# Asian HLA benchmark — reference release 2026-10-20

## Start with the truth labels

[Download the truth table](truth/haplotype_assignments.tsv) | [Understand the evidence and columns](truth/README.md)

This is the assembly-derived reference-label dataset used to build the benchmark: 147 donors, eight HLA loci, 2,351 available gene copies. One expected Arab gene copy is missing. Labels retain matching allele alternatives and the resolution supported by sequence evidence. These labels are benchmark truth candidates; independent raw-read/phase certification and an untouched evaluation cohort are still outstanding. Method predictions are not included in the truth table.

## Choose what you need

| Task | Files |
| --- | --- |
| Obtain the reference labels | [truth/haplotype_assignments.tsv](truth/haplotype_assignments.tsv) |
| Inspect genomic/CDS/protein evidence | [truth/](truth/) |
| Inspect extracted sequences and source annotations | [inputs/](inputs/) |
| Check sample sources and pinned database version | [metadata/](metadata/) |
| Audit Arab/Saudi source binding | [provenance/](provenance/) |
| Download everything needed for label regeneration | [Complete benchmark archive](https://bio2vec.net/data/Asian-HLA-Consortium/downloads/asian-hla-benchmark-2026-10-20.tar.gz) |
| Reproduce the assignments | [Code and commands](https://github.com/Asian-HLA-Consortium/hla-assembly-benchmark) |
| Compare caller predictions against these labels | [Prediction catalogue](https://bio2vec.net/data/Asian-HLA-Consortium/predictions/README.md) |

## Scope and validation

86 HPRC donors (50 East Asian, 36 South Asian), 52 Arab donors and nine Saudi donors. Loci: A, B, C, DPA1, DPB1, DQA1, DQB1, DRB1. The HPRC inputs contain both phased regional assemblies and all 172 original source GTFs. Arab/Saudi inputs contain retained gene intervals and their JSON annotations and haplotype/source mappings; no source GTF is claimed for those intervals.

IPD-IMGT/HLA 3.65.0 is pinned. HPRC copies: 1,201 genomic matches, 173 CDS-only, one protein-only and one unresolved. Arab/Saudi copies: 813 genomic-record matches, 150 spliced CDS matches and 12 unresolved. An unmatched sequence is not automatically a novel allele. A stored genomic record may not certify the outer biological gene boundaries.

The reorganized inputs were rerun through the generator and checked against the complete historical per-copy semantic hashes. Run `sha256sum -c SHA256SUMS` in the extracted release directory before using the files. `MANIFEST.json` lists file sizes and hashes. Reference IPD files are downloaded from their original provider using the repository's pinned downloader; they and whole-genome reads are excluded here.

This release separates truth, inputs and predictions. It has explicit paths, with no moving directory aliases. Use this dated release URL and file checksums when citing an analysis. Later completed caller outputs receive their own immutable snapshot URLs and do not alter the truth release.

[Google benchmark tracking Sheet](https://docs.google.com/spreadsheets/d/18F0YMcjL8NUnPrjghI8z567m0oV4vYD7vHiM6KFASGo/edit) | [Consortium guide](https://bio2vec.net/data/Asian-HLA-Consortium/README.md)