Datasets:

lordim
/

BenchRep-T

This datasheet follows the framework of Gebru et al., "Datasheets for Datasets" (CACM 2021). Because BenchRep-T aggregates repertoires from multiple independently collected source cohorts, several questions are answered per source cohort. A summary of the sources appears under Composition.

Motivation

For what purpose was the dataset created?

BenchRep-T was created to provide a fixed, reproducible benchmark for comparing TCRβ repertoire-based disease-classification methods under identical inputs, preprocessing, and evaluation splits. Prior comparisons were confounded by differences in cohorts, preprocessing, and data splits. BenchRep-T freezes cleaned repertoire datasets, common model-facing sequence fields, and fixed three-fold assignments so that methods can be compared on the same examples.

Who created the dataset and on behalf of which entity?

BenchRep-T was created by Alejandro Buendia*, Chiho Im*, Anshul Kundaje, Scott D Boyd, and Liel Cohen-Lavi (* co-first authors) for the NeurIPS 2026 Evaluations and Datasets Track paper "BenchRep-T: A Systematic Evaluation of T-Cell Repertoire-Based Disease Diagnostics."

Who funded the creation of the dataset?

Composition

What do the instances represent?

Each instance is one TCRβ repertoire from one biological specimen. A repertoire is a table of TCRβ sequences with CDR3 amino-acid sequences and V/J gene calls. What a row stands for depends on the platform: Mal-ID files have one row per sequencing read that passed preprocessing, whereas immunoSEQ files have one row per unique rearrangement with Adaptive's template or read count. A participant may contribute more than one specimen. Repeated specimens from the same participant are explicitly linked by participant_label and are kept in the same evaluation fold.

How many instances are there in total?

Across the six non-overlapping source collections, BenchRep-T contains 2,352 labeled repertoire specimens from 2,264 source participant identifiers. This total counts each source repertoire once and assumes that the independently collected source cohorts do not share participants. The common-method headline comparisons contain 1,300 unique evaluated specimens: 550 Mal-ID, 196 Mitchell, the 204-specimen Rawat evaluation fold, 90 complete-case RA specimens, 140 TB specimens, and the 120-specimen Emerson validation cohort.

Per-cohort and per-task counts:

Source cohort	Condition(s)	Platform	N specimens	Positives / total (per task)	Base rate
Mal-ID (Zaslavsky et al.)	COVID-19	Custom cDNA TCRβ AIRR-seq / Illumina MiSeq	255	58 / 255	22.75%
Mal-ID	HIV	Custom cDNA TCRβ AIRR-seq / Illumina MiSeq	295	98 / 295	33.22%
Mal-ID	Influenza vaccination	Custom cDNA TCRβ AIRR-seq / Illumina MiSeq	234	37 / 234	15.81%
Mal-ID	Lupus	Custom cDNA TCRβ AIRR-seq / Illumina MiSeq	261	64 / 261	24.52%
Mal-ID + Mitchell et al.	T1D (pooled)	Mal-ID cDNA/MiSeq + Mitchell Adaptive immunoSEQ gDNA TCRβ	489	267 / 489	54.60%
Savola/Kelkka et al.	Rheumatoid arthritis	Adaptive immunoSEQ of gDNA from sorted T-cell fractions	91[^ra]	71 / 91[^ra]	78.02%[^ra]
Musvosvi et al.	TB progression	Adaptive immunoSEQ, bulk PBMC gDNA TCRβ	140	63 / 140	45.00%
Rawat et al. 2026	T1D (external)	Adaptive immunoSEQ hsTCRB_v4, bulk PBMC gDNA TCRβ	614 (204 in evaluation fold)	142 / 204 (evaluation fold)	69.61%
Emerson et al. 2017	CMV serostatus	Adaptive immunoSEQ, bulk gDNA TCRβ	641 discovery / 120 validation	51 / 120 (validation cohort)	42.50%

The per-task Mal-ID counts must not be summed because all five tasks reuse the same 197 healthy controls. The benchmark contains 550 unique Mal-ID specimens from 542 participants: 197 healthy controls, 58 COVID-19, 98 HIV, 37 influenza-vaccination recipients, 64 lupus, and 96 T1D specimens. The pooled T1D task combines 293 Mal-ID specimens (96 T1D and 197 controls) with 196 Mitchell specimens (171 T1D and 25 controls), producing 489 specimens and 267 positives.

