Datasets:

neurips-2026-dataset
/

BenchRep-T

license: mit
pretty_name: BenchRep-T
size_categories:
  - 1K<n<10K
tags:
  - immunology
  - t-cell-receptor
  - tcr-beta
  - immune-repertoire
  - AIRR
  - immunoseq
  - repertoire-classification

BenchRep-T

BenchRep-T is a benchmark for TCRβ (T-cell receptor β-chain) immune-repertoire classification. It bundles the multi-disease Mal-ID cohort together with five external disease cohorts (sequenced on the Adaptive Biotechnologies immunoSEQ platform), providing per-specimen repertoire files, harmonized sample metadata, and auxiliary files used for the sequencing-depth scaling-law and driver-sequence experiments.

Each specimen (one sequenced repertoire) is one example; the classification target is the specimen-level disease label.

Dataset structure

BenchRep-T/
├── Mal-ID/
│   ├── metadata.tsv                             # 550 specimen annotations + labels
│   ├── repertoires/                             # 550 per-specimen TCRβ rearrangement tables (.tsv.gz)
│   ├── scaling_exp_depth_indices_max75k.json.gz # sampling indices for the depth scaling-law experiment
│   └── vdjdb_minervina_driver_seq_matches.csv   # repertoire ↔ public/driver-clone matches (driver-seq experiment)
└── immunoSEQ/
    ├── Savola_RA/         # Rheumatoid Arthritis
    │   ├── metadata.tsv
    │   └── repertoires/   # 91 .tsv.gz
    ├── Musvosvi_TB/       # Tuberculosis (progression)
    │   ├── metadata.tsv
    │   └── repertoires/   # 140 .tsv.gz
    ├── Rawat_T1D/         # Type 1 Diabetes
    │   ├── metadata.tsv
    │   └── repertoires/   # 614 .tsv.gz
    ├── Mitchell_T1D/      # Type 1 Diabetes
    │   ├── metadata.tsv
    │   └── repertoires/   # 196 .tsv.gz
    └── Emerson_CMV/       # Cytomegalovirus exposure
        ├── metadata.tsv
        └── repertoires/   # 761 .tsv.gz

Cohorts

Group	Path	Classification task	Repertoires	Label composition	Repertoire format	Source study
Mal-ID	`Mal-ID/`	Multi-disease vs. Healthy/Background	550	Healthy/Background 197, HIV 98, Lupus 64, Covid19 58, Influenza 37, T1D 96	AIRR-style rearrangement table	Zaslavsky et al. 2025
Savola_RA	`immunoSEQ/Savola_RA/`	Rheumatoid Arthritis vs. Healthy	91	RA 71, Healthy 20	immunoSEQ export	Savola et al. 2017
Musvosvi_TB	`immunoSEQ/Musvosvi_TB/`	TB Progressor vs. Controller	140	Progressor 63, Controller 77	immunoSEQ export	Musvosvi et al. 2023
Rawat_T1D	`immunoSEQ/Rawat_T1D/`	Type 1 Diabetes vs. Control	614	T1D 426, Healthy/Background 188	immunoSEQ export (reduced)	Rawat et al. 2026
Mitchell_T1D	`immunoSEQ/Mitchell_T1D/`	Type 1 Diabetes vs. Healthy	196	T1D 171, Healthy/Background 25	immunoSEQ export	Mitchell et al. 2022
Emerson_CMV	`immunoSEQ/Emerson_CMV/`	CMV+ vs. Healthy/Background	761	CMV 340, Healthy/Background 421	immunoSEQ export	Emerson et al. 2017

Files

repertoires/*.tsv.gz — one gzip-compressed, tab-separated file per specimen; each row is a TCRβ clone/rearrangement. The file stem is the specimen identifier (for the immunoSEQ cohorts it matches the specimen_label/sample_name/filename column in that cohort's metadata.tsv; Rawat_T1D file names carry a _TCRB suffix).
metadata.tsv — one row per specimen with the classification label and per-cohort annotations (schemas differ across cohorts — see below).
Mal-ID/scaling_exp_depth_indices_max75k.json.gz — precomputed read-sampling indices (up to 75k reads/specimen) defining the subsampled repertoires used in the sequencing-depth scaling-law experiment.
Mal-ID/vdjdb_minervina_driver_seq_matches.csv — matches between repertoire clones and known public / antigen-specific "driver" TCRs (VDJdb + Minervina et al.), used for the driver-sequence experiment. Columns: disease, sample_cdr3, sample_vgene, sample_jgene, public_clone_cdr3, public_clone_vgene, public_clone_jgene, similarity, score, filename.

