DogoHLA: catalogue content and graph contribution
Updated UTC: 2026-10-20T23:45:42.318225+00:00
Execution status: complete_bounded_typing_experiments

Scope: 86 Asian donors for catalogue tests; 86 for genomic graph-retrieval;
60 for original-reference retrieval and 10%/25% read-fraction linkage.
This is exposed developmental CV.
An execution-complete result is not an untouched confirmatory benchmark.
Counts below cover completed available outputs; compare methods through
their matched contrasts, not through unequal availability denominators.
Known family groups are resampled together. Donors absent from the
relationship metadata are treated as singletons; independence is not certified.

1. Catalogue typing
ipd: donors=86; named=271/541 (50.1%); unnamed recovered=0/175; represented in training=0/0.
genomic: donors=86; named=489/541 (90.4%); unnamed recovered=0/175; represented in training=0/0.
genomic_hprc: donors=86; named=423/541 (78.2%); unnamed recovered=52/175; represented in training=52/60.
genomic_hprc_asian: donors=86; named=444/541 (82.1%); unnamed recovered=61/175; represented in training=61/69.
genomic_nonasian: donors=86; named=282/541 (52.1%); unnamed recovered=6/175; represented in training=6/11.
genomic_core_hprc: donors=86; named=491/541 (90.8%); unnamed recovered=33/175; represented in training=33/60.
genomic_core_hprc_asian: donors=86; named=485/541 (89.6%); unnamed recovered=39/175; represented in training=39/69.
genomic_core_nonasian: donors=86; named=468/541 (86.5%); unnamed recovered=4/175; represented in training=4/11.
Matched genomic minus ipd: +40.30 percentage points (n=541; family bootstrap95=[35.99257884972171, 44.64944649446495]; family sign-flip p=4.999750012499375e-05).
Matched genomic_hprc_asian minus genomic: -8.32 percentage points (n=541; family bootstrap95=[-11.41804788213628, -5.37037037037037]; family sign-flip p=4.999750012499375e-05).
Matched genomic_core_hprc_asian minus genomic: -0.74 percentage points (n=541; family bootstrap95=[-1.8281535648994516, 0.18726591760299627]; family sign-flip p=0.2928853557322134).
Matched genomic_core_hprc_asian minus genomic_core_nonasian: +3.14 percentage points (n=541; family bootstrap95=[1.308411214953271, 5.018587360594796]; family sign-flip p=0.0030998450077496125).
Matched genomic_core_hprc_asian minus genomic_hprc_asian: +7.58 percentage points (n=541; family bootstrap95=[4.562043795620438, 10.674157303370787]; family sign-flip p=4.999750012499375e-05).

The core-trimmed diagnostic trims flanks and deduplicates sequences; the
builder also skips exact IPD-identical sequences. It is a combined
reference-normalization intervention, not an isolated flank experiment.
Some retained cores have known genomic labels but different sequence
boundaries; this is not a strictly unnamed-sequence-only augmentation.
The pre-existing flanked contrast remains reported; core-trimmed results
are a later diagnostic, not a retroactively chosen primary endpoint.

