prescience-bio/hla-hprc-plus

HLA benchmark subsample

Twelve participants with PacBio and ONT WGS, paired Illumina WGS file locations, and detailed HLA typing metadata. The HG002 HLA-Resolve demo is HLA-targeted PacBio hybrid-capture data, separate from the whole-genome sequencing (WGS) inputs. Download progress is recorded per file in downloads/; the HG002 PacBio WGS BAM is already present locally.

This collection is self-contained and follows the format-first layout of dna-format-zoo. It does not require the old ../metadata directory or files in dna-format-zoo. The old metadata has been retained for review; it has not been deleted.

Participants

HG002, HG005, HG01106, HG01258, HG01928, HG02055, HG02630, HG03492, HG03579, NA19240, NA20129, NA21309.

participants.tsv gives the selected platforms, reported PacBio coverage, Illumina source, BioSample identity where available, and capture availability. participant-aliases.tsv contains supported aliases. KIR availability is not a selection criterion, and KIR annotation data is omitted.

Layout

subsample/
  README.md
  VERSION
  catalog.tsv
  participants.tsv
  participant-aliases.tsv
  pacbio-selection.tsv
  pending-capture.tsv
  source-provenance.tsv
  validation.json
  downloads/
    hprc-pacbio.tsv
    nygc-illumina.tsv
    hprc-illumina.tsv
    hprc-assemblies.tsv
    fufihla.tsv
    immuannot.tsv
    lai.tsv
    hla-resolve.tsv
  data/<format>/<technology-or-collection>/<participant>/<case>/
    artifact.yaml
    <sequencing reads or assembly files>
    checksums.sha256
    checksums.blake3
  labels/<source>/<participant>/<case>/
    artifact.yaml
    <HLA typing tables or annotation files>
    checksums.sha256
    checksums.blake3
  tools/
    download.py
    download-all.sh
    download-hg002.sh
    verify.py

Each source has its own file manifest in downloads/. The columns give the artifact/participant, file role, URL, destination, expected size/checksums when available, status, and provenance. catalog.tsv indexes artifact directories under data/ and labels/. artifact.yaml uses JSON serialization, which is valid YAML 1.2 and can also be read using Python's standard JSON library.

Selected files

Source manifest	Selection	Present now
hprc-pacbio.tsv	40 HiFi BAMs across 12 participants	No
nygc-illumina.tsv	18 FASTQs: 9 paired NYGC datasets	No
hprc-illumina.tsv	6 FASTQs: HG002, HG005 and NA21309	No
hprc-assemblies.tsv	24 phased assemblies plus 66 source-listed sidecars, stored under data/fasta/assemblies/<participant>/phased/	No
fufihla.tsv	12 per-person six-gene HLA truth tables	Yes
immuannot.tsv	24 per-haplotype HLA-only GTF annotation files	Yes
lai.tsv	12 per-person additional assembly-derived HLA truth tables	Yes
hla-resolve.tsv	HG002 HLA-targeted hybrid-capture demo FASTQ (not WGS), adapters and upstream README	Yes

The three label sources under labels/ are separate: differences in methods, assemblies, reference databases and unresolved allele fields are retained. None is silently treated as independent experimental consensus truth. The FuFiHLA tables provide 12 actual allele calls per participant for HLA-A, B, C, DRB1, DQA1 and DQB1. Their allele-slot order does not establish maternal/paternal phase. :new and other source values are preserved.

What each participant has

Every participant has the same benchmark shape: one curated PacBio HiFi WGS input set, one ONT WGS input set, one paired Illumina WGS input set, phased assembly candidates, and three small HLA label/annotation sources. HG002 additionally has the small HLA-Resolve capture demo FASTQ from the upstream repository.

Participant	Population	PacBio HiFi WGS	PacBio files	PacBio reported depth	Illumina WGS	Assemblies	Labels available	HLA-Resolve capture demo
HG002	N/A	BAM, Revio	1	32.6×	paired FASTQ, HPRC Illumina; filenames indicate 30× PCR-free NovaSeq	phased FASTA, HPRC R2/Q100-style HG002 assembly files plus sidecars	FuFiHLA six-gene HLA typing; Lai assembly HLA typing; Immuannot R1 HLA GTF annotations	Yes, repository demo FASTQ plus adapters/README
HG005	N/A	BAM, Sequel	4	30.1×	paired FASTQ, HPRC Illumina; filenames indicate 30× PCR-free NovaSeq	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG01106	PUR	BAM, Sequel	3	34.6×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG01258	CLM	BAM, Sequel	4	35.9×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG01928	PEL	BAM, Sequel	4	35.1×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG02055	ACB	BAM, Revio	4	34.9×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG02630	GWD	BAM, Sequel	3	36.0×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG03492	PJL	BAM, Revio	4	36.3×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R1 fallback	FuFiHLA; Lai; Immuannot R1	pending full SRA only
HG03579	MSL	BAM, Sequel	3	35.7×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
NA19240	YRI	BAM, Sequel	3	30.7×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
NA20129	ASW	BAM, Sequel	4	33.0×	paired FASTQ, NYGC/ENA	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only
NA21309	MKK	BAM, Sequel	3	30.3×	paired FASTQ, HPRC Illumina	phased FASTA, HPRC R2 plus sidecars	FuFiHLA; Lai; Immuannot R1	pending full SRA only

Illumina depth is not normalized in this collection. HG002 and HG005 HPRC Illumina filenames explicitly describe 30× PCR-free NovaSeq. Other Illumina rows retain upstream FASTQ file sizes and MD5s where ENA provides them, but a single curated depth field has not yet been assigned.

