How dismech Represents Antigens on B and T Cells: Gap Analysis (2026-09-03)

A census of how the knowledge base records antigen identity in immune and autoimmune entries, and whether the recorded antigen can be attributed to the lymphocyte lineage that recognises it.

Scope, as of 2026-09-03: all 2,535 files in kb/disorders/ plus all 167 in kb/modules/ (2,702 entries), narrowed to a 573-entry immune cohort. Every number below is regenerable with scripts/immune_antigen_audit.py, but this is a dated snapshot and the KB grows daily — a run today scans more entries and will not reproduce these totals verbatim. The conclusions are robust to that drift: re-run on 2026-09-07 against 2,881 entries, the headline 64.0% / 36.0% split and the zero-bindable-HLA-rows result of §5 were unchanged. Nothing here is gated, and no KB or schema file was changed.

just immune-antigen-audit                                # the run behind this report
just immune-antigen-audit --format tsv --out /tmp/antigen.tsv
just immune-antigen-audit --entry Celiac_Disease

The finding in one line

The schema has no antigen concept. The string antigen does not appear once in any of the 21 files under src/dismech/schema/ (7 top-level YAML, 13 more under classifications/, and a README) — not as a class, not as a slot, not as an enum value. Every antigen in the knowledge base is free text.

The measurable consequence, stated over the denominator that can actually carry the answer: only three classes own a cell-type slot — Pathophysiology, Biochemical and ExperimentalModel (plus FunctionalEffect.affected_cell_types). Of the 3,122 objects that name an antigen, 428 are instances of one of those classes, and 274 of them (64.0%) name no B, T or antigen-presenting lineage. So for roughly two thirds of the antigen mentions that could say which lymphocyte sees the antigen, the answer is not recorded and can only be had by reading the sentence.

Antigen-naming objects in a class that can own a cell-type slot	Count	Share
naming a B, T or antigen-presenting lineage	154	36.0%
naming none	274	64.0%

The remaining 2,694 (86.3% of 3,122) are evidence items, treatments, phenotypes, references, genetic rows and descriptors — classes with no cell-type slot at all. For those, an absent lineage is not a curation gap, it is structurally impossible, so they are excluded from the rate above.

Counted across all 3,122 objects the figure is 2,967 (95.0%) with no lineage, of which 2,892 have an empty or absent cell-type slot and 75 have one naming a non-lymphoid cell. That number is real but inflated by its denominator; 64% is the one a design decision should be taken on.

Antigen-naming objects (all 3,122)	Count	Share
with a B-lineage cell type	58	1.9%
with a T-lineage cell type	87	2.8%
with an antigen-presenting cell	49	1.6%
no B/T/APC lineage — cell-type slot empty or absent	2,892	92.6%
no B/T/APC lineage — slot set, names a non-lymphoid cell	75	2.4%

Per entry: of the 323 immune-cohort entries that name an antigen anywhere, 226 (70%) attribute none of them to a B or T lineage.

1. The cohort, and how little of it is structurally identifiable

Cohort membership established by	Entries
`classifications.harrisons_chapter: IMMUNE_RHEUMATOLOGIC`	104
prose only — matches `autoimmun` / `autoantibod` / `autoantigen`	263
prose only — matches `immunodeficien` / `immune-mediated` / `hypersensitiv` / `vasculit`	206
total	573

Only 18% of the immune cohort is identifiable from a curated classification. The rest had to be found by grep. That is a prior gap, not the subject of this report, but it bounds what any antigen-focused query could return today: a consumer asking "show me the autoimmune diseases" gets 104 of 573.

Within the cohort, 323 entries name an antigen and 250 name none — including entries where an antigen is the defining fact of the disease (see §6).

2. Where antigen text actually lives

Objects in the immune cohort whose own scalar fields name an antigen:

Top-level slot	Objects
`pathophysiology`	1,083
`treatments`	412
`biochemical`	239
`phenotypes`	208
`discussions`	195
`references`	146
`genetic`	138
`diagnosis`	112
`mechanistic_hypotheses`	98
`has_subtypes`	92
`environmental`	87

Raw pattern frequency across the cohort: antigen (generic) 2,005; bare anti-X 1,480; anti-X antibody 1,331; autoantigen 213; epitope 200; antigen present* 191; molecular mimicry 122; neoantigen 43; citrullinat* 33; superantigen 30; deamidat* 20; epitope spreading 11.