2. Population content and its limits
hprc = East/South Asian HPRC training donors; hprc_asian additionally
includes Arab/Saudi donors. nonasian is matched by donor count, not by
unique sequence count. These labels must not be conflated.
all: 86 donors; 175 gene copies lack a genomic IPD match; training-core coverage={'hprc': 60, 'hprc_asian': 69, 'nonasian': 11}; extra from Arab/Saudi=9.
EAS: 50 donors; 99 gene copies lack a genomic IPD match; training-core coverage={'hprc': 27, 'hprc_asian': 29, 'nonasian': 8}; extra from Arab/Saudi=2.
SAS: 36 donors; 76 gene copies lack a genomic IPD match; training-core coverage={'hprc': 33, 'hprc_asian': 40, 'nonasian': 3}; extra from Arab/Saudi=7.
hprc mean per-fold catalogue size across eight loci: unique_flanked_sequences=482.2, unique_cores=375.0, unique_unnamed_cores=104.8.
hprc_asian mean per-fold catalogue size across eight loci: unique_flanked_sequences=802.6, unique_cores=597.0, unique_unnamed_cores=201.6.
nonasian mean per-fold catalogue size across eight loci: unique_flanked_sequences=525.8, unique_cores=431.0, unique_unnamed_cores=121.6.
This is an oracle coverage ceiling, not typing accuracy or certification of novel alleles.
Matched unnamed-copy recovery, genomic_hprc minus genomic_nonasian: 52 vs 6 / 175; change=+26.29pp; CI95=[18.96551724137931, 34.193548387096776]; family p=4.9998e-05.
Matched unnamed-copy recovery, genomic_hprc_asian minus genomic_nonasian: 61 vs 6 / 175; change=+31.43pp; CI95=[23.863636363636363, 39.24050632911393]; family p=4.9998e-05.
Matched unnamed-copy recovery, genomic_hprc_asian minus genomic_hprc: 61 vs 52 / 175; change=+5.14pp; CI95=[1.7647058823529411, 8.938547486033519]; family p=0.011719.
Matched unnamed-copy recovery, genomic_hprc_asian minus genomic: 61 vs 0 / 175; change=+34.86pp; CI95=[27.564102564102566, 42.42424242424242]; family p=4.9998e-05.
Matched unnamed-copy recovery, genomic_core_hprc_asian minus genomic_core_nonasian: 39 vs 4 / 175; change=+20.00pp; CI95=[13.793103448275861, 26.630434782608695]; family p=4.9998e-05.
Matched unnamed-copy recovery, genomic_core_hprc_asian minus genomic: 39 vs 0 / 175; change=+22.29pp; CI95=[15.853658536585366, 29.069767441860463]; family p=4.9998e-05.
Core-reference source provenance was reconstructed across merged contexts; no held-out donors were found in the reconstructed training sources.
Excluding DRB1, genomic_hprc_asian: unnamed recovery 23/65 (86 donors).
Excluding DRB1, genomic_core_hprc_asian: unnamed recovery 21/65 (86 donors).
Excluding DRB1, genomic_core_nonasian: unnamed recovery 3/65 (86 donors).

