immrep23_negatives: 832 IMMREP decoy models, and provenance READMEs for the VDJdb arms

The receptor benchmark caps most epitopes at 20-35 decoys while the VDJdb positive
pool holds hundreds to thousands, so the balanced cohort size is set by the decoys.
This adds a second, independent decoy source: 832 of a planned 2,985 TCRmodel2 models
of IMMREP mismatched TCR-epitope pairings (CINGVCWTV 299, ATDALMTGF 278, GILGFVFTL
255), fetched from aldan3 while the run is still producing. All 832 verified
md5-identical to the cluster copy.

No TCR appears against more than one epitope, and none of the 778 distinct
(CDR3beta, epitope) pairs occurs as a VDJdb positive -- so the decoys neither
cross-contaminate cohorts nor conflict with the positive labels.

vdjdb_positives/ and vdjdb_negatives/ get the READMEs they never had: what each arm
is, that the decoys are modelled de novo rather than peptide-swapped, the +0.028 ipTM
gap between the arms, and the 1,000 AppleDouble stubs a naive *.pdb glob picks up.

README.md

@@ -23,6 +23,8 @@ no notebooks, figures, or analysis tables.
 | `Native2022/`, `PolyV2022/` | 2022-paper reproduction | structure sets (`.gz`) |
 | `tcrvdb/` | TCR-ranking (specificity) | 618 TCRmodel2 models (`<hash>.pdb`) |
 | `vdjdb_binder_benchmark/` | TCR-ranking (real binders vs mock) | 1,523 PDBs (`.tar.gz`) + `metadata.tsv` (labels, ipTM, tcren scores, covered/uncovered flag); see its `README.md` |
+| `vdjdb_positives/`, `vdjdb_negatives/` | TCR-ranking source arms | 1,000 real VDJdb pairs + 1,000 mismatched decoys, same TCRmodel2 batch; each has a `README.md` |
+| `immrep23_negatives/` | TCR-ranking, extra decoys | 832 of 2,985 modelled IMMREP mismatched pairings (run in progress); lifts the per-epitope decoy ceiling |
 | `cpl/pdb_cpl/` | peptide-ranking (epitope) | peptide-swap best/worst, 5 TCRs |
 | `as_case/` | MHC-ranking (B\*27:05 vs :02) | `as_case.tar.gz` + `content.md` |
 | `Bobisse/`, `Bigot/` | neoantigen | cohort structures (`.gz`) |