2a. The free-text form cannot distinguish an autoantigen from a drug target

The same anti-X string carries four different kinds of fact, and nothing in the YAML separates them. Ranking the captured X by how many entries use it:

Outside `treatments` (candidate autoantigens)	Inside `treatments` / `clinical_trials` (drug targets)
TNF (10), neutrophil cytoplasmic (10), NMDAR (9), CD20 (7), GBM (5), AChR (4), nuclear (4), Mullerian (4), AQP4 (3), GM1 (3), Sm (3), Ro (3), La (3), Hu (3), TPO (3), DNase (3)	CD20 monoclonal (26), TNF (25), CD20 (20), PD-1 (7), IL-6 (6), IgE (4), C5 (4), VEGF (4), IL-5 (4)

Reading down the left column: neutrophil cytoplasmic, NMDAR, AChR, AQP4, Sm, Ro, La, Hu and TPO are genuine autoantigens; TNF and CD20 are therapeutic targets that leaked out of treatments via evidence prose; DNase is anti-DNase B titre, an anti-streptococcal serology and not a self-antigen at all; and Mullerian is anti-Müllerian hormone, which is not an antibody target in any sense — the pattern matched anti-Müllerian.

An autoantigen, a pathogen antigen, a hormone name and a monoclonal-antibody target are indistinguishable to any consumer of this data. That is the cost of having no slot: the distinction exists only in a curator's head.

3. Sense 1 — autoantigen identity

116 autoantibody rows in biochemical across 66 entries. Of those, 11 carry a biomarker_term; the other 105 are name strings.

Where a binding does exist it is almost always to a phenotype, not to the antigen. Adult-Onset_Myasthenia_Gravis is the best-curated case in the KB:

biochemical:
- name: Anti-Acetylcholine Receptor Antibody
  biomarker_term:
    preferred_term: Anti-acetylcholine receptor antibody positivity
    term: {id: HP:6001064, label: Anti-acetylcholine receptor antibody positivity}

This is a real improvement over a bare string, and HPO supports it further than the KB uses it — 33 HP terms with antibody in the label are already in cache/hp/terms.csv — 26 of them of the form ... antibody positivity, headed by HP:0030057 Autoimmune antibody positivity — covering ANA, anti-dsDNA-adjacent, ANCA/MPO/PR3, anti-Ro/SS-A, anti-cardiolipin, anti-β2GPI, anti-MuSK and more. Three entries use them.

But note what the HP binding says: it asserts the patient is seropositive. It does not identify the antigen as a molecular entity, so it cannot be joined to the gene that encodes it, to the tissue that expresses it, or to a T-cell response against the same protein.

biomarker_term bindings across kb/disorders/ and kb/modules/ by prefix — CHEBI 452, NCIT 269, HP 71, hgnc 4, GO 2, plus 22 blocks carrying no term.id at all — show the slot is already used heterogeneously, so an antigen-as-gene-product binding would not be unprecedented. It would just be undeclared.

3a. The `genetic` block is being used as an antigen slot, with a disclaimer

GeneDiseaseRelationshipEnum has ten values — CAUSATIVE, RISK_FACTOR, PROTECTIVE, MODIFIER, SUSCEPTIBILITY, SOMATIC_DRIVER, COOPERATING, BIOMARKER, DISPUTED, UNKNOWN. None of them means "the protein this gene encodes is the target of the autoimmune response."

Two entries record the autoantigen gene anyway and work around the missing value in free text:

Anti-GBM_Disease — COL4A3 (alpha-3 type IV collagen) autoantigen target gene, with association: Autoantigen target (not a causal germline mutation) and relationship_type left empty. The curator wrote the disclaimer into the data because the enum could not carry it.
Antisynthetase_Syndrome — HARS1 (autoantigen; anti-Jo-1 target), association: Autoantigen, with a note that "HARS1 is the autoantigen, not a mutated disease gene."

A third case shows the failure mode this invites. Membranous_Nephropathy records PLA2R1 susceptibility locus with association: GWAS. PLA2R1 is both a GWAS hit and the defining autoantigen; the entry captures the first fact and silently drops the second, because genetic has a value for one and not the other. Its own biochemical block, meanwhile, carries Anti-PLA2R autoantibodies as an unbound string. The same protein appears twice in one file, in two vocabularies, with no link between them.