[^ra]: The processed RA cohort contains 91 specimens and 71 positives. The common-method comparison excludes RA6-SF, which is absent from the Mal-ID prediction output, leaving 90 specimens, 70 positives, and a 77.78% base rate in the performance table labeled (n=90).

Does the dataset contain all possible instances or a sample?

BenchRep-T contains task-specific, quality-controlled subsets of the source collections rather than every specimen collected by the original studies:

Mal-ID: BenchRep-T includes all 550 specimens with paired IgH and TRB data used in the original Mal-ID paired-repertoire evaluation. Sixty-six BCR-only specimens in the supplied metadata (34 lupus, 27 healthy, and 5 COVID-19) are not eligible for this TCR benchmark. No additional participant-level biological exclusion was applied after selecting the target conditions and TCR-available specimens.
Mitchell T1D: one repertoire per participant was retained: 143 new-onset T1D validation participants, 28 DAISY progressors, and 25 DAISY controls. One additional DAISY case (310264) was excluded because it contained only five productive rearrangements, leaving 196 repertoires.
RA: 91 of the 94 downloaded repertoires passed the minimum-depth rule. RA23-SF-CD4, RA2-CD8+Vb1+, and RA69 were excluded because they contained fewer than 1,000 productive rearrangements after preprocessing. No further disease-status exclusion was applied.
TB: BenchRep-T uses all 140 released bulk repertoires that passed the source study's QC. The source study had already excluded one ACS bulk sample that did not align with the participant's other samples and ten GC6-74 samples that failed material-transfer or within-participant alignment QC.
Rawat T1D: BenchRep-T selects 426 participants with T1D and 188 unrelated, islet-autoantibody-negative controls from Rawat cohort 1. It excludes 625 first-degree relatives, 59 second-degree relatives, and 95 autoantibody-positive participants without diabetes. These categories account for all 1,393 cohort-1 repertoires.
Emerson CMV: all 761 specimens with known CMV serostatus are retained. Twenty-five discovery-cohort specimens with unknown CMV status remain documented in the source metadata but are excluded from supervised classification. The headline external evaluation is restricted to the original 120-person validation cohort.

What data does each instance consist of?

The guaranteed model-facing sequence fields are:

cdr3_aa: the amino-acid sequence used by the benchmark;
v_call and j_call: harmonized gene-level V and J calls;
sequence: nucleotide rearrangement, where available;
num_reads and source abundance/frequency columns, where supplied by the source platform;
repertoire_id and participant_label, where present in the cohort adapter.

cdr3_aa, rather than junction_aa, is the benchmark sequence field. In the Mal-ID conversion, junction_aa is an empty AIRR placeholder; in Adaptive cohorts, the conserved leading cysteine and terminal phenylalanine/tryptophan are removed so that cdr3_aa follows the Mal-ID sequence convention. The cleaned files preserve additional source-specific columns, so their complete schemas are not identical. Clone abundance is also represented differently across platforms. Adaptive exports have one row per unique rearrangement, with its template/read count in num_reads. Mal-ID files are not deduplicated: each row is one read, identical reads repeat (specimen M124-S014, for example, has 170,507 rows but 96,178 distinct nucleotide sequences), and the source's clone_id groups a specimen's reads into clones. Mal-ID files have no count column, so their abundance is carried by row multiplicity. Evaluators that request duplicate_count fall back to unit weight when that field is absent, so methods that count rows weight a Mal-ID clone by its number of reads but count each immunoSEQ rearrangement once. The immuneML-ready export under immuneML/ adds junction/junction_aa, rebuilt from the nucleotide data where they translate back to cdr3_aa; see immuneML/README.md.