immrep23_negatives/README.md

@@ -0,0 +1,61 @@
+# immrep23_negatives — modelled IMMREP non-binder TCR:pMHC complexes
+
+TCRmodel2 models of **IMMREP mismatched TCR–epitope pairings** (`label = 0`), built to supply
+*additional negatives* for the receptor-ranking task (`vdjdb_binder_benchmark/`), whose mock-negative
+pool caps most epitopes at 20–35 decoys while the VDJdb positive pool holds hundreds to thousands.
+
+> ⚠ **Snapshot of a run still in progress.** 832 of a planned **2,985** complexes were modelled when
+> this was fetched (**2026-07-27**). The remaining ~2,150 are being generated on aldan3; re-pull to
+> extend. The folder name follows the project's `immrep23` label; every upstream file is stamped
+> `2022` (`immrep_2022_negatives/`, `immpred2022_negative_data_subsampled_generation.tsv`) — the
+> naming difference is upstream's, not a second dataset.
+
+## Contents
+
+| file | what |
+|---|---|
+| `immrep23_negatives.tar.gz` | **832** PDBs, flat, `<CDR3α>_<CDR3β>_<epitope>_<V/J genes>.pdb` |
+| `immrep2022_negatives.tsv` | one row per **modelled** structure (832): pairing, chain sequences, MHC, and the TCRmodel2 metrics `ranking_confidence, plddt, ptm, iptm, tcr-pmhc_iptm` |
+| `immpred2022_negative_data_subsampled_generation.tsv` | the **2,985-row generation manifest** (`index_seq`, TRA/TRB/MHC-α sequences, epitope) — the full target list, i.e. what the run will contain when it finishes |
+
+Chains are already canonical, matching `vdjdb_positives/` and `vdjdb_negatives/`:
+**A** = TCRα, **B** = TCRβ, **C** = peptide, **D** = MHC α (+ no B2M chain).
+
+## Modelled so far (832)
+
+| epitope | MHC | n | mean ipTM | in the 2,985-row target |
+|---|---|--:|--:|--:|
+| CINGVCWTV | HLA-A\*02:01 | 299 | 0.777 | 501 |
+| ATDALMTGF | HLA-A\*01:01 | 278 | 0.756 | 306 |
+| GILGFVFTL | HLA-A\*02:01 | 255 | 0.766 | 1200 |
+
+The manifest targets **16 epitopes** in total; the 13 not yet started are `GLCTLVAML` (166),
+`NLVPMVATV` (149), `TTDPSFLGRY` (83), `KSKRTPMGF` (78), `YLQPRTFLL` (75), `LLWNGPMAV` (156),
+`LTDEMIAQY` (73), `GPRLGVRAT` (46), `SPRWYFYYL` (44), `HPVTKYIM` (42), `RAQAPPPSW` (19),
+`TPRVTGGGAM` (18), `NQKLIANQF` (13).
+
+## Provenance
+
+- **Origin of record**: aldan3 `/projects/structures/immrep_2022_negatives/` — `all_pdbs/` (structures +
+  `immrep2022_negatives.tsv`) and the generation manifests in the parent directory. The per-complex
+  AlphaFold output trees (688 GiB) stay on the cluster; only the retained model per complex is here.
+- **Pairings are COMPUTED, not measured.** Rows are IMMREP benchmark *train-split* negatives — a real
+  TCR re-paired with an epitope it was not observed against. They are **not** experimentally verified
+  non-binders. Every row carries `label = 0`, `source = train`.
+- **Structures are COMPUTED** (TCRmodel2/AlphaFold, one retained model per complex); `iptm`, `ptm`,
+  `plddt`, `ranking_confidence` are the generator's own metrics, not measurements.
+- 829 distinct TCRs over 832 rows (3 appear twice under different V/J calls); **no TCR is used against
+  more than one epitope**, so decoys do not cross-contaminate epitope cohorts.
+- **Zero label conflicts with VDJdb**: 0 of the 778 distinct (CDR3β, epitope) pairs here occur as a
+  positive in `vdjdb_structure_models`. 245 of 744 CDR3β do appear in VDJdb against *other* epitopes,
+  which is the mismatched-pairing signature.
+
+## Fetch / refresh
+
+```zsh
+aldan3 pull /projects/structures/immrep_2022_negatives/all_pdbs <local>
+aldan3 pull /projects/structures/immrep_2022_negatives/immpred2022_negative_data_subsampled_generation.tsv <local>
+```
+
+Fetched 2026-07-27; all 832 PDBs verified **byte-identical** to the cluster copy (md5 over every file,
+0 differences).

immrep23_negatives/immrep23_negatives.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b485975812cc22e3799e2cead4bd9e2e98ebb815af8e583d1c58f8f7133cd752
+size 49827680

vdjdb_negatives/README.md

@@ -0,0 +1,53 @@
+# vdjdb_negatives — modelled mismatched TCR:pMHC decoys (the matched negative arm)
+
+**1,000** TCRmodel2 models of TCR–epitope pairs that were **never observed together**: each real VDJdb
+TCR from `vdjdb_positives/` is re-paired with a mismatched epitope and the whole ternary complex is
+predicted from scratch. Same pipeline, same batch, same settings as the positives.
+
+| | |
+|---|---|
+| structures | 1,000 (`vdjdb_negatives.tar.gz`) |
+| epitopes / MHC alleles | 41 / 12, all MHC-I |
+| mean ipTM / pLDDT | 0.755 / 92.52 (positives: 0.783 / 92.97) |
+| most frequent epitopes | FLYALALLL 35, ELAGIGILTV 33, RPIIRPATL 33, YLQPRTFLL 32, TTDPSFLGRY 31 |
+
+`negative_samples_meta.tsv` — as the positives, plus `original_TCR_hash`, which records the real TCR
+each decoy was built from (1,000/1,000 map to a positive) and `hla_pmhc` for the assigned pMHC.
+
+Chains: **A** = TCRα, **B** = TCRβ, **C** = peptide, **D** = MHC α. File stem is
+`<original_TCR_hash>_<epitope>_<MHC>`.
+
+## Provenance — read before describing these
+
+- **Decoys are COMPUTED, not measured.** A mismatched pairing is a *presumed* non-binder, not an
+  experimentally verified one.
+- **Structures are COMPUTED de novo** — **not peptide swaps.** The complex is predicted whole from the
+  re-paired sequences; nothing is grafted onto a positive's pose. Verified: uniform pLDDT B-factors on
+  both arms, same generation command as the positives.
+- **No relaxation confound.** Both arms are raw and unrelaxed. The old `generation_method` column
+  (`TCRmodel_no_relax`) was data-join metadata, never a processing distinction, and was dropped
+  2026-07-24.
+- **Origin of record**: aldan3 `/projects/structures/VDJdb_negatives/VDJdb_negatives_all_pdb/`.
+  Reconciled against the Hub 2026-07-23: filename sets, meta TSV md5 and sampled contents all identical.
+
+```zsh
+aldan3 pull /projects/structures/VDJdb_negatives/VDJdb_negatives_all_pdb <local>
+```
+
+> ⚠ **The tarball carries 1,000 macOS AppleDouble stubs** (`._<name>.pdb`) alongside the 1,000 real
+> PDBs — filter `basename` startswith `._`, or extract with `--exclude '._*'`.
+
+## The ipTM gap is a real confound
+
+Decoys average **0.755** ipTM against **0.783** for positives. Any evaluation that treats ipTM as a
+baseline inherits that gap, and it favours ipTM. Report it whenever these two arms are compared; see
+`vdjdb_binder_benchmark/README.md`, which quantifies it per template-coverage stratum.
+
+## Relation to the other deposits
+
+- `vdjdb_positives/` — the matched real arm (the source TCRs).
+- `vdjdb_binder_benchmark/` — the shipped receptor benchmark: 940 of these decoys plus 583 positives
+  drawn from `isalgo/vdjdb_structure_models`, re-featurised with the current `tcren`.
+- `immrep23_negatives/` — a **second, independent** decoy source (IMMREP mismatched pairings). Do not
+  mix the two generators inside one epitope cohort without saying so; one generator per epitope keeps
+  the comparison clean.

