# Results: completed baselines and bounded typing experiments

Updated 2026-10-21. OptiType, SpecHLA and HLA*LA each completed all **86 donors**.
The completed developmental experiments support a strong contribution from full
genomic reference sequences and additional recovery from population-derived
sequences. The genomic-reference graph comparison has no clear aggregate gain,
and the tested linkage prior has a null aggregate result.

The [final native-caller snapshot](https://bio2vec.net/data/Asian-HLA-Consortium/predictions/native-callers/7b20524bb11cfcb30271e07a94d1448ab621a78380b47032ee1c6199e91469a8/)
contains finalized predictions, scores and provenance. The
[experiment release](https://bio2vec.net/data/Asian-HLA-Consortium/experiments/2026-10-21/)
contains the full interpretation, machine-readable verdict and matched contrasts
under `scores-and-interpretation/`, with baseline summaries in `baselines/`.
[Assembly-derived reference labels](https://bio2vec.net/data/Asian-HLA-Consortium/releases/2026-10-20/truth/)
remain subject to independent raw-read, phasing and boundary certification.

## Software and database versions

| Baseline or pipeline | Software version | Database/reference version used |
| --- | --- | --- |
| T1K | v1.0.6-r216; runtime binary hashes pinned | IPD-IMGT/HLA v3.65.0, with shortened/default and full-genomic configurations reported separately; fold-specific population additions where stated |
| OptiType | v1.5.0 | Bundled historical class-I database; exact IPD release unknown; reference and allele-table hashes pinned |
| SpecHLA | v1.0.12 | Annotation IPD-IMGT/HLA v3.65, renewal commit `8707b783b02cef7141db4dbf8382d9449cdab5d3`; mapping/binning v3.38; renewed G annotation v3.65.0 |
| HLA*LA | v1.0.4 | Historical `PRG_MHC_GRCh38_withIMGT` graph; exact IPD release unknown; archive/serialized-graph hashes pinned; native G annotation v3.32.0; common-G scoring uses current v3.65 projection |
| DogoHLA graph pipeline | Development snapshot pinned by source/exporter hashes; vg v1.76.1-89-gf2ed0bbd4 (Cactus v3.3.0 distribution) | IPD-IMGT/HLA v3.65.0 native catalogue; fold-specific HPRC or HPRC plus Arab/Saudi observed-haplotype graphs |
| HLA-HD | Unavailable; not run | Unavailable; not run |

[SOFTWARE_DATABASES.json](SOFTWARE_DATABASES.json) records the exact source and
checksum pins, including the IPD v3.65.0 commit and historical graph archive.
Unknown historical release labels remain explicit. The database version used
for inference and the nomenclature projection used for scoring are recorded
separately.

## Supported-resolution native baselines

A scored unit is a donor–locus diploid genotype pair. Each row below compares
that caller and genomic-reference T1K on the same eligible pairs and resolution.
`Called` counts eligible pairs with an admissible prediction; completed donors
measure execution coverage. Missing calls contribute to the eligible denominator.

| Caller | Donors completed | Supported endpoint | Correct / eligible | Called / eligible | T1K correct on matched pairs |
| --- | ---: | --- | ---: | ---: | ---: |
| OptiType | 86/86 | Class-I numeric two-field | 217/245 | 218/245 | 241/245 |
| SpecHLA | 86/86 | Numeric two-field, eight loci | 514/552 | 552/552 | 543/552 |
| HLA*LA | 86/86 | Common G classes, eight loci | 505/552 | 531/552 | 552/552 |

| Caller minus matched T1K | Difference (percentage points) | Family-bootstrap 95% interval | Family sign-flip p |
| --- | ---: | --- | ---: |
| OptiType, two-field | −9.80 | [−13.93, −5.76] | 0.000050 |
| SpecHLA, two-field | −5.25 | [−7.59, −3.11] | 0.000050 |
| HLA*LA, common G | −8.51 | [−10.56, −6.55] | 0.000050 |

These are historical or mixed-database comparator benchmarks. OptiType uses its
bundled class-I database. SpecHLA uses v3.65 annotation with v3.38 mapping/binning.
HLA*LA uses its historical native graph, with both methods projected through
current v3.65 G definitions for the common-G endpoint. Numeric ambiguity of a
native G class is retained as a resolution diagnostic. Comparisons use each
row's supported endpoint; the common-G and two-field scores answer different
questions. The scores use 4,000 family-bootstrap replicates and exact or
20,000-draw Monte Carlo family sign flips. Relationship metadata are incomplete
for **26 donors**, which use explicitly reported singleton fallback.

