Add retrieval README and shard index for missing_pdb_afdb

predictions/missing_pdb_afdb/README.md

@@ -0,0 +1,148 @@
+# missing_pdb_afdb_training — AlphaFold 3 predicted structures
+
+Structures for the **6,540 sequences** in `missing_pdb_afdb_training.fasta`: the
+training-split subset of the GPBridge expanded-GO dataset that has neither a PDB
+entry nor an AlphaFold DB model, so it had to be predicted rather than fetched.
+
+Sequence ids are `seq_<sha256(sequence)[:20]>`, the same convention as the rest
+of the dataset, so a structure always maps back to its exact sequence.
+
+> **Status: in progress.** Predictions are published per shard as they finish.
+> `shard_index.tsv` lists all 6,540 sequences; check which archives actually
+> exist before assuming a sequence is available.
+
+## Layout
+
+```
+predictions/missing_pdb_afdb/part_00000.tar.zst   500 sequences
+...
+predictions/missing_pdb_afdb/part_00013.tar.zst    40 sequences
+shard_index.tsv                                    seq_id -> archive, length, split
+```
+
+Shards hold 500 sequences each and are **grouped by length**, not by any
+biological property — AlphaFold 3 recompiles per length bucket, so keeping a
+shard length-homogeneous amortises compilation. The consequence is that a part
+number tells you nothing except size, which is what `shard_index.tsv` is for.
+
+Files are bundled per shard rather than stored loose: 6,540 predictions is
+~46,000 files, and a repo of loose objects at that scale is unusable.
+
+## Get the structure for a sequence
+
+```python
+import csv, io, json, tarfile, zstandard
+from huggingface_hub import hf_hub_download
+
+REPO = "tonynzh2/afdb"
+
+idx = {r["seq_id"]: r["archive"] for r in
+       csv.DictReader(open(hf_hub_download(REPO, "predictions/missing_pdb_afdb/shard_index.tsv",
+                                           repo_type="dataset")), delimiter="\t")}
+
+def get(seq_ids):
+    """Yields (seq_id, cif_text, summary_dict). Groups by archive so each is read once."""
+    by_archive = {}
+    for s in seq_ids:
+        by_archive.setdefault(idx[s], set()).add(s)
+    for archive, wanted in by_archive.items():
+        path = hf_hub_download(REPO, archive, repo_type="dataset")
+        cif, summary = {}, {}
+        with open(path, "rb") as fh:
+            stream = zstandard.ZstdDecompressor().stream_reader(fh)
+            with tarfile.open(fileobj=stream, mode="r|") as tar:   # streaming, low memory
+                for m in tar:
+                    if not m.isfile() or "/seed-" in m.name:
+                        continue
+                    sid = m.name.split("/")[0]
+                    if sid not in wanted:
+                        continue
+                    if m.name.endswith("_model.cif"):
+                        cif[sid] = tar.extractfile(m).read().decode()
+                    elif m.name.endswith("_summary_confidences.json"):
+                        summary[sid] = json.load(tar.extractfile(m))
+        for sid in cif:
+            yield sid, cif[sid], summary.get(sid, {})
+```
+
+Whole archive from the shell:
+
+```bash
+hf download tonynzh2/afdb predictions/missing_pdb_afdb/part_00000.tar.zst \
+    --repo-type=dataset --local-dir .
+zstd -dc part_00000.tar.zst | tar -xf -     # GNU tar <1.31 has no --zstd
+```
+
+## What is in each prediction
+
+```
+seq_<id>/
+├── seq_<id>_model.cif                  top-ranked structure  <- use this
+├── seq_<id>_summary_confidences.json   ptm, ranking_score, has_clash, ...
+├── seq_<id>_confidences.json           per-atom pLDDT, PAE matrix
+├── seq_<id>_data.json                  the MSAs and templates used as input
+├── seq_<id>_ranking_scores.csv         score per diffusion sample
+├── seed-20260916_sample-{0..4}/        all 5 samples, each with its own cif
+└── TERMS_OF_USE.md
+```
+
+`_data.json` is most of the bytes — it embeds a 50,000-row UniProt MSA. Drop it
+unless you need to reproduce or re-run the input. Per-residue pLDDT is in the
+B-factor column of the mmCIF.
+
+## Filter on confidence
+
+These are predictions, not experimental structures, and the spread is wide:
+in spot checks `ptm` ranged 0.19 to 0.85 within a single shard, tracking MSA
+depth. From `*_summary_confidences.json`:
+
+| field | use |
+|---|---|
+| `ranking_score` | AlphaFold 3's own overall ranking — the primary filter |
+| `ptm` | fold-level confidence |
+| `has_clash` | drop anything non-zero |
+| `fraction_disordered` | high values mean much of the chain is unstructured |
+
+Look at the distribution before choosing a cutoff rather than inheriting one.
+On one measured shard, `ranking_score >= 0.7` with no clash kept 97%, and
+`>= 0.8` kept 74%.
+
+## Protocol
+
+Stock open-source AlphaFold 3, defaults throughout — no protocol deviations.
+
+| Setting | Value |
+|---|---|
+| Code | google-deepmind/alphafold3 v3.0.4 (`c0f97ed`) |
+| Genetic search | full standard pipeline: UniRef90, MGnify, small BFD, UniProt |
+| Templates | searched (PDB seqres + mmCIF, `max_template_date` 2021-09-30) |
+| Diffusion samples | 5 (default) |
+| Recycles | 10 (default) |
+| Model seed | 20260916 |
+| Databases | the versions used in the AlphaFold 3 paper |
+
+Both MSAs AlphaFold 3 builds are kept — unpaired and paired (UniProt). The
+paired MSA is retained even though every input is a single chain, because
+AlphaFold 3 featurises it as `msa_all_seq` regardless of chain count and its
+presence also changes the unpaired crop size; omitting it would not be the
+standard protocol.
+
+## Related
+
+- `gpbridge/join.tsv.gz` — ties every structure in this repo back to a dataset
+  sequence. Nothing in the filenames does.
+- `gpbridge/afdb/`, `gpbridge/pdb/` — the fetched AlphaFold DB and PDB
+  structures for the rest of the dataset.
+- `gpbridge/afdb_no_model_needs_af3.fasta` — 1,508 sequences the source bundle
+  counted as having an AlphaFold DB model that in fact has none. They are not in
+  `missing_pdb_afdb_training.fasta` either, so they also need prediction.
+
+## Terms of use
+
+These are AlphaFold 3 Output, subject to the
+[AlphaFold 3 Output Terms of Use](../../OUTPUT_TERMS_OF_USE.md) — non-commercial
+use, and redistribution must carry the terms. Each archive also contains
+AlphaFold 3's own `TERMS_OF_USE.md`.
+
+Cite: Abramson, J. et al. Accurate structure prediction of biomolecular
+interactions with AlphaFold 3. *Nature* **630**, 493–500 (2024).