vdjdb_positives/README.md

@@ -0,0 +1,45 @@
+# vdjdb_positives — modelled real VDJdb TCR:pMHC pairs (the matched positive arm)
+
+**1,000** TCRmodel2 models of TCR–epitope pairs **observed in VDJdb**, generated by the same pipeline
+and in the same batch as `vdjdb_negatives/`. The pair is the matched real-vs-decoy arm of the
+receptor-ranking task; the decoys reuse these same TCRs (`vdjdb_negatives/…original_TCR_hash`).
+
+| | |
+|---|---|
+| structures | 1,000 (`vdjdb_positives.tar.gz`) |
+| epitopes / MHC alleles | 26 / 10, all MHC-I |
+| mean ipTM / pLDDT | 0.783 / 92.97 |
+| most frequent epitopes | NLVPMVATV 429, KLGGALQAK 342, GILGFVFTL 103, LLWNGPMAV 20, LLAGIGTVPI 15 |
+
+`positives_samples_meta.tsv` — one row per structure: `cdr3.alpha/beta`, V/J, `antigen.epitope`,
+`mhc.a/mhc.b`, full TRA/TRB/MHC chain sequences, `TCR_hash`, and the TCRmodel2 metrics
+`ranking_confidence, plddt, ptm, iptm, tcr-pmhc_iptm`. It is the TCR-annotated (`*_TCR_added`)
+version — per-chain TRA/TRB sequences added downstream of generation.
+
+Chains: **A** = TCRα, **B** = TCRβ, **C** = peptide, **D** = MHC α.
+
+## Provenance
+
+- **Pairings are EXPERIMENTAL** — real TCR:epitope specificities curated in VDJdb.
+- **Structures are COMPUTED** — TCRmodel2/AlphaFold, **raw and unrelaxed** (B-factors carry pLDDT).
+  `iptm`, `ptm`, `plddt`, `ranking_confidence` are the generator's own confidence metrics.
+- **Origin of record**: aldan3 `/projects/structures/VDJdb_negatives/VDJdb_positives_all_pdb/`.
+  Reconciled against the Hub 2026-07-23: filename sets identical, meta TSV md5 identical, sampled
+  contents identical — nothing exists on the cluster that is not here.
+
+```zsh
+aldan3 pull /projects/structures/VDJdb_negatives/VDJdb_positives_all_pdb <local>
+```
+
+> ⚠ **The tarball carries 1,000 macOS AppleDouble stubs** (`._<name>.pdb`) alongside the 1,000 real
+> PDBs. A naive `*.pdb` glob after extraction picks up 2,000 files, half of them not PDBs. Filter
+> `basename` startswith `._`, or extract with `--exclude '._*'`.
+
+## Relation to the other deposits
+
+- `vdjdb_negatives/` — the decoy arm: the *same* TCRs re-paired with mismatched epitopes.
+- `vdjdb_binder_benchmark/` — the shipped receptor benchmark. It does **not** use these 1,000 as its
+  positives: it draws them 1:1 per epitope from the larger `isalgo/vdjdb_structure_models` pool
+  (TCRvdb-excluded) so the positive count can be matched to the decoy count per epitope.
+- `immrep23_negatives/` — additional non-binder models (IMMREP mismatched pairings) that lift the
+  per-epitope decoy ceiling for a few epitopes.