3. Graph-proposed alignment contribution
GRAPH_RETRIEVAL_ABLATION.json: 120/120 donor-panel arms.
hprc/E1: union 278/377 vs native-only 269/377; change=+2.39pp CI95=[1.0526315789473684, 3.926701570680628]; p=0.0039062; unnamed recovery 34 vs 33.
hprc/E2: union 274/377 vs native-only 268/377; change=+1.59pp CI95=[0.0, 3.3854166666666665]; p=0.146; unnamed recovery 22 vs 27.
hprc/E3: union 284/377 vs native-only 275/377; change=+2.39pp CI95=[0.5333333333333333, 4.370179948586118]; p=0.033203; unnamed recovery 27 vs 28.
hprc_asian/E1: union 287/377 vs native-only 282/377; change=+1.33pp CI95=[-0.2724795640326976, 3.0303030303030303]; p=0.22656; unnamed recovery 41 vs 40.
hprc_asian/E2: union 288/377 vs native-only 281/377; change=+1.86pp CI95=[-0.26666666666666666, 3.926701570680628]; p=0.14346; unnamed recovery 26 vs 33.
hprc_asian/E3: union 303/377 vs native-only 290/377; change=+3.45pp CI95=[1.0582010582010581, 5.974025974025974]; p=0.012549; unnamed recovery 30 vs 37.
GRAPH_GENOMIC_RETRIEVAL_ABLATION.json: 172/172 donor-panel arms.
hprc/E1: union 437/541 vs native-only 448/541; change=-2.03pp CI95=[-4.174573055028463, 0.0]; p=0.090195; unnamed recovery 51 vs 52.
hprc/E2: union 445/541 vs native-only 459/541; change=-2.59pp CI95=[-5.147058823529412, -0.1834862385321101]; p=0.064747; unnamed recovery 30 vs 38.
hprc/E3: union 427/541 vs native-only 432/541; change=-0.92pp CI95=[-2.536231884057971, 0.7233273056057866]; p=0.36285; unnamed recovery 36 vs 46.
hprc_asian/E1: union 449/541 vs native-only 459/541; change=-1.85pp CI95=[-3.910614525139665, 0.19193857965451055]; p=0.13469; unnamed recovery 60 vs 61.
hprc_asian/E2: union 451/541 vs native-only 466/541; change=-2.77pp CI95=[-5.291970802919708, -0.18832391713747645]; p=0.049948; unnamed recovery 36 vs 46.
hprc_asian/E3: union 450/541 vs native-only 447/541; change=+0.55pp CI95=[-0.7352941176470589, 1.8796992481203008]; p=0.58105; unnamed recovery 43 vs 55.
Competitive comparison to full-genomic IPD (matched donors):
DogoHLA/hprc/E1 minus genomic: n=377; gains=12, losses=70; change=-15.38pp; CI95=[-19.733333333333334, -11.256544502617801]; p=4.9998e-05.
DogoHLA/hprc/E2 minus genomic: n=377; gains=11, losses=73; change=-16.45pp; CI95=[-20.786516853932586, -12.335958005249344]; p=4.9998e-05.
DogoHLA/hprc/E3 minus genomic: n=377; gains=13, losses=65; change=-13.79pp; CI95=[-18.230563002680967, -9.585492227979275]; p=4.9998e-05.
DogoHLA/hprc_asian/E1 minus genomic: n=377; gains=13, losses=62; change=-13.00pp; CI95=[-16.93548387096774, -8.994708994708995]; p=4.9998e-05.
DogoHLA/hprc_asian/E2 minus genomic: n=377; gains=11, losses=59; change=-12.73pp; CI95=[-16.711590296495956, -8.785529715762275]; p=4.9998e-05.
DogoHLA/hprc_asian/E3 minus genomic: n=377; gains=12, losses=45; change=-8.75pp; CI95=[-12.631578947368421, -5.121293800539084]; p=4.9998e-05.
DogoHLA_genomic/hprc/E1 minus genomic: n=541; gains=13, losses=65; change=-9.61pp; CI95=[-12.86231884057971, -6.4935064935064934]; p=4.9998e-05.
DogoHLA_genomic/hprc/E2 minus genomic: n=541; gains=14, losses=58; change=-8.13pp; CI95=[-11.04868913857678, -5.273069679849341]; p=4.9998e-05.
DogoHLA_genomic/hprc/E3 minus genomic: n=541; gains=16, losses=78; change=-11.46pp; CI95=[-15.053763440860216, -7.899461400359066]; p=4.9998e-05.
DogoHLA_genomic/hprc_asian/E1 minus genomic: n=541; gains=13, losses=53; change=-7.39pp; CI95=[-10.018552875695732, -4.909090909090909]; p=4.9998e-05.
DogoHLA_genomic/hprc_asian/E2 minus genomic: n=541; gains=13, losses=51; change=-7.02pp; CI95=[-9.742647058823529, -4.251386321626617]; p=4.9998e-05.
DogoHLA_genomic/hprc_asian/E3 minus genomic: n=541; gains=18, losses=57; change=-7.21pp; CI95=[-10.30741410488246, -4.308797127468582]; p=9.9995e-05.
Both arms retain the same catalogue, fallback, rescoring model and gates.
Only graph-proposed placements are removed in the native-only arm.
A positive result identifies a contribution of this graph mapper within
this pipeline; it does not show that every graph method is superior.
The original-reference replay does not establish superiority to genomic IPD.

