timodonnell commited on
Commit
382ad32
·
verified ·
1 Parent(s): 25b024b

Add root dataset card; correct pdb-2026-04-22 reference in snapshot README

Browse files
Files changed (2) hide show
  1. README.md +184 -0
  2. processed/pdb-2026-08-08/README.md +4 -1
README.md ADDED
@@ -0,0 +1,184 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Helico Training Data
3
+ license: other
4
+ license_name: mixed-upstream
5
+ language:
6
+ - en
7
+ size_categories:
8
+ - 100K<n<1M
9
+ tags:
10
+ - biology
11
+ - structural-biology
12
+ - protein-structure-prediction
13
+ - protein-folding
14
+ - alphafold3
15
+ - pdb
16
+ - protein
17
+ viewer: false
18
+ ---
19
+
20
+ # Helico Training Data
21
+
22
+ Preprocessed, model-ready training data for
23
+ [Helico](https://github.com/Open-Athena/helico) — an AlphaFold3 clone built from
24
+ scratch in PyTorch. The core of this repository is **236,326 structures from the
25
+ RCSB PDB**, tokenized in AF3 convention, pickled, and published as versioned
26
+ snapshots.
27
+
28
+ > ⚠️ **This is not a `datasets`-loadable dataset.** The payload is Python
29
+ > pickles inside split tar archives; `load_dataset()` will not work and the
30
+ > dataset viewer is disabled. Unpickling requires the `helico` package to be
31
+ > importable, and carries the usual pickle code-execution caveat. See
32
+ > [Getting the data](#getting-the-data).
33
+
34
+ ## Contents
35
+
36
+ | Path | What it is |
37
+ |---|---|
38
+ | [`processed/pdb-2026-08-08/`](processed/pdb-2026-08-08) | **Current snapshot.** 236,326 structures with contacts, 83.4 GB. [Full documentation →](processed/pdb-2026-08-08/README.md) |
39
+ | [`processed/pdb-2026-04-22/`](processed/pdb-2026-04-22) | Superseded, **metadata only** — see the warning below |
40
+ | `processed/ccd_cache.pkl` | Parsed Chemical Component Dictionary, 117.8 MB. Shared across snapshots; needed for inference |
41
+ | `processed/latest.json` | Names the current snapshot per source type |
42
+ | `benchmarks/FoldBench/` | Vendored [FoldBench](https://github.com/BEAM-Labs/FoldBench) evaluation suite, 2.24 GB |
43
+
44
+ ### Snapshots
45
+
46
+ Data is published as **immutable, date-stamped snapshots** rather than edited in
47
+ place. `processed/latest.json` maps a source type to the current snapshot id:
48
+
49
+ ```json
50
+ {"schema_version": 1, "sources": {"pdb": "pdb-2026-08-08"}}
51
+ ```
52
+
53
+ Each snapshot directory carries a `SOURCE.json` recording the raw data sources,
54
+ the preprocessing parameters, and the `git_sha` of the Helico commit that
55
+ produced it.
56
+
57
+ | Snapshot | Structures | Contacts | Status |
58
+ |---|---|---|---|
59
+ | `pdb-2026-08-08` | 236,326 | ✅ pyconfind `contacts-v1` | **Current** |
60
+ | `pdb-2026-04-22` | — | ❌ | ⚠️ Metadata only |
61
+
62
+ > ⚠️ **`pdb-2026-04-22` contains no structures.** Its `SOURCE.json` claims
63
+ > 236,326, but the `structures.tar.*` chunks were never uploaded — a publishing
64
+ > bug silently packed an empty directory. Only its manifest, MSA indices, and
65
+ > provenance record are present. Use `pdb-2026-08-08`.
66
+
67
+ ## Getting the data
68
+
69
+ Helico is not published to PyPI; install it from source:
70
+
71
+ ```bash
72
+ git clone https://github.com/Open-Athena/helico && cd helico
73
+ uv pip install -e ".[dev]"
74
+ ```
75
+
76
+ Then let it resolve, fetch and extract everything:
77
+
78
+ ```bash
79
+ helico-download # follows latest.json
80
+ helico-download --snapshot pdb-2026-08-08 # pin explicitly
81
+ helico-download --subset ccd-only # just the CCD cache
82
+ ```
83
+
84
+ Files land in `~/.cache/helico/data/` by default; override with `--data-dir` or
85
+ the `HELICO_DATA_DIR` env var. Note that `helico-download` deliberately flattens
86
+ the snapshot id away — the snapshot is a publishing concern, and the training
87
+ code sees a flat `processed/` layout.
88
+
89
+ To bypass the CLI, download `processed/<snapshot>/structures.tar.*` and
90
+ concatenate them in order:
91
+
92
+ ```bash
93
+ cat structures.tar.* | tar -xf -
94
+ ```
95
+
96
+ ## What's in a structure
97
+
98
+ One pickled `TokenizedStructure` per PDB entry, tokenized in AF3 convention —
99
+ one token per protein residue, one per nucleotide, one per **heavy atom** for
100
+ ligands. Each carries per-token atom names, elements, coordinates, CCD reference
101
+ coordinates, and charges, plus sparse edge lists for covalent bonds and
102
+ residue–residue contacts.
103
+
104
+ The corpus was built by parsing 252,091 mmCIF files from the RCSB archive,
105
+ dropping water and hydrogens, and keeping entries at **≤ 9.0 Å resolution** with
106
+ at least one polymer chain — **236,326 passed**.
107
+
108
+ The current snapshot additionally carries side-chain contacts computed by
109
+ [pyconfind](https://github.com/timodonnell/pyconfind) with the same
110
+ `contacts-v1` parameters that [MarinFold](https://github.com/Open-Athena/MarinFold)
111
+ uses, densified at load time into a three-state (contact / no-contact /