Per-specimen metadata include a pseudonymous specimen identifier, a participant identifier, condition label, fixed fold, and source-specific demographics.

Is there a label or target associated with each instance?

Yes. Each task supplies a binary condition, serostatus, or outcome label: case versus healthy/background control, TB progressor versus controller, or CMV-seropositive versus CMV-seronegative. Fold assignments are fixed. For cohorts containing repeated specimens, the assignment is made at the participant level so that all specimens from one participant remain in one fold.

Is any information missing from individual instances?

Yes. Demographic and HLA completeness varies substantially:

Cohort	Age	Sex	Ancestry/race	HLA in released benchmark metadata
Mal-ID (550 specimens)	509/550 (92.5%)	451/550 (82.0%)	361/550 (65.6%)	No
Mitchell (196)	143/196 (73.0%)	196/196 (100%)	Not reliably available[^mitchell-ancestry]	No
Rawat (614)	614/614 (100%)	614/614 (100%)	No ancestry field	585/614 (95.3%)
RA (91)	Not included	Not included	Not included	Not included
TB (140)	Not included	Not included	Not included	Not included
Emerson labeled set (761)	674/761 (88.6%)	761/761 (100%)	548/761 (72.0%)	Not included

Among Mal-ID specimens, 87 of 98 HIV specimens (88.8%) have recorded ancestry and all 87 are labeled African; 52 of 64 lupus specimens (81.3%) are female. These skews motivate the confounding analyses and must not be interpreted as population prevalence estimates.

[^mitchell-ancestry]: The Mitchell conversion currently writes White for every participant as a cohort-level harmonization value, regardless of heterogeneous source race/ethnicity values. It must not be treated as verified individual-level ancestry or used for ancestry-confounding analyses.

Are relationships between instances made explicit?

Yes. participant_label links repeated specimens within Mal-ID, RA, and TB, and the folds prevent participant leakage. The Rawat classification subset excludes the source study's first- and second-degree-relative groups and retains unrelated controls; there are no known family links among the included benchmark participants, although BenchRep-T did not independently perform genetic kinship inference.

No duplicated specimen identifiers were found across source protocols. Clone-set comparisons show only low public-clone overlap: the reported maximum cross-source repertoire Jaccard similarity in pooled T1D is 0.013, with no pair above 0.10, and the aggregate Mal-ID–Mitchell CDR3-amino-acid Jaccard similarity is 0.035. These sequence-overlap statistics are supporting checks, not proof of participant identity.

Are there recommended data splits?

Yes. The supplied participant-grouped, fixed three-fold assignments are the intended benchmark splits. Main within-cohort numbers are computed from pooled out-of-fold predictions. Emerson cohort 2 is the held-out CMV validation fold; Rawat fold 2 is the fixed 204-specimen external evaluation fold used in the reported comparison. Users should not create new random splits when attempting to reproduce reported results.

Are there errors, sources of noise, or redundancies?

AIRR-seq data contain expected PCR, sequencing, annotation, and sampling noise. Sequencing depth and assay chemistry differ across cohorts; Mal-ID is RNA/cDNA-based, while the external immunoSEQ cohorts are genomic-DNA based. The RA data are primarily sorted T-cell fractions rather than bulk unsorted PBMC repertoires. These platform and cell-composition differences may create cohort effects.

Additional known caveats are:

incomplete V/J resolution and non-standard CDR3 characters are removed during preprocessing;
allele annotations are stripped, and biologically or technically indistinguishable V genes are collapsed;
abundance is not harmonized: Mal-ID rows are individual reads, immunoSEQ rows are unique rearrangements with counts, and some evaluators therefore operate on unique-clonotype presence;
demographic fields are incomplete and use harmonized category vocabularies;
the Mitchell ancestry value is a conversion-time cohort label rather than reliable individual-level metadata;
Mal-ID AIRR files retain lowercase t/f booleans and empty junction/junction_aa placeholders, so strict AIRR-schema validation requires conversion to uppercase T/F (the immuneML/ export is converted and passes the AIRR schema validator);
exact sequencing depth differs considerably across specimens and cohorts; depth sensitivity is evaluated explicitly in the benchmark;
25 Emerson specimens with unknown CMV status are metadata-only and cannot be used for supervised evaluation.