Pemphigus_Vulgaris shows the opposite failure: DSG1 and DSG3 — arguably the most precisely characterised autoantigen pair in medicine — appear nowhere in its genetic block, which holds only three HLA rows. They exist solely inside strings like Anti-Desmoglein 3 Antibodies.

4. Sense 2 — B versus T attribution in the pathograph

Across the immune cohort's pathophysiology nodes:

	Nodes
carrying a B-lineage cell type	203
carrying a T-lineage cell type	384
carrying both	56

Per entry: 87 have both B and T nodes, 132 have T only, 45 have B only, and 309 of 573 have neither.

The cell-type layer itself is in good shape — CL:0000236 B cell (163 uses), CL:0000084 T cell (150), CL:0000625 CD8-positive, alpha-beta T cell (82), CL:0000624 CD4-positive, alpha-beta T cell (72), CL:0000786 plasma cell (58), CL:0000844 germinal center B cell (11), CL:0000980 plasmablast (9). The gap is not the vocabulary. It is that a node says which cell is present and, separately, prose says which antigen is involved, with no edge between them.

The 56 both-lineage nodes are where this bites hardest: a node carrying T cell and B cell together is exactly the node where the two lineages might be seeing different antigens, and it is structurally incapable of saying so.

The KB's only antigen-centric module makes the point. molecular_mimicry_autoimmunity collapses the entire lymphocyte response into one node:

- name: Cross-Reactive Autoreactive Lymphocyte Activation
  cell_types: [T cell, B cell]

Its own description says the cross-reactive antigen "activates autoreactive T and/or B cells" — the and/or is doing load-bearing work that the data model cannot express. The module names no antigen structurally, and its downstream node Epitope Spreading and Autoimmune Amplification — the mechanism by which the antigen set changes over time — has no antigens to spread between.

5. Senses 3 and 4 — HLA restriction and surface markers

HLA, and why almost none of it is backfillable. 1,289 prose mentions across 106 cohort entries; 48 entries bind an HLA-* gene via gene_term. At the row level there are 100 genetic rows naming HLA, of which 59 are bound and 41 are not.

The unbound 41 look like a backlog and are not one. Not a single one names an HGNC gene symbol:

Of the 41 unbound HLA rows, the name is	Rows
an HGNC gene symbol (`HLA-DRB1`, `HLA-B`, …) — bindable	0
a serotype (`HLA-DQ2`, `HLA-B27`, `HLA-DR3`), an allele (`HLA-DRB103:01`, `HLA-DQB106:02`), a haplotype (`HLA-DR3-DQ2`) or a region	41

A serotype is not a gene. HLA-DQ2 is a serological specificity carried by an HLA-DQA1/HLA-DQB1 haplotype, so there is no single HGNC identifier that means it, and binding one would assert something false. Six of the 41 rows say exactly this in their own notes, and the two entries where HLA restriction is the mechanism are among them:

Celiac_Disease — "No gene_term is bound because DQ2 is a serotype encoded by an HLA-DQA1/HLA-DQB1 haplotype rather than a single gene."

Type_I_Diabetes — "HLA-DQ2 is a haplotype-level risk label, not a single HGNC-resolvable gene."

That is the ontology term contract working as intended — no term beats a bad one — and it is prior art, not a gap. The other 35 unbound rows have no such note, but they are the same kind of name; what they are missing is the written-down reasoning, not a binding.

The real gap here is a different one, and no amount of gene_term backfill touches it: the KB has nowhere to put a serotype, an allele or a haplotype. HLA-DQ2, HLA-B27 and DRB1*15:01 are typed into a free-text name because GeneDescriptor binds HGNC genes and nothing binds an HLA allele. Even where a gene does bind, it loses the allele — hgnc:4948 is HLA-DRB1, which cannot distinguish the DR15 haplotype of anti-GBM disease from the shared-epitope DRB1*04 of rheumatoid arthritis. Nowhere in the KB is a peptide linked to the allele that presents it.