4. Linkage at reduced read fractions
dsf100/hprc linked vs independent: 37/480 selected candidate pairs changed; eligible named calls corrected=8, harmed=7.
dsf100/hprc linked vs shuffled: 41/480 selected candidate pairs changed; eligible named calls corrected=8, harmed=8.
dsf100/hprc_asian linked vs independent: 46/480 selected candidate pairs changed; eligible named calls corrected=8, harmed=7.
dsf100/hprc_asian linked vs shuffled: 53/480 selected candidate pairs changed; eligible named calls corrected=7, harmed=8.
dsf100/hprc linked minus independent: donors=60; n=377; change=+0.27pp; CI95=[-1.5915119363395225, 2.1220159151193636]; p=1; status=complete.
dsf100/hprc linked minus shuffled: donors=60; n=377; change=+0.00pp; CI95=[-2.099737532808399, 2.1164021164021163]; p=1; status=complete.
dsf100/hprc_asian linked minus independent: donors=60; n=377; change=+0.27pp; CI95=[-1.5831134564643798, 2.1563342318059298]; p=1; status=complete.
dsf100/hprc_asian linked minus shuffled: donors=60; n=377; change=-0.27pp; CI95=[-2.1220159151193636, 1.6129032258064515]; p=1; status=complete.
dsf250/hprc linked vs independent: 17/480 selected candidate pairs changed; eligible named calls corrected=3, harmed=3.
dsf250/hprc linked vs shuffled: 18/480 selected candidate pairs changed; eligible named calls corrected=2, harmed=2.
dsf250/hprc_asian linked vs independent: 24/480 selected candidate pairs changed; eligible named calls corrected=4, harmed=4.
dsf250/hprc_asian linked vs shuffled: 24/480 selected candidate pairs changed; eligible named calls corrected=3, harmed=4.
dsf250/hprc linked minus independent: donors=60; n=377; change=+0.00pp; CI95=[-1.2987012987012987, 1.3054830287206267]; p=1; status=complete.
dsf250/hprc linked minus shuffled: donors=60; n=377; change=+0.00pp; CI95=[-1.0554089709762533, 1.0638297872340425]; p=1; status=complete.
dsf250/hprc_asian linked minus independent: donors=60; n=377; change=+0.00pp; CI95=[-1.3404825737265416, 1.5706806282722514]; p=1; status=complete.
dsf250/hprc_asian linked minus shuffled: donors=60; n=377; change=-0.27pp; CI95=[-1.5873015873015872, 1.0810810810810811]; p=1; status=complete.

Current statistical interpretation (provisional until execution is complete):
- genomic minus ipd: positive aggregate difference in this tested configuration (n=541).
- genomic_hprc_asian minus genomic: negative aggregate difference in this tested configuration (n=541).
- genomic_core_hprc_asian minus genomic: no clear aggregate evidence under both family interval and sign-flip checks (n=541).
- genomic_core_hprc_asian minus genomic_core_nonasian: positive aggregate difference in this tested configuration (n=541).
- genomic_core_hprc_asian minus genomic_hprc_asian: positive aggregate difference in this tested configuration (n=541).
- GRAPH_RETRIEVAL_ABLATION.json hprc/E3 union minus native-only: positive aggregate difference in this tested configuration (n=377).
- GRAPH_RETRIEVAL_ABLATION.json hprc_asian/E3 union minus native-only: positive aggregate difference in this tested configuration (n=377).
- GRAPH_GENOMIC_RETRIEVAL_ABLATION.json hprc/E3 union minus native-only: no clear aggregate evidence under both family interval and sign-flip checks (n=541).
- GRAPH_GENOMIC_RETRIEVAL_ABLATION.json hprc_asian/E3 union minus native-only: no clear aggregate evidence under both family interval and sign-flip checks (n=541).
- dsf100 hprc linked minus independent: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf100 hprc linked minus shuffled: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf100 hprc_asian linked minus independent: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf100 hprc_asian linked minus shuffled: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf250 hprc linked minus independent: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf250 hprc linked minus shuffled: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf250 hprc_asian linked minus independent: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- dsf250 hprc_asian linked minus shuffled: no clear aggregate evidence under both family interval and sign-flip checks (n=377).
- unnamed recovery: genomic_hprc minus genomic_nonasian: positive aggregate difference in this tested configuration (n=175).
- unnamed recovery: genomic_hprc_asian minus genomic_nonasian: positive aggregate difference in this tested configuration (n=175).
- unnamed recovery: genomic_hprc_asian minus genomic_hprc: positive aggregate difference in this tested configuration (n=175).
- unnamed recovery: genomic_hprc_asian minus genomic: positive aggregate difference in this tested configuration (n=175).
- unnamed recovery: genomic_core_hprc_asian minus genomic_core_nonasian: positive aggregate difference in this tested configuration (n=175).
- unnamed recovery: genomic_core_hprc_asian minus genomic: positive aggregate difference in this tested configuration (n=175).
- DogoHLA_genomic/hprc/E1 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
- DogoHLA_genomic/hprc/E2 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
- DogoHLA_genomic/hprc/E3 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
- DogoHLA_genomic/hprc_asian/E1 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
- DogoHLA_genomic/hprc_asian/E2 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
- DogoHLA_genomic/hprc_asian/E3 minus full-genomic IPD: negative aggregate difference in this tested configuration (n=541).
Read fractions are 10% and 25% of the original paired reads, not measured
absolute coverage levels. Linked/independent/shuffled arms share the
same read likelihoods and marginal candidate priors.