112
+ unknown) token × token matrix. This supports contact-conditioned, MSA-free
113
+ folding.
114
+
115
+ **[→ Full field-by-field documentation, provenance, and generation pipeline](processed/pdb-2026-08-08/README.md)**
116
+
117
+ ## Train/val split
118
+
119
+ The split is applied at load time from `release_date` in the manifest — it is
120
+ not baked into the files. Helico's defaults match AF3, Protenix v1, and
121
+ OpenFold3-preview2 so metrics are directly comparable:
122
+
123
+ | Split | Rule | Count |
124
+ |---|---|---|
125
+ | Train | `release_date < 2021-09-30` | 170,926 |
126
+ | Val | `2022-05-01 ≤ release_date ≤ 2023-01-12` | 9,716 |
127
+
128
+ Structures in the 2021-09-30 → 2022-05-01 gap are in neither split — AF3's
129
+ deliberate leakage-prevention design.
130
+
131
+ > ⚠️ **A temporal split is not a redundancy split.** 38.2% of validation
132
+ > structures share at least one chain sequence *verbatim* with training, and
133
+ > 18.4% share every chain sequence. Any holdout built from these files by date
134
+ > needs an explicit sequence-identity filter before it will support a
135
+ > generalization claim. Details in the
136
+ > [snapshot README](processed/pdb-2026-08-08/README.md#-a-temporal-split-is-not-a-redundancy-split).
137
+
138
+ Note also that no cluster-based weighted sampling, deduplication, or
139
+ molecule-type rebalancing is applied — the corpus is published as-is.
140
+
141
+ ## MSAs
142
+
143
+ Precomputed alignments are **not hosted here** — the two `*_msa_index.pkl` files
144
+ in each snapshot are byte-offset maps enabling O(1) random reads into archives
145
+ you obtain separately:
146
+
147
+ - `rcsb_raw_msa.tar` (131 GB) — `https://boltz1.s3.us-east-2.amazonaws.com/rcsb_raw_msa.tar`
148
+ - `openfold_raw_msa.tar` (88 GB) — `https://boltz1.s3.us-east-2.amazonaws.com/openfold_raw_msa.tar`
149
+
150
+ Lookup is content-addressed by sequence hash, not PDB ID. The OpenFold index is
151
+ published for completeness but is not usable through Helico's current lookup
152
+ path (its members are keyed by UniProt accession, and that mapping is not
153
+ implemented).
154
+
155
+ ## Benchmarks
156
+
157
+ `benchmarks/FoldBench/` is a vendored checkout of the
158
+ [FoldBench](https://github.com/BEAM-Labs/FoldBench) evaluation suite (cloned
159
+ 2026-03-04, ground-truth CIFs from 2025-05-20), including precomputed MSAs for
160
+ its targets and vendored copies of baseline algorithm code. It is bundled for
161
+ reproducible benchmarking; FoldBench is MIT licensed and carries its own
162
+ `LICENSE`, as does each vendored algorithm. See the FoldBench paper in
163
+ [Nature Communications](https://doi.org/10.1038/s41467-025-67127-3).
164
+
165
+ ## Provenance and terms
166
+
167
+ This repository aggregates data from several upstream sources under differing
168
+ terms, which is why no single license is declared:
169
+
170
+ | Component | Origin |
171
+ |---|---|
172
+ | Structures | [RCSB PDB](https://www.rcsb.org/) — archive distributed without copyright restriction (CC0 1.0) |
173
+ | Chemical Component Dictionary | [wwPDB](https://www.wwpdb.org/) |
174
+ | MSA archives (indexed, not hosted) | Published by [Boltz](https://github.com/jwohlwend/boltz); OpenFold alignments derive from [OpenFold](https://github.com/aqlaboratory/openfold) |
175
+ | Contacts | [pyconfind](https://github.com/timodonnell/pyconfind) |
176
+ | `benchmarks/FoldBench/` | [FoldBench](https://github.com/BEAM-Labs/FoldBench), MIT; includes vendored third-party algorithm code under its own licenses |
177
+
178
+ If you use this data, please cite the underlying PDB entries and the upstream
179
+ projects above. Usage is governed by the terms of those sources.
180
+
181
+ ## Related
182
+
183
+ - **Code and training recipes:** [`Open-Athena/helico`](https://github.com/Open-Athena/helico)
184
+ - **Model weights:** [`timodonnell/helico`](https://huggingface.co/timodonnell/helico)
processed/pdb-2026-08-08/README.md CHANGED
@@ -337,4 +337,7 @@ usage follows the terms of those upstream sources.
337
 
338
  - Code, training recipes, and full documentation: [`Open-Athena/helico`](https://github.com/Open-Athena/helico)
339
  - Model weights: [`timodonnell/helico`](https://huggingface.co/timodonnell/helico)
340
- - Previous snapshot: [`pdb-2026-04-22`](../pdb-2026-04-22) — same corpus, without contacts
 
 
 
 
337
 
338
  - Code, training recipes, and full documentation: [`Open-Athena/helico`](https://github.com/Open-Athena/helico)
339
  - Model weights: [`timodonnell/helico`](https://huggingface.co/timodonnell/helico)
340
+ - Previous snapshot: [`pdb-2026-04-22`](../pdb-2026-04-22) — ⚠️ **metadata only.** Its
341
+ `SOURCE.json` claims 236,326 structures, but no `structures.tar.*` was ever
342
+ uploaded (a publishing bug that silently packed nothing). Use this snapshot
343
+ instead; `pdb-2026-04-22` is retained only for its manifest and provenance record.