Surface / lineage markers. These divide sharply by whether a CL term happens to encode them. Counted over the immune cohort, by whether the marker appears in a curated identity field (name / preferred_term / label) or in prose:

Marker	Mentions	In an identity field	In prose
CD4	952	226 (23.7%)	726
CD8	737	183 (24.8%)	554
CD27 †	245	12 (4.9%)	233
CD3	141	3 (2.1%)	138
CD20	140	4 (2.9%)	136
CD19	104	4 (3.8%)	100

The table stops at 100 mentions; below it the script also reports CD21 (25), CD38 (10), CD79a (3), CD138 (1) and CD22 (1).

CD4 and CD8 are the best-represented, but only as a side effect: they ride along inside CL labels such as CD8-positive, alpha-beta T cell. CD19, CD20 and CD3 — the B-lineage and pan-T markers, and the ones that matter for therapy — have no CL term to ride on and are therefore almost entirely prose. 79 entries mention rituximab; the CD20 it depletes is a sentence in a description, not a target.

† CD27 is in the table because it outranks the three rows below it, but it is not the same signal, and it should not be read as one. 217 of its 245 mentions (89%) come from a single entry, CD27-related_lymphoproliferative_and_immune_disorder, which is named after the gene; the other 28 are scattered across 12 entries. And none of its 12 identity-field hits is a CL label — they are the disease name, its MONDO disease_term, a pathophysiology node name, two diagnosis names, and HGNC gene descriptors under pathophysiology, genetic and diagnosis.

That makes CD27 a different failure mode from CD4/CD8, and a sharper one for this report's argument. Where CD27 is bound at all, it is bound as a gene. The KB can say "the CD27 gene" and it can say "a memory B cell" — it has no way to say that the cell displays the molecule. A lineage marker is exactly a molecule displayed by a cell, so the one relation that would make it queryable is the one missing, and binding the gene does not supply it.

6. Worked exemplars

`Celiac_Disease` — the textbook B/T divergence, entirely unstructured

Coeliac disease is the canonical case: the T-cell antigen is a deamidated gliadin peptide presented on HLA-DQ2/DQ8, while the dominant B-cell antigen is tissue transglutaminase (TG2), the very enzyme that performs the deamidation. Two different molecules, two different lineages, one disease. What the entry actually holds:

pathophysiology: Gluten-Triggered Immune Response (cell type: T Helper Cell) and Autoantibody Production (cell type: Plasma Cell) — the two lineages are correctly separated into distinct nodes, and neither node names its antigen.
biochemical: Anti-tTG IgA, Anti-Endomysial Antibodies, Anti-DGP Antibodies — three unbound strings, with no biomarker_term key on any of them. Note Anti-DGP is a B-cell readout of the T-cell antigen; nothing records that.
genetic: TGM2, the B-cell autoantigen, is absent. The block holds HLA-DQ2 and HLA-DQ8 with gene_term: null, plus ten bound susceptibility genes (IL2, IL21, BACH2, PTPN22 …).
environmental: Gluten Exposure has no exposure_term and no influences_mechanisms, so the T-cell antigen is not in the pathograph at all. Only Gastrointestinal Infections and Gut Microbiome are linked.

The entry is not badly curated — the node split is right and the genetics are bound. It is that the one fact this report is about has nowhere to go.

`Type_I_Diabetes` — the autoantibody panel is simply missing

T1D is clinically defined by islet autoantibody status, and the antigens are named in every guideline. In the entry:

GAD65, IA-2, ZnT8 — zero occurrences each.
biochemical holds Blood Glucose, Hemoglobin A1c (HbA1c) and an NMR metabolomic risk score. No autoantibody rows at all.
autoantibody appears 16 times, all in prose, and eight of those are evidence explanation fields apologising for the gap — "Supports T-cell mediated cytotoxicity but does not discuss autoantibody production", and seven near- identical siblings.

The node Autoimmune Destruction of Beta Cells lists seven cell types (CD8+ cytotoxic T, CD4+ T, Treg, B cell, DC, macrophage) on a single node — a rich cellular picture with no antigen attached to any of it.

`Pemphigus_Vulgaris` — a compartmentalised antigen story held in prose

The entry curates the desmoglein compensation theory well, including a dedicated node Desmoglein Compensation and Lesion Distribution whose description states that mucosal-dominant PV has anti-Dsg3 only while mucocutaneous PV has anti-Dsg3 plus anti-Dsg1. This is precisely an antigen-stratified subtype claim — and it is a paragraph. has_subtypes is absent, DSG1/DSG3 are not in genetic, and the two biochemical antibody rows are unbound.

