# CPL — combinatorial peptide library structures (cpl_structures.tar.gz)

7 TCR clones (1e6, 4c6, 868, ila1, mel5, mel8, sb27), best- vs worst-peptide AlphaFold C-beta peptide-swap models — 2103 PDB files (per-clone best/worst dirs inside the archive).

All seven clones now carry a model of their OWN cognate epitope in the best half:
1e6/ALWGPDPAAA, 4c6/LWMRLLPLL, 868/SLYNTVATL, ila1/ILAKFLHWL, mel5/ELAGIGILTV, mel8/EAAGIGILTV,
sb27/LPEPLPQGQLTAY. The sb27 one (added 2026-07-27) was modelled after the other 2102 on the same
receptor and MHC coordinates — chains A/B/D are byte-for-byte the other sb27 models', only the
13-mer peptide differs — but it is NOT part of the original library deposit, so anything that
counts "structures per clone" should mark it separately.

868's cognate is SLYNTVATL (HIV-1 Gag p17 SL9), NOT SIFLFIARV. SIFLFIARV is one of its own library
peptides and was recorded as the cognate until 2026-07-27; because the clone → crystal lookup was
keyed on that string it matched nothing, and the pipeline reported that 868 had no deposited
crystal. It has four: 5NME (the index complex, SLYNTVATL), 5NMD (unligated receptor), 5NMF (6I) and
5NMG (3F6I8V).

# cpl_native_structures.tar.gz — the deposited complexes

27 RCSB entries covering all seven clones. The cognate complex per clone is
1e6→3UTS, 4c6→7NMG, 868→5NME, ila1→5MEN, mel5→3HG1, mel8→7Q9B, sb27→2AK4; the rest are the same
receptors on variant peptides. MATCH ON THE RECEPTOR (CDR3α/CDR3β pair), never on the epitope: a
peptide identifies neither side uniquely, and EAAGIGILTV in particular occurs both in MEL8's own
7Q9B and in 4QOK, whose receptor is MEL5. 7Q9B is held here as PDB (transferred from Native2026);
it was previously mmCIF-only, which made it invisible to a `*.pdb` glob.

Peptide-ranking benchmark: AUC of -TCRen(TCR:peptide), best(+) vs worst(-), per clone and pooled.
Provenance: structures COMPUTED (AF peptide-swap); best/worst labels from EXPERIMENTAL combinatorial peptide-library positional scans (Sewell/Cole lab; see DATASETS.md per-clone references).
Source: aldan3 /projects/structures/CPL_data_to_hf/pdb_cpl/.

# How to use this folder

```bash
mkdir -p cpl_pdb && tar -xzf cpl_structures.tar.gz -C cpl_pdb
find cpl_pdb -name '*.pdb' | wc -l          # 2103
ls cpl_pdb                                   # 14 dirs: <clone>_{best,worst}, 7 clones
```

The archive is the only copy. A loose, uncompressed `pdb_cpl/` folder used to sit beside it holding
1,556 of the same files -- five clones out of seven, byte-identical to their entries in the archive.
It was removed on 2026-08-23. It cost 392 MB against the archive's 127 MB, it duplicated data the
dataset already had, and it was worse than merely wasteful: a script that globbed the loose folder
instead of untarring silently benchmarked 5 of the 7 clones and reported a clean run. If you want
the structures, untar.

# Counts, and why 2103 is not 2102

| clone | best | worst | total | assay rows |
|---|--:|--:|--:|--:|
| 1e6  | 161 | 164 | 325 | 325 |
| 4c6  | 161 | 160 | 321 | 321 |
| 868  | 161 | 160 | 321 | 321 |
| ila1 | 161 | 164 | 325 | 325 |
| mel5 | 161 | 164 | 325 | 325 |
| mel8 | 164 |  96 | 260 | 260 |
| sb27 | 162 |  64 | 226 | 225 |
| **total** | **1131** | **972** | **2103** | **2102** |

`CPL_stats/cpl_data_stats_new_struct.tsv` carries one row per (clone, peptide) -- 2,102 rows, no
duplicates -- with the graded activation score `cpl_score`, the best/worst half, and the TCRmodel2
confidences. The single extra structure is sb27's own cognate epitope LPEPLPQGQLTAY, modelled after
the library deposit and therefore absent from the assay table; it joins to nothing and any count of
"structures per clone" should say so.