STRUCTURES.md

Structures for the GPBridge expanded-GO dataset

One structure per sequence for 322,992 sequences, assembled from three sources. They are meant to be used together, so the first thing to read is structure_index.tsv: it resolves any seq_id to a file, whichever source provides it.

Sequence ids are seq_<sha256(sequence)[:20]>, so an id always identifies an exact sequence.

Coverage

Source	Sequences	What it is
gpbridge/afdb/	304,787	AlphaFold DB, AlphaFold 2, model v6
gpbridge/pdb/	11,495	experimental structures from RCSB
predictions/missing_pdb_afdb/	6,540	AlphaFold 3, predicted here
no structure	11,075	neither source has one

10,904 sequences have both an AFDB model and a PDB entry — the index exposes both so you can choose per use case.

These come from three different methods. AF2, AF3 and experiment are not interchangeable, and the split is not random: see Provenance below before training on the union.

Resolve a sequence to a file

structure_index.tsv — one row per sequence, all sources on the same row:

seq_id  length  split  best  af3_archive  afdb_archive  afdb_file
                             pdb_archive  pdb_file  pdb_all_ids  note

best is a suggested default, not a judgement: af3 where one was predicted, otherwise afdb, otherwise pdb. That ordering follows the source bundle's own guidance — AFDB is the workhorse, while a PDB hit may be a construct, a single domain, or one chain of a complex and is not guaranteed to cover the full sequence. For experimental ground truth, prefer pdb_archive explicitly.

import csv, io, json, tarfile, zstandard
from huggingface_hub import hf_hub_download
REPO = "tonynzh2/afdb"

idx = {r["seq_id"]: r for r in csv.DictReader(
    open(hf_hub_download(REPO, "structure_index.tsv", repo_type="dataset")),
    delimiter="\t")}

def structure(seq_id, prefer=None):
    """Returns (source, text). prefer='pdb' to force experimental when present."""
    r = idx[seq_id]
    order = [prefer] if prefer else []
    order += ["af3", "afdb", "pdb"]
    for src in order:
        if src == "af3" and r["af3_archive"]:
            return "af3", _from_zst(r["af3_archive"], seq_id)
        if src == "afdb" and r["afdb_archive"]:
            return "afdb", _from_tar(r["afdb_archive"], r["afdb_file"], gz=True)
        if src == "pdb" and r["pdb_archive"]:
            return "pdb", _from_tar(r["pdb_archive"], r["pdb_file"], gz=True)
    return "none", None

def _from_tar(archive, member, gz=False):
    import gzip
    path = hf_hub_download(REPO, archive, repo_type="dataset")
    with tarfile.open(path) as tar:                    # plain tar, members are .gz
        raw = tar.extractfile(member).read()
    return gzip.decompress(raw).decode() if gz else raw.decode()

def _from_zst(archive, seq_id):
    path = hf_hub_download(REPO, archive, repo_type="dataset")
    with open(path, "rb") as fh:
        stream = zstandard.ZstdDecompressor().stream_reader(fh)
        with tarfile.open(fileobj=stream, mode="r|") as tar:
            for m in tar:
                if m.isfile() and m.name == f"{seq_id}/{seq_id}_model.cif":
                    return tar.extractfile(m).read().decode()
    return None

The AFDB and PDB archives are plain .tar holding already-gzipped .cif.gz members; the AF3 archives are .tar.zst holding whole output directories. Both are bundled by shard because 350,000+ loose objects would make the repo unusable.

Provenance — read before combining

The three sources differ in method, and which source a sequence gets is not random:

• Reconstructed ancestral sequences have no UniProt accession, so they can never have an AFDB model. They are predicted or nothing.
• Well-studied proteins are the ones with PDB entries.
• The 6,540 AF3 predictions are, by construction, the sequences with neither a PDB entry nor an AFDB model — i.e. the thinnest-homology corner of the dataset.

So source correlates with sequence properties. If a model trained on the union behaves differently on AF2-labelled versus AF3-labelled sequences, that is confounded with those properties and not attributable to the structures alone. Record which source each label came from — the index's best column does this — and check it as a covariate rather than assuming exchangeability.

Version details worth recording: AFDB files are model v6 (not the v4 many papers cite), and the AF3 predictions use v3.0.4 with stock settings, seed 20260916, 5 diffusion samples, 10 recycles, full 4-database MSA and template search. Per-shard details are in predictions/missing_pdb_afdb/README.md.

Known gaps

Issue	Count	Handling
AFDB accession linked but no model exists at any version	1,508	in gpbridge/afdb_no_model_needs_af3.fasta; not in any fetch list and not in missing_pdb_afdb_training.fasta either — they need prediction
PDB entries withdrawn by RCSB	7	fetched into gpbridge/pdb_obsolete/, flagged in the index's note
No structure from any source	11,075	best = none

The 1,508 are worth emphasising: the source bundle's coverage table counts them as having an AlphaFold DB model. They do not — they are live UniProt entries of ordinary length returning 404 on both v6 and v4. Nothing in the original pipeline covered them.

Not part of this dataset

The repository also holds predictions/00001_00256/ and predictions/00257_00512/. These are a different dataset - the GO + ASR catalog (292,804 sequences: Gene Ontology couplings and ancestral sequence reconstructions), predicted with the same stock AlphaFold 3 settings but unrelated to the GPBridge expanded-GO set documented here.

They are deliberately absent from structure_index.tsv, which covers only the 322,992 GPBridge sequences. Do not join them by seq_id and assume membership: both datasets use seq_<sha256(sequence)[:20]> ids, so an id collision means the two sets genuinely share that exact sequence, not that the row belongs here. That campaign is partially complete and its shards are named by length band.

Layout

structure_index.tsv                      every sequence -> its file(s)
gpbridge/join.tsv.gz                     original join table from the source bundle
gpbridge/coverage.csv                    source bundle's coverage, all 13 populations
gpbridge/BUNDLE_README.md                source bundle's own documentation
gpbridge/afdb/<2-char>.tar               211 archives, AF2 model v6
gpbridge/pdb/<2-char>.tar              1,088 archives, experimental
gpbridge/pdb_obsolete/                     7 withdrawn entries
gpbridge/afdb_no_model.tsv               the 1,508, with their 404s
gpbridge/afdb_no_model_needs_af3.fasta   their sequences
predictions/missing_pdb_afdb/            AF3 output + its own README and index
predictions/00001_00256/                 GO+ASR catalog - a DIFFERENT dataset
predictions/00257_00512/                 GO+ASR catalog - a DIFFERENT dataset

Completeness

predictions/missing_pdb_afdb/ is still being produced: 6 of 14 archives are published at the time of writing. structure_index.tsv lists all 6,540 of those sequences regardless, because the index describes the dataset, not the current upload state. Check which archives exist before assuming a sequence is retrievable:

from huggingface_hub import HfApi
have = {x.rfilename for x in HfApi().repo_info("tonynzh2/afdb",
                                               repo_type="dataset").siblings}
missing = {s: r for s, r in idx.items()
           if r["af3_archive"] and r["af3_archive"] not in have}

The AFDB and PDB portions are complete: 211 and 1,088 archives respectively.

Terms of use

AlphaFold 3 output in predictions/ is subject to the AlphaFold 3 Output Terms of Use — non-commercial, and redistribution must carry the terms. AlphaFold DB structures are CC-BY-4.0 from EMBL-EBI. PDB entries are in the public domain.

Cite AlphaFold 3: Abramson, J. et al. Nature 630, 493–500 (2024). Cite AlphaFold DB: Varadi, M. et al. Nucleic Acids Research (2024).