Is the dataset self-contained or does it link to external resources?

The benchmark requires cleaned repertoire files, metadata, and fixed folds. The preprocessing code also records the upstream Synapse and immuneACCESS sources needed to reconstruct the cleaned form. Raw source access is not required to run benchmark methods once an authorized copy of the cleaned benchmark has been obtained.

Does the dataset contain confidential or personally identifiable data?

No names, medical-record numbers, postal addresses, or direct contact identifiers are included. Identifiers are source-issued pseudonyms. The data should nevertheless be treated as potentially re-identifiable biomedical data: immune repertoires and HLA genotypes are high-dimensional and individual-specific, and combinations of exact age, sex, disease, HLA, and collection date can act as quasi-identifiers.

The Rawat metadata originally included each specimen's exact collection date. No benchmark task uses it, and together with the day-precise age it would reveal each participant's date of birth, so it was removed from the released metadata on 2026-10-29. Only fields required for the benchmark should be released. No formal quantitative re-identification-risk analysis has been performed.

Does the dataset contain sensitive data (health, demographics)?

Yes. It contains disease or immune-status labels and, depending on cohort, age, sex, ancestry/race categories, HLA genotype, and specimen timing. These values originate from source metadata and are mapped into a shared vocabulary; no demographic attribute is inferred from TCR sequences. The exception is the Mitchell ancestry column described above, which is a conversion-time cohort label and should be removed or marked unavailable rather than presented as individual-level source data.

Collection Process

How was the data acquired, and by whom?

BenchRep-T did not collect biological specimens. All samples were collected and sequenced by the source studies:

Source	Original study	Original acquisition
Mal-ID	Zaslavsky et al., Science 2025	Peripheral-blood RNA; random-hexamer-primed cDNA; separate TCRβ PCR; paired-end Illumina MiSeq sequencing; IgBLAST annotation
Mitchell T1D	Mitchell et al., JCI Insight 2022	Genomic DNA from longitudinal peripheral-blood specimens and a new-onset T1D validation cohort; Adaptive immunoSEQ TCRβ sequencing
RA	Savola et al., Nature Communications 2017 and Kelkka et al., Frontiers in Immunology 2020	Genomic DNA from sorted T-cell fractions, primarily peripheral-blood CD8+ cells with some CD4+ or synovial-fluid fractions; Adaptive immunoSEQ TCRβ
TB	Musvosvi et al., Nature Medicine 2023	Longitudinal unstimulated PBMC specimens from M. tuberculosis-infected progressors and controllers in the ACS and GC6-74 cohorts; bulk Adaptive immunoSEQ TCRβ
Rawat T1D	Rawat et al., Science Advances 2026	Genomic DNA from bulk PBMCs; bias-controlled Adaptive immunoSEQ hsTCRB_v4 assay at shallow depth for cohort 1
Emerson CMV	Emerson et al., Nature Genetics 2017	Genomic DNA TCRβ immunosequencing of discovery and independent validation blood-donor cohorts using Adaptive immunoSEQ

Over what timeframe was the data collected?

Collection dates are not uniformly reported in a common format, so BenchRep-T does not assign one global collection window. Rawat cohort 1 was collected from 2010 through 2018. Mitchell includes four longitudinal childhood time points in DAISY plus a cross-sectional new-onset validation cohort. The TB source follows progressors and controllers longitudinally over two years, with released bulk specimens at study days 0, 180, 360, and 540 where available. The RA collection includes newly diagnosed specimens and a small number of follow-up or synovial-fluid specimens. The component Mal-ID studies and Emerson cohorts were collected over the windows reported in their respective publications and supplements.