All 86 HLA*LA predictions passed final validation. The completion record includes
22 post-native validator failures recovered through validation of native allele
alternative lists, and eight compressed-input failures completed after retry.
Native predictions and failure/recovery provenance are retained. HLA-HD remains
blocked by unavailable distribution.

## Catalogue content and graph contribution

The catalogue endpoint scores exact named diploid genotypes over **541 eligible
donor–locus pairs** from 86 donors. A separate recovery endpoint covers **175
gene copies** without a genomic IPD match. The denominators distinguish genotype
accuracy from recovery of assembly sequences absent from the genomic database.

| Reference configuration | Named genotypes correct / 541 | Unmatched gene copies recovered / 175 |
| --- | ---: | ---: |
| Default IPD | 271/541 | 0/175 |
| Full genomic IPD | 489/541 | 0/175 |
| Full genomic IPD + flanked Asian HPRC/Arab/Saudi sequences | 444/541 | 61/175 |
| Full genomic IPD + normalized Asian HPRC/Arab/Saudi cores | 485/541 | 39/175 |
| Full genomic IPD + normalized donor-count-matched non-Asian cores | 468/541 | 4/175 |

Full genomic IPD improves named accuracy by **40.30 percentage points** relative
to default IPD (95% family interval [35.99, 44.65], p=0.000050). This establishes
reference completeness as a major contributor in the tested configuration.
Population-derived cores recover additional unmatched sequences while retaining
named accuracy near the genomic baseline: the Asian-core configuration differs
by −0.74 points (95% interval [−1.83, 0.19], p=0.293). Its unmatched-copy recovery
exceeds the normalized non-Asian control by 35/175 copies, or 20.00 percentage
points (95% interval [13.79, 26.63], p=0.000050).

Core normalization combines flank trimming, sequence deduplication and exclusion
of exact IPD-identical sequences. The normalized-core result is a later diagnostic
with a combined reference intervention. Training-source reconstruction found no
held-out donors among reconstructed training sources. The Asian configuration
includes East/South Asian HPRC and Arab/Saudi sequences; its non-Asian control is
matched by donor count. Exact recovered cores remain candidates for independent
sequence and boundary validation, particularly at DRB1.

With full genomic reference sequences and the expanded Asian panel, graph-union
retrieval in mode E3 gives **450/541** named genotypes versus **447/541** with
native-only placements: +0.55 percentage points, 95% interval [−0.74, 1.88],
p=0.581. Both arms share catalogue, fallback, rescoring and gates. The matched
full-genomic T1K baseline gives **489/541**; graph union differs by −7.21 points
(95% interval [−10.31, −4.31], p=0.000100). The results support reference and
catalogue improvements while leaving the aggregate graph contribution uncertain
in this pipeline. Original-reference experiments over 60 donors and 377 eligible
pairs are reported separately in the experiment release.

## Linkage and scope

At 10% and 25% of the original paired reads, the linked, independent and shuffled
arms share read likelihoods and marginal candidate priors. All eight matched
linkage contrasts across two panels and two fractions use **60 donors and 377
eligible genotype pairs**. Differences range from −0.27 to +0.27 percentage
points, every family interval includes zero, and every family sign-flip p is 1.
The tested linkage prior therefore has a null aggregate effect at these read
fractions. Absolute coverage was not measured for this contrast.

These results come from exposed developmental cross-validation. Regional read
recruitment starts from linear-reference MHC, alternative and HLA alignments
with mate rescue; the measured graph intervention acts after recruitment.
Untouched confirmatory evaluation and independent truth certification remain
next steps. The bounded execution is complete; broader DeepVariant-matched
variant, fold-safe RCCX and novel-recombinant simulation experiments remain
outside this completed scope.

## Evidence and verification

The authoritative sources are `A3_SCORES.json` in the final native snapshot and
`RESULTS.txt` / `SCIENTIFIC_VERDICT.json` under `scores-and-interpretation/` in
the experiment release. The original analysis-source hashes
and the extracted headline scores are recorded in
[RESULTS_SOURCES.json](RESULTS_SOURCES.json). Release manifests bind detailed
outputs and provenance. The [tracking sheet](https://docs.google.com/spreadsheets/d/18F0YMcjL8NUnPrjghI8z567m0oV4vYD7vHiM6KFASGo/edit)
provides the consortium status view.