timodonnell/helico-data

Helico Training Data

Preprocessed, model-ready training data for Helico — an AlphaFold3 clone built from scratch in PyTorch. The core of this repository is 236,326 structures from the RCSB PDB, tokenized in AF3 convention, pickled, and published as versioned snapshots.

⚠️ This is not a datasets-loadable dataset. The payload is Python pickles inside split tar archives; load_dataset() will not work and the dataset viewer is disabled. Unpickling requires the helico package to be importable, and carries the usual pickle code-execution caveat. See Getting the data.

Contents

Path	What it is
processed/pdb-2026-08-08/	Current snapshot. 236,326 structures with contacts, 83.4 GB. Full documentation →
processed/pdb-2026-04-22/	Superseded, metadata only — see the warning below
processed/ccd_cache.pkl	Parsed Chemical Component Dictionary, 117.8 MB. Shared across snapshots; needed for inference
processed/latest.json	Names the current snapshot per source type
benchmarks/FoldBench/	Vendored FoldBench evaluation suite, 2.24 GB

Snapshots

Data is published as immutable, date-stamped snapshots rather than edited in place. processed/latest.json maps a source type to the current snapshot id:

{"schema_version": 1, "sources": {"pdb": "pdb-2026-08-08"}}

Each snapshot directory carries a SOURCE.json recording the raw data sources, the preprocessing parameters, and the git_sha of the Helico commit that produced it.

Snapshot	Structures	Contacts	Status
pdb-2026-08-08	236,326	✅ pyconfind contacts-v1	Current
pdb-2026-04-22	—	❌	⚠️ Metadata only

⚠️ pdb-2026-04-22 contains no structures. Its SOURCE.json claims 236,326, but the structures.tar.* chunks were never uploaded — a publishing bug silently packed an empty directory. Only its manifest, MSA indices, and provenance record are present. Use pdb-2026-08-08.

Getting the data

Helico is not published to PyPI; install it from source:

git clone https://github.com/Open-Athena/helico && cd helico
uv pip install -e ".[dev]"

Then let it resolve, fetch and extract everything:

helico-download                            # follows latest.json
helico-download --snapshot pdb-2026-08-08  # pin explicitly
helico-download --subset ccd-only          # just the CCD cache

Files land in ~/.cache/helico/data/ by default; override with --data-dir or the HELICO_DATA_DIR env var. Note that helico-download deliberately flattens the snapshot id away — the snapshot is a publishing concern, and the training code sees a flat processed/ layout.

To bypass the CLI, download processed/<snapshot>/structures.tar.* and concatenate them in order:

cat structures.tar.* | tar -xf -

What's in a structure

One pickled TokenizedStructure per PDB entry, tokenized in AF3 convention — one token per protein residue, one per nucleotide, one per heavy atom for ligands. Each carries per-token atom names, elements, coordinates, CCD reference coordinates, and charges, plus sparse edge lists for covalent bonds and residue–residue contacts.

The corpus was built by parsing 252,091 mmCIF files from the RCSB archive, dropping water and hydrogens, and keeping entries at ≤ 9.0 Å resolution with at least one polymer chain — 236,326 passed.

The current snapshot additionally carries side-chain contacts computed by pyconfind with the same contacts-v1 parameters that MarinFold uses, densified at load time into a three-state (contact / no-contact / unknown) token × token matrix. This supports contact-conditioned, MSA-free folding.

→ Full field-by-field documentation, provenance, and generation pipeline

Train/val split

The split is applied at load time from release_date in the manifest — it is not baked into the files. Helico's defaults match AF3, Protenix v1, and OpenFold3-preview2 so metrics are directly comparable:

Split	Rule	Count
Train	release_date < 2021-09-30	170,926
Val	2022-05-01 ≤ release_date ≤ 2023-01-12	9,716

Structures in the 2021-09-30 → 2022-05-01 gap are in neither split — AF3's deliberate leakage-prevention design.

⚠️ A temporal split is not a redundancy split. 38.2% of validation structures share at least one chain sequence verbatim with training, and 18.4% share every chain sequence. Any holdout built from these files by date needs an explicit sequence-identity filter before it will support a generalization claim. Details in the snapshot README.

Note also that no cluster-based weighted sampling, deduplication, or molecule-type rebalancing is applied — the corpus is published as-is.

MSAs

Precomputed alignments are not hosted here — the two *_msa_index.pkl files in each snapshot are byte-offset maps enabling O(1) random reads into archives you obtain separately:

• rcsb_raw_msa.tar (131 GB) — https://boltz1.s3.us-east-2.amazonaws.com/rcsb_raw_msa.tar
• openfold_raw_msa.tar (88 GB) — https://boltz1.s3.us-east-2.amazonaws.com/openfold_raw_msa.tar

Lookup is content-addressed by sequence hash, not PDB ID. The OpenFold index is published for completeness but is not usable through Helico's current lookup path (its members are keyed by UniProt accession, and that mapping is not implemented).

Benchmarks

benchmarks/FoldBench/ is a vendored checkout of the FoldBench evaluation suite (cloned 2026-03-04, ground-truth CIFs from 2025-05-20), including precomputed MSAs for its targets and vendored copies of baseline algorithm code. It is bundled for reproducible benchmarking; FoldBench is MIT licensed and carries its own LICENSE, as does each vendored algorithm. See the FoldBench paper in Nature Communications.

Provenance and terms

This repository aggregates data from several upstream sources under differing terms, which is why no single license is declared:

Component	Origin
Structures	RCSB PDB — archive distributed without copyright restriction (CC0 1.0)
Chemical Component Dictionary	wwPDB
MSA archives (indexed, not hosted)	Published by Boltz; OpenFold alignments derive from OpenFold
Contacts	pyconfind
benchmarks/FoldBench/	FoldBench, MIT; includes vendored third-party algorithm code under its own licenses

If you use this data, please cite the underlying PDB entries and the upstream projects above. Usage is governed by the terms of those sources.

Related

• Code and training recipes: Open-Athena/helico
• Model weights: timodonnell/helico