Were ethical review or IRB processes conducted?

The original studies report institutional ethics approval and informed consent under their respective protocols. For example, Rawat cohort-1 samples were de-identified under University of Florida IRB201400703. BenchRep-T is a secondary computational analysis of previously collected, de-identified or pseudonymized data and involved no new recruitment or biospecimen collection.

The repository audit did not identify a separate institutional determination or exemption number covering BenchRep-T itself. Therefore, this datasheet does not claim a BenchRep-T-specific IRB exemption number; one should be added only if the authors have an applicable institutional determination.

Was consent obtained, and for what?

The source publications report informed consent, with parental permission and participant assent where applicable. Consent scope and downstream-use conditions remain those of each source study and repository. BenchRep-T does not broaden those permissions.

The local repository does not contain source-author correspondence or a source-by-source legal determination authorizing unrestricted redistribution of every cleaned repertoire. Consequently, secondary analytical use can be described as permitted under the accessed source terms, but unrestricted public redistribution of the sequence files should be claimed only for cohorts with documented redistribution permission. For any cohort lacking such permission, users should obtain the source data directly and run the released preprocessing scripts.

Preprocessing / Cleaning / Labeling

Was any preprocessing done?

Yes. The model-facing fields were harmonized to cdr3_aa, v_call, and j_call.

For Mal-ID, preprocessing retains productive rearrangements, requires v_score > 80, removes missing CDR3/V/J fields and non-standard amino-acid characters, uppercases and strips IgBLAST spacing, removes TRBV25/OR9-2*01, fixes/collapses V calls that cannot be distinguished under the FR3 primer scheme (TRBV12-4→TRBV12-3, TRBV6-3→TRBV6-2), and converts the cleaned internal tables to AIRR-style columns. The Mal-ID source had already assigned each person's reads to clones (clone_id); the reads remain separate rows rather than being collapsed into one row per clone.

For Adaptive/immunoSEQ cohorts, preprocessing:

retains productive (sequenceStatus == "In") rearrangements;
removes unresolved, missing, empty, family-only, or otherwise non-gene-level V/J calls;
maps Adaptive gene names to IMGT/AIRR form and strips allele suffixes;
harmonizes orphon names and collapses the same indistinguishable V genes used for Mal-ID;
removes the conserved first and last CDR3 residues to match the Mal-ID cdr3_aa convention;
removes CDR3s containing non-standard amino acids;
preserves nucleotide sequence and source template/read count as sequence and num_reads where available; and
excludes specimens below 1,000 usable productive rearrangements where required by the cohort protocol.

The preprocessing does not deduplicate rows, within or across cohorts: Mal-ID rows remain individual reads with their clone_id, and immunoSEQ rows remain Adaptive's unique rearrangements with their counts. Fixed folds are participant-grouped, so repeated specimens from one participant never cross training and test partitions.

Was the raw data saved in addition to the cleaned data?

Raw source data remain at their original repositories, including Synapse and Adaptive immuneACCESS. BenchRep-T's repository contains preprocessing scripts and local paths/manifests needed to regenerate cleaned files for users who are independently authorized to access the sources.

Is the preprocessing software available?

Yes. Cleaning scripts, cohort adapters, fixed-fold metadata, and per-method environments are provided in the public code repository at https://github.com/boyd-lab/BenchRep-T. The immuneML-ready export is produced by utils/immuneml_export.py and utils/immuneml_tasks.py in the same repository.

Uses

What has the dataset been used for?

BenchRep-T benchmarks nine TCRβ repertoire-classification methods across four tasks: binary disease/condition classification, sequencing-depth robustness, antigen-specific driver-sequence recovery, and demographic-confounding sensitivity.

What other tasks could it be used for?

Potential uses include AIRR machine-learning method development, transfer and out-of-distribution studies, analysis of cohort effects, robustness to sequencing depth, and confounding-aware evaluation. An immuneML-ready AIRR export, with the fixed folds and example specifications, is provided under immuneML/.

