0]
shard = hf_hub_download(repo, prefix + "blob/data/" + row["shard"], repo_type="dataset")
store = BlobStore(str(Path(shard).parent))
record = store.get_bytes(row["shard"], int(row["offset"]), int(row["length"]))
Path("example.cif").write_bytes(record)
print(row["example_id"], "saved to example.cif")

This reads one indexed structure, not a training-sampler draw. BlobStore handles per-record decompression. For training, use the component's matching sampling table and the loader/configuration in the training guide.

AtomWorks' indexed-blob loader supports local storage and S3 ranged reads. Direct Hugging Face ranged reads are not a drop-in loader backend; download the selected component locally. Keep its index and shards from the same revision.

Filters, provenance, and reproducibility

The inventory records each component's sampling table, blob directory, filters, proposed share, and object sizes. Component READMEs and provenance/ folders hold additional processing details.

For a reproducible training run, record the dataset commit, model/code revisions, component selection, runtime filters, mixture weights, and sampler settings. File 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: recorded filters include deposition before January 1, 2024, resolution below 9 Å, and the inventory's bond, ligand, interface, and size gates. Inherited interface definitions and geometry-dependent flags were not recertified on new coordinates.
• PDB–DNA interfaces: added on 2026-10-13; the table (12,812 rows) reads structures from the existing PDB blobs, with matching dna_binding_interfaces_20261013 sampling files.
• AFDB monomers: predicted structures filtered to at most 256 standard amino-acid residues; they are not experimental measurements.
• Missing counts: null means unknown or unaudited, not zero. Read component metadata for the precise atom/residue-count conventions before filtering.

Intended use and limitations

Use this snapshot for training, fine-tuning, data inspection, and reproducibility studies with compatible molecular loaders. Predicted structures are not experimental measurements. Coverage and errors inherit the biases of the upstream experimental and predicted sources. Large numbers of rows do not imply equivalent diversity or independent samples.

Before constructing an evaluation, audit overlap with your targets at the appropriate sequence, structure, complex, and source-date levels. Neither the Hub metadata configuration nor the proposed mixture automatically performs that audit.

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. 640 annotated PDB structures (128 each of protein–ligand, modified residue, glycan, DNA and RNA examples, all deposited on or after 2024-01-01) 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 (caplacizumab, ozoralizumab, vobarilizumab, h-NbBCII10).
• metrics/cdr_torsion_data/: reference data for the cdr_torsion_naturalness antibody metric: native CDR phi/psi samples (VH/VL and VHH) and their kernel-density parameters. The metric downloads it on first use.
• media/: images and video used by the project page and the code repository's 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:

• PDB entries come from the wwPDB archive, whose data are available under the CC0 1.0 waiver.
• AlphaFold DB structures (EMBL-EBI / Google DeepMind) are available under CC BY 4.0, which requires attribution.
• 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). All upstream sources above are compatible with CC BY 4.0 attribution. The model's BSD-3-Clause license is separate and does not govern these data. Credit the upstream sources as well as this dataset, and check the component README before reuse.

When reporting work, cite the RFD4-Proteína dataset repository at the exact commit, credit each upstream source used, and describe any filtering or resampling. The inventory and component READMEs describe how each component was processed.

Model and examples · Dataset history · Code and issues