pdb_cpl/ held 1,556 uncompressed PDBs that are a strict subset of cpl_structures.tar.gz: five of
the seven clones, byte-identical to their entries in the archive, 392 MB against the archive's 127.
Verified before removal -- no file in the loose folder is absent from the archive, and sampled
contents hash the same.
It was not only redundant. A producer that globbed the loose folder instead of untarring silently
benchmarked 5 of 7 clones and reported a clean run, which is how the duplicate was found.
content.md now gives the untar command, the per-clone best/worst counts against the assay table's
rows, and the reason the archive holds 2,103 structures where the assay table has 2,102: sb27's own
cognate epitope was modelled after the library deposit and joins to no assay row.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deposit only what the paper reports: the S8 panel and the HPVGEADYFEY control
vdjdb_binder_benchmark is cut from 1,523 structures over 38 epitopes to the
1,089 in the 22 well-powered cohorts that Table 4, Fig. 6 and Supplementary
Figs. S8/S10 actually report (523 positives / 566 negatives, 7 HLA alleles).
The 434 dropped rows sat in cohorts too small or too one-sided to enter any
reported average, so the folder and the paper now describe the same set;
metadata.tsv is filtered to match and in_eval/well_powered are true throughout.
vdjdb_mut2023_benchmark is new: HPVGEADYFEY / HLA-B*35:01 only, 81 structures
-- 4 crystals (3mv7, 3mv8, 3mv9, 4pri), 56 Rosetta mutants descending from
them, 21 TCRmodel2 mock decoys -- with the three scores the figure plots, the
receptor junctions (crystal junctions read off their own coordinates, since the
2023 archive annotates only its models) and the mutation each model carries.
Its README says plainly that it is a control and not a benchmark: every
trajectory converges on one receptor, so the positive side has an effective n
of …
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.
cpl: hold 7Q9B as PDB, and correct 868's cognate in the notes
7Q9B (MEL8 TCR / HLA-A*02 / EAAGIGILTV) was the only entry in the native
archive held solely as mmCIF, which made it invisible to a *.pdb glob and to a
fixed-column ATOM parser. Replaced with the PDB-format copy from Native2026;
same receptor (CDR3a AVQKLV, CDR3b ASSYSFTEATYEQY), one copy in the AU instead
of two. Archive stays at 27 entries.
content.md: 868's cognate is SLYNTVATL, not SIFLFIARV, and the per-clone
cognate-complex map is now recorded here so a consumer does not have to
re-derive it from peptide strings -- which is not safe, since EAAGIGILTV occurs
both in MEL8's 7Q9B and in 4QOK, whose receptor is MEL5.
cpl: add the sb27 cognate model, completing 7 of 7 clones
sb27's own cognate 13-mer LPEPLPQGQLTAY was the one clone epitope absent from
the peptide-swap library, so the deposit covered the cognate of 6 of the 7
clones. Modelled on the same receptor and MHC coordinates as the other 161
sb27 models (chains A/B/D identical, only the peptide differs).
cpl_structures.tar.gz 2102 -> 2103 PDB (sb27_best/LPEPLPQGQLTAY.pdb)
CPL_stats/cpl_data_stats_new_struct.tsv 2101 -> 2102 rows (cpl_status 1)
Source: aldan3 /projects/structures/CPL_data_to_hf/, 2026-07-26.
vdjdb_binder_benchmark: fold ranking_confidence and pLDDT into metadata.tsv
The deposit shipped ipTM alone, so reproducing the receptor result also needed
vdjdb_structure_models and vdjdb_negatives, each with its own key normalisation.
Three locations for one benchmark is an arrangement that rots quietly: the pools
can change without the deposit noticing. This folder is now the only input.
Join checked biologically, not by row count -- the epitope each confidence row
carries must equal the epitope the benchmark assigns that structure (0
disagreements over 1,523) -- and the pre-existing ipTM column verified against
its own upstream source (max |difference| 0.00e+00). The downstream recompute is
bit-identical across all 117 columns after the switch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Add vdjdb_binder_benchmark: final receptor-task set (real binders vs mock)
583 real VDJdb positives (1:1 from the 15k pool, TCRvdb-excluded) + 940 mock
negatives, re-featurised with tcren. metadata.tsv carries labels, ipTM, tcren
scores, Q_geom, combo, and the covered/uncovered (native_structure_exists) flag.
Reproduces the manuscript C42 template split exactly (covered 6 epi combo
0.752/0.737 vs ipTM 0.692/0.693). See vdjdb_binder_benchmark/README.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
vdjdb pos/neg: expand 997->1000 + TCR-annotated meta; drop junk generation_method
Fetched the current 1000/1000 VDJdb real+mock TCRmodel2 complexes from aldan3
(/projects/structures/VDJdb_negatives/VDJdb_{positives,negatives}_all_pdb), superseding
the 997/997 tarballs. The 997 prior structures are byte-identical; +3 new each side.
Meta replaced with the *_TCR_added versions (per-chain TRA/TRB annotation added), which
also drop the generation_method column (a junk data-join artifact, not a real
processing distinction: positives and negatives are the same raw unrelaxed TCRmodel2
pipeline). Removed the stray root-level negative_samples_meta.tsv duplicate.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
cleanup: remove 997 stray root-level negative PDBs (mis-published to root; kept in vdjdb_negatives.tar.gz)
Mirrors Models_VDJdb.tar.gz (3,240 PDBs: 1,605 min + 1,606 repacked + 29 cleaned crystal templates) plus the mutation annotation and Rosetta total interface energies from Shcherbinin et al. 2023, Front Immunol 10.3389/fimmu.2023.1224969.
Structures and energies are COMPUTED; no experimental binding ground truth.
Upload cpl/cpl_native_structures.tar.gz with huggingface_hub
All structures placed in the canonical TCR-pMHC frame (PCA: z=MHC->TCR, y=peptide,
x=thin axis) with chains renamed A=Valpha B=Vbeta C=peptide D=MHCa E=MHCb/b2m. Native2026
is the comprehensive 2026 TCR:pMHC set (374 originals, peptide-preserving, geometry-
validated). The lossy renumbered Tcr3d2026 CIFs are removed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Native2022: restore original 243 paper-era set; add Tcr3d2026 (TCR3D)
Native2022 reverts to the original 2022 paper PDB database (243 structures,
repo commit 840b843) after an over-broad expansion. The newer/expanded structures
are represented by the 2026 TCR3D set, added as Tcr3d2026/ (353 renumbered CIFs).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Native2022: expand to full 395-structure PDB set (paper superset)
Replace the 163-structure subset with the complete data/PDB_structures set
(239 paper-era + post-2022 additions), enabling TCRen rederivation on the full
non-redundant set. Add Native2022/*.pdb LFS glob.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