This is a good benchmark layout because it combines orthogonal evidence without pretending it is all the same kind of truth: long reads for HLA-resolving sequence context, short reads for conventional WGS callers, assemblies for haplotype-level inspection, and multiple assembly-derived HLA label sources for cross-checking. The 30×-ish PacBio subset keeps storage practical while preserving whole upstream run files instead of downsampling. Separating data/ from labels/ also makes it clear which files are caller inputs and which files are expected answers or annotations.

PacBio choice

The provisional WGS shortlist targets at least 30× source-reported coverage per participant, using whole sequencing runs. It prefers Revio where at least 30× with known coverage is available; otherwise it uses Sequel HiFi. Within the chosen instrument group it selects the fewest files reaching 30×, then the lowest total coverage among those choices. No reads were downsampled. HG03492's Revio classification comes from the historical Revio workflow/file context because its old manifest instrument field is blank.

The results range from 30.1× to 36.3×; see pacbio-selection.tsv for every selected filename and the exact rule. HG002 uses one 32.6× Revio BAM. DeepConsensus alternatives, methylation-specific mirrors, unknown-coverage runs and unneeded extra runs are not included. This is a practical provisional input selection, not a reconstruction of the exact FuFiHLA benchmark inputs. Source coverage is approximate sequence yield, not guaranteed local HLA coverage or file size.

Illumina uses paired FASTQ for all 12 participants, with both mates retained. NYGC FASTQs are the ENA-generated representation of the NYGC runs. CRAM duplicates are omitted, so an exact CRAM-decoding reference is not required for these inputs.

Assembly candidates are HPRC R2, except HG03492 uses an R1 fallback. Immuannot annotations refer to R1 contigs and IPD-IMGT/HLA 3.52.0; those coordinates must not be applied directly to a different assembly or GRCh38. Matching each truth source to its precise assembly remains a benchmark-design task.

HG002 HLA-targeted hybrid-capture demo (not WGS)

The two HG002 PacBio inputs serve different purposes:

Input	Scope	Local path
HLA-Resolve demo FASTQ (~46 MiB)	HLA-targeted hybrid-capture reads	data/fastq/pacbio/HG002/HG002.hifi_reads.fastq.gz
HPRC Revio HiFi BAM (~41 GiB)	Whole-genome reads; source-reported 32.6× coverage	data/bam/pacbio/HG002/hifi-wgs/m84011_220902_175841_s1.hifi_reads.bam

The demo is a separate targeted-capture dataset, not an established computational HLA slice of the WGS BAM. Use --scheme hybrid_capture for the demo when running HLA-Resolve.

Demo file details:

• 48,324,079 compressed bytes.
• 88,955 reads, containing 348,125,272 bases.
• Entire gzip stream and FASTQ structure checked; sequence/quality lengths agree for every record.
• Compressed file, adapters and README matched the pinned Git blob hashes from commit 0c9c4460bfefe3916cb4b89c2901a3e3c135e8d2.
• Local SHA-256 and BLAKE3 hashes recorded.

This is PacBio hybrid capture, not PCR amplicon data. The repository demo is not established as identical to the full capture SRA run. pending-capture.tsv records the 12 full-run accessions separately. All 12 SRA pages said unreleased when checked on 2026-09-09; the other 11 participants therefore have no capture payload in this collection yet. Pending accessions are not included as executable download URLs.

ONT WGS selection

downloads/hprc-ont.tsv lists 28 planned BAM files across all 12 participants, stored under data/bam/ont/<participant>/wgs/. ont-selection.tsv records the file count, source-reported coverage, chemistry and selection policy for each participant.

For the 11 participants in the HPRC R2 ONT catalog, the selection uses the fewest whole run files reaching at least 30× reported coverage, then minimizes excess coverage. These are unaligned ONT reads; no reads have been downsampled. HG002 uses one R9.4.1 file, 03_08_22_R941_HG002_4.bam, with 31.76× total coverage (12.17× from reads over 100 kb). This is not the separate modern GIAB R10 dataset. NA19240, NA20129 and NA21309 use R10.4.1; the other current-catalog participants use R9.4.1.

