20eddibae/sceptr-tcr-benchmark

CLEST-TCRdist benchmark data

The VDJdb snapshot behind the numbers in the CLEST-TCRdist paper, re-hosted so the benchmark reproduces exactly. VDJdb grows between releases, and few-shot AUROC moves with the size and composition of the negative pool, so a result quoted against "VDJdb" with no snapshot is not reproducible. Pinning it here is the point of this repository.

Nothing here is new data. Every receptor comes from VDJdb, which is public and citable in its own right, and all of it was already de-identified by the studies VDJdb aggregates.

Files

File	Rows	What it is
finetune/vdjdb_full.txt	139,744	The raw VDJdb TSV, 35 columns, all species. The paper protocol curates this at load time.
tcr_data/preprocessed/benchmarking/train_filtered.csv	42,292	Cleaned paired αβ receptors across 20 epitopes. The fine-tuning split, and the negative pool of the leak-free protocol.
tcr_data/preprocessed/benchmarking/valid_filtered.csv	4,688	Held out of fine-tuning. The positive pool of the leak-free protocol.

Paths are kept as they were in the project tree so the loader drops each file where the benchmark expects it.

Curation applied to the raw TSV

The paper protocol builds its evaluation set from vdjdb_full.txt at load time, following Nagano et al. 2025: keep human entries, drop anything traceable to the 10x Genomics whitepaper (known dextramer false positives, which removes several otherwise well-populated epitopes), require both CDR3α and CDR3β, standardise V genes to IMGT with tidytcells and drop receptors whose gene symbol cannot be resolved to an allele, then deduplicate on (CDR3α, CDR3β, epitope). That leaves 53,269 paired receptors across 1,455 epitopes.

The six benchmark epitopes are GILGFVFTL (influenza A M1), NLVPMVATV (CMV pp65), SPRWYFYYL, TFEYVSQPFLMDLE, TTDPSFLGRY and YLQPRTFLL (SARS-CoV-2).

Use it

from clest_tcrdist import download_benchmark_data
paths = download_benchmark_data(dest="data/benchmark")

Then reproduce the headline result:

python -m clest_tcrdist.benchmark --pretrained --data-dir data/benchmark

That gives mean AUROC over the six epitopes of 0.701, 0.765, 0.794 and 0.814 at 1, 10, 50 and 200 reference TCRs. Code and the full protocol are at github.com/20eddibae/CLEST-TCRDist.

One caveat about this pool

It is larger and more current than the one the reference protocol was defined on, and that is not automatically in a new method's favour. Of the 39,080 background receptors, 29,719 carry a single specificity (SLLMWITQV), deposited mostly by one study that postdates the protocol. The nominal 1,449 background specificities have a median size of one receptor. Concentration like that affects how much averaging over a reference set helps, so numbers from this pool and numbers from the original 7,168-receptor curation are not interchangeable.

Citation

Cite VDJdb (Bagaev et al., Nucleic Acids Research 2020) for the data itself, and the CLEST-TCRdist repository for this snapshot and the curation. CC BY 4.0, matching VDJdb's terms.