Repertoire schemas

Mal-ID (AIRR-style, ~120 columns) — key fields: sequence_id, repertoire_id, locus, v_call, d_call, j_call, cdr3, cdr3_aa, junction, junction_aa, productive, participant_label, specimen_time_point, plus full nucleotide alignment / insertion-deletion columns.

immunoSEQ cohorts (Savola_RA, Musvosvi_TB, Mitchell_T1D; Adaptive immunoSEQ export, ~54 columns) — key fields: nucleotide, cdr3_aa, count (templates/reads), frequencyCount (%), v_call, j_call, vGeneName, jGeneName, sequenceStatus, estimatedNumberGenomes, sequence, num_reads, repertoire_id, participant_label.

Rawat_T1D (reduced immunoSEQ-derived, 17 columns): cdr3_aa, count (templates/reads), frequency, nucleotide, v_call, d_gene, j_call, sequenceStatus, v_family, d_family, j_family, v_resolved, d_resolved, j_resolved, sequence, num_reads.

Emerson_CMV (minimal, 7 columns): cdr3_aa, v_call, j_call, sequence, num_reads, repertoire_id, participant_label.

Metadata schemas

All metadata.tsv files share participant_label, specimen_label, and disease (the classification target). Cross-validation fold assignments are provided per cohort (CV_fold / fold / malid_cross_validation_fold_id_when_in_test_set). Additional per-cohort columns:

Mal-ID / Mitchell_T1D: specimen_time_point, study_name, available_gene_loci, disease_subtype, age, sex, ancestry.
Savola_RA / Musvosvi_TB: immunoSEQ sample statistics (total_templates, productive_templates, fraction_productive, productive_simpson_clonality, sample_tags, sku, test_name, …).
Rawat_T1D: subject clinical/HLA fields (diabetes_status, sex, age, duration, HLA A/B/C/DPB1/DQB1/DRB1…, hla_high_risk_type, autoantibody statuses, ML_class).
Emerson_CMV: same base fields as Mal-ID/Mitchell_T1D (specimen_time_point, study_name, available_gene_loci, disease_subtype, age, sex, ancestry) plus repertoire_file, emerson_subject_id, cohort, cohort_name, race, ethnicity, race_and_ethnicity, known_cmv_status, metadata_source.

Usage

from huggingface_hub import snapshot_download
import pandas as pd, glob, os

# Download one cohort (metadata + repertoires)
local = snapshot_download(
    repo_id="neurips-2026-dataset/BenchRep-T",
    repo_type="dataset",
    allow_patterns="immunoSEQ/Savola_RA/*",
)

meta = pd.read_csv(f"{local}/immunoSEQ/Savola_RA/metadata.tsv", sep="\t")
rep_files = glob.glob(f"{local}/immunoSEQ/Savola_RA/repertoires/*.tsv.gz")
rep = pd.read_csv(rep_files[0], sep="\t")   # pandas reads .gz transparently

License

Released under the MIT License.

References

Source studies for each cohort:

Mal-ID — Zaslavsky, Maxim E., et al. "Disease diagnostics using machine learning of B cell and T cell receptor sequences." Science 387.6736 (2025): eadp2407.
Savola_RA — Savola, Paula, et al. "Somatic mutations in clonally expanded cytotoxic T lymphocytes in patients with newly diagnosed rheumatoid arthritis." Nature Communications 8.1 (2017): 15869.
Musvosvi_TB — Musvosvi, Munyaradzi, et al. "T cell receptor repertoires associated with control and disease progression following Mycobacterium tuberculosis infection." Nature Medicine 29.1 (2023): 258-269.
Rawat_T1D — Rawat, Puneet, et al. "Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood." Science Advances 12.7 (2026): eadx7448.
Mitchell_T1D — Mitchell, Angela M., et al. "Temporal development of T cell receptor repertoires during childhood in health and disease." JCI Insight 7.18 (2022): e161885.
Emerson_CMV — Emerson, Ryan O., et al. "Immunosequencing identifies signatures of cytomegalovirus exposure history and HLA-mediated effects on the T cell repertoire." Nature Genetics 49.5 (2017): 659-665.

Citation

This dataset accompanies a manuscript currently under double-blind peer review. Author and citation details will be added once the review process is complete.