Is there anything about its composition or collection that could cause unfair treatment or harm?

Yes. The cohorts are small, demographically skewed, and collected with different platforms and biological sampling protocols. For example, the recorded HIV subset is predominantly African ancestry and the lupus subset is predominantly female. Disease status is therefore entangled with ancestry, age, sex, source study, sequencing chemistry, cell fraction, and collection context. Models may exploit those variables rather than disease biology. The dataset is not population representative, and subgroup estimates may be unstable.

Are there tasks for which the dataset should not be used?

Yes. BenchRep-T must not be used for clinical diagnosis, treatment decisions, insurance or employment decisions, participant identification, ancestry inference, or any other deployed high-stakes decision. It is a research benchmark, not a clinically validated assay.

Distribution

How is the dataset distributed?

The cleaned repertoires, per-cohort metadata with fixed fold assignments, and auxiliary task files are distributed through the Hugging Face Hub at https://huggingface.co/datasets/lordim/BenchRep-T. Code is available at https://github.com/boyd-lab/BenchRep-T. Raw source data remain at Synapse or immuneACCESS.

Before public release, the authors should publish a manifest that identifies, for each file, its source, checksum, applicable source terms, redistribution status, and preprocessing version. Croissant metadata should be claimed only after a Croissant record is actually included in the release. For the immuneML export, immuneML/MANIFEST.tsv records each file's checksum and, for repertoires, the native source file and its checksum; source terms and redistribution status are not yet recorded per file.

When will it be distributed?

The code and metadata are available through the review artifact. A permanent public release is planned for the camera-ready version, subject to source-specific redistribution permissions and removal or generalization of unnecessary quasi-identifiers such as exact collection dates. The permanent release should be deposited as an immutable, versioned archive with a DOI.

License / terms of use?

No single blanket license can currently be asserted for all constituent sequence files:

the BenchRep-T article is CC BY 4.0, but an article license does not automatically relicense third-party data;
the local Mal-ID source bundle contains a CC BY-NC-SA 4.0 license, which is not compatible with presenting the Mal-ID-derived sequence bundle as unrestricted CC BY 4.0 without separate permission;
immuneACCESS and Synapse data remain subject to their repository- and dataset-specific terms; availability for download does not by itself establish a right to redistribute a transformed copy; and
BenchRep-T-authored code, metadata, and documentation may be licensed separately only to the extent that the authors own those materials.

The release should therefore preserve per-source terms and attribution. A CC BY 4.0 label may be applied to BenchRep-T-authored metadata and documentation, but it should not be applied to the aggregate sequence files unless written permission or a compatible upstream license has been documented for every constituent cohort.

Have any third parties imposed IP-based or other restrictions?

Potentially. Synapse controlled-access data may not be redistributed, and immuneACCESS projects can carry project-specific terms. BenchRep-T's cleaned-versus-raw distinction does not automatically remove those restrictions. For any source that prohibits redistribution, the compliant release pattern is to distribute code, checksums, manifests, labels/folds where permitted, and automated preprocessing instructions while requiring each user to retrieve the source repertoire through the original access mechanism.

Maintenance

Who will maintain the dataset, and how can they be contacted?

Will the dataset be updated?

Yes. Planned updates include larger and more demographically balanced cohorts, completed full-cross-validation results for external cohorts, stricter schema validation, and corrections documented in the changelog. Updates must not silently change the benchmark: every release should record cohort membership, checksums, folds, preprocessing commit, and metric corrections.

Will older versions be maintained?

Yes. The intended policy is semantic versioning with immutable archived releases. Each public release should receive a DOI; superseded versions should remain accessible, and the default landing page should identify the current version and link its changelog. This policy is prospective until the first archive DOI is minted.

Is there an erratum?

No. There are currently no known errors in the released dataset requiring an erratum. Corrections to manuscript text, figures, or reported method-performance results are documented separately and do not constitute dataset errata.