5. Basic explanation and limits of any negative result
- Full-genomic IPD repairs a major reference-completeness confound.
  Improvements over shortened default IPD must not be assigned to graphs.
- Population haplotypes add relevant gene sequences missing from IPD.
  This can improve unnamed-sequence recovery even when named accuracy
  does not improve; the named endpoint excludes unnamed truth.
- The current graph caller selects observed candidate paths. It cannot
  synthesize an allele absent from every candidate, even if some of its
  component variation is in the graph. Missing-catalogue failures are
  therefore distinct from mapping or inference failures.
- Graph retrieval can help particular loci; aggregate gains must survive
  the native-only placement control and the strong genomic baseline.
- Regional recruitment starts from linear-reference MHC, alternative
  and HLA alignments, with mate rescue. Graph inference cannot recover
  pairs discarded at recruitment. These tests therefore do not measure
  a possible benefit of graph-based recruitment from the full WGS.
- A null LD result applies to the tested prior, read evidence and depths.
  It does not establish absence of biological linkage disequilibrium.
- DRB1 dominates many unmatched counts. Exact annotation-core matches
  need independent raw-read/boundary validation before novelty claims.
- All comparisons remain developmental. Multiple modes and diagnostic
  endpoints are reported rather than selecting the best result.

6. Supported-resolution comparator benchmarks
OptiType/SpecHLA use two-field comparisons; HLA*LA uses common G classes.
Unresolved numeric projections of HLA*LA G classes are not wrong two-field calls.
SpecHLA uses 3.65 annotation with 3.38 mapping/binning; OptiType
uses its bundled class-I database; HLA*LA uses its historical graph.
They are not fully matched 3.65-database benchmarks.
Detailed locus, region and stricter-resolution results: baseline-completion-20261020/A3_SCORES.txt.
optitype: completed donors=86/86; two-field correct=217/245; called=218; matched with genomic T1K=245 eligible pairs.
  optitype minus genomic T1K: -9.80pp; CI95=[-13.934426229508198, -5.761316872427984]; family p=4.9998e-05.
spechla: completed donors=86/86; two-field correct=514/552; called=552; matched with genomic T1K=552 eligible pairs.
  spechla minus genomic T1K: -5.25pp; CI95=[-7.592592592592593, -3.1135531135531136]; family p=4.9998e-05.
hlala: completed donors=86/86; common-G correct=505/552; called=531; matched with genomic T1K=552 eligible pairs.
  hlala minus genomic T1K: -8.51pp; CI95=[-10.564663023679417, -6.545454545454546]; family p=4.9998e-05.
HLA-HD remains blocked by unavailable distribution; no access-request message was sent.

Broader research plan disposition:
- HLA_LA: All86 common-G baseline complete; historical graph and label-resolution limitations remain; see A3 receipts
- HLA_HD_SpecHLA_OptiType: SpecHLA and OptiType all86 supported-resolution benchmarks complete; HLA-HD distribution unavailable; see A3 status
- truth_raw_HiFi_QC: Not yet certified; assembly-based endpoints exploratory
- untouched_confirmatory_set: Not frozen/admitted
- matched_DeepVariant_variant_H2: Not implemented/run; old vg/bcftools results confounded
- fold_safe_RCCX_H3: Not yet rebuilt/scored
- novel_recombinant_simulation_H1: Not yet implemented/run