HG03492 is absent from the current ONT catalog. All nine BAMs listed in its historical polishing inputs are retained, with coverage and chemistry marked unknown; a 30× minimum has not been verified for that participant. File sizes and publisher checksums were not provided in these metadata sources. URLs come from the source manifests; their current availability has not been checked.

Both tools/download-hg002.sh and tools/download-all.sh include hprc-ont. Preview before downloading:

tools/download-hg002.sh --dry-run
# ONT only, HG002 (preview):
python3 tools/download.py --source hprc-ont --participant HG002

Add --execute to the Python command to fetch ONT only. The shell wrapper downloads its full source selection unless --dry-run is supplied. No ONT sequencing payload was downloaded while adding this metadata. The initial counts in validation.json predate this ONT addition; this adds 12 artifacts and 28 planned files.

Preview and download

Commands can run from any working directory when given the correct path to the scripts. They require Python 3.9+, curl, and b3sum for BLAKE3 verification. No Python packages are required.

From this directory, preview HG002 WGS downloads:

python3 tools/download.py --source hprc-pacbio --source hprc-ont --source hprc-illumina --participant HG002

Add --execute to actually fetch that selection. The tool resumes .partial files, verifies available upstream sizes/MD5s/Git blob hashes, and records successful downloads. Omitting --participant selects every participant in the chosen source manifests. Sources can be repeated. It never downloads anything without --execute.

For all 12 WGS inputs, the source selection is:

python3 tools/download.py --source hprc-pacbio --source hprc-ont --source nygc-illumina --source hprc-illumina

Convenience wrappers are also provided. They execute by default; pass --dry-run to preview.

# All directly downloadable files in the collection manifests.
tools/download-all.sh --dry-run
tools/download-all.sh

# HG002 only, including the HLA-Resolve repository demo sample if absent.
tools/download-hg002.sh --dry-run
tools/download-hg002.sh

Assemblies are a separate selection, --source hprc-assemblies. The small HLA metadata is already materialized, so its manifest rows have blank direct download URLs and public provenance URLs instead. No parent-cohort workbook is needed to use the local tables.

Sources combined here

Source	Used for	Links
Human Pangenome Reference Consortium / HPRC intermediate assembly catalogs	PacBio HiFi WGS BAMs, HPRC Illumina FASTQs, HPRC R2 assembly FASTAs and sidecars	repository, R2 HiFi catalog, R2 Illumina catalog, R2 assembly catalog
HPRC historical/pre-release metadata	Older PacBio records and HG03492 R1 fallback assembly context	HiFi pre-release catalog, HG03492 R1 workflow input JSON
ENA / NYGC Illumina read records	Paired Illumina FASTQs for nine participants	PRJEB31736 file report, PRJEB36890 file report
HLA-Resolve repository	HG002 PacBio hybrid-capture demo FASTQ, adapters, upstream README, Lai supplementary workbook link, and pending full-capture SRA accession list	repository, pinned demo directory, capture validation doc, Lai supplementary workbook
FuFiHLA paper supplement	Six-gene per-participant HLA typing labels, extracted from the paper's Actual/ground-truth columns	paper/supplement page, supplement download endpoint
Immuannot paper supplement	Per-haplotype HLA-only GTF annotation derivatives from HPRC R1 assemblies	paper/supplement page, supplement download endpoint

source-provenance.tsv records retrieval dates, source payload sizes and SHA-256 hashes for source metadata used during preparation. Per-file URLs, expected upstream hashes and status live in the downloads/*.tsv manifests and corresponding artifact.yaml files.

Integrity and status

./tools/verify-artifacts.sh          # SHA-256 + BLAKE3 + available upstream hashes
./tools/verify-artifacts-blake.sh    # Faster BLAKE3-only content pass

This verifies catalog/manifest agreement, participant membership, path uniqueness, available upstream hashes, and SHA-256/BLAKE3 for all stored files. Missing files marked pending are reported rather than treated as unexpected corruption. Use --require-all when all planned files should be present. --refresh regenerates local hash manifests after checking available upstream hashes; it is not a replacement for checking old hashes when corruption is suspected.

Both entry points check metadata consistency, expected sizes, and coverage of both checksum manifests. The BLAKE3-only pass skips SHA-256 and upstream content hashing. verify-artifacts-blake3.sh is an alias matching the dna-format-zoo name. The main script also accepts --sha256; python3 tools/verify.py remains available. Checksum refresh requires the default full mode.

Empty checksum manifests in planned artifact directories mean no payload is stored there yet. Publisher checksums are distinct from local checksums. HPRC file sizes/MD5s absent from source metadata remain unknown; downloaded assembly .md5 sidecars are retained but not automatically interpreted as verified assembly checksums. Successful HTTP transfer and a new local hash alone do not establish a match to a missing publisher digest.

validation.json records the initial preparation result. Current download state lives in each per-source manifest and artifact.yaml. The initial collection contains 51 stored files and 154 planned file entries. No full WGS or genome assembly payload was fetched during preparation.