`Adult-Onset_Myasthenia_Gravis` — the current ceiling

Three autoantibody rows bound to HP autoantibody-positivity terms; three HLA genes bound to HGNC; a node correctly separating germinal-centre B cells and T follicular helper cells in the hyperplastic thymus, whose description explains that thymic myoid cells express the autoantigen in situ. This is as far as the present schema goes, and it still cannot state that the AChR seen by the B cell and the AChR peptide seen by the T cell are the same protein.

7. What already exists to build on

Most of this requires no new ontology, but two things do, and they are not the same size. HLA is cheap: MRO exists and is one ols: configuration line (§7a). Naming an antigen by protein accession is not — dismech has no UniProt binding at all, and OLS/OAK do not serve UniProt, so that one needs a real decision rather than a config change (§7a).

Need	Existing resource
Antigen as a protein	`GeneProductDescriptor` (NCIT gene-product hierarchy), `ProteinComplexDescriptor`
Antigen as a gene	`GeneDescriptor` (HGNC), already used for COL4A3, HARS1, PLA2R1
Antigen as a small molecule / hapten	`ChemicalEntityDescriptor` (CHEBI)
Dietary or environmental antigen	`ExposureDescriptor` (ECTO), `FoodDescriptor` (FOODON)
Seropositivity	26 of those 33 cached HP terms are `... antibody positivity` terms
Recognising lineage	B/T-lineage CL bindings already on 531 cohort pathophysiology nodes (2,158 carry some `cell_types`)
Attaching an antigen to a node	the `<kind>#<name>` entity-reference grammar
Recording how an antigen acts	the `ModelMechanismLink` / `influences_mechanisms` link-object pattern
HLA serotype, allele, haplotype, presenting complex	MRO (IEDB's MHC Restriction Ontology) — not yet in `conf/oak_config.yaml`; see §7a
Antigen as a protein, by accession	UniProt is the form IEDB publishes antigens in, and dismech has no slot that accepts it — see §7a

The pattern the KB already uses for exactly this shape of problem is the link object: treatments.target_mechanisms, environmental.influences_mechanisms, and animal_models.modeled_mechanisms all attach a typed, evidence-bearing edge from an entity to a pathograph node. An antigen link would be the same shape — antigen descriptor, target node, a recognising-lineage value (B / T / both), and its own evidence — and would inherit the pathograph rendering and export for free.

7a. Alignment with IEDB

Added 2026-09-07. All IEDB figures in this section were queried on that date against https://query-api.iedb.org; they are a snapshot, like the rest of this document.

The four senses this report separates are not novel distinctions. The Immune Epitope Database has modelled them for twenty years, its data is public domain, and its identifiers are already in the CURIE form dismech uses. Two things follow that change the recommendations above.

The HLA gap has an ontology, and §5 stops one step short of it

§5 concludes that "the KB has nowhere to put a serotype, an allele or a haplotype." That is true of dismech's descriptors and remains the finding. But the resource that fills it exists, is maintained by IEDB, and reaches dismech through the adapter pattern conf/oak_config.yaml already uses:

§5 calls this unbindable	MRO term
`HLA-DQ2` (serotype)	`MRO:0000283` HLA-DQ2 serotype
`HLA-B27` (serotype)	`MRO:0000217` HLA-B27 serotype
`HLA-DRB1*03:01` (allele)	`MRO:0000703` HLA-DRB1*03:01 chain
the presenting molecule	`MRO:0001284` HLA-DRB1*03:01 protein complex
`HLA-DR3-DQ2` (haplotype)	`MRO:0000005` MHC haplotype, with the `MRO:0000000` haplotype member of relation †

MRO (the MHC Restriction Ontology, PMID:26759709) is an IEDB product. It carries 51,436 terms at version 2026-08-26, is loaded in OLS, and resolves today through the same ols: adapter as HP, GO and CL:

uv run runoak -i ols:mro info MRO:0000283 MRO:0001620 MRO:0000217
# MRO:0000283 ! HLA-DQ2 serotype
# MRO:0001620 ! HLA protein complex with DQ2 serotype
# MRO:0000217 ! HLA-B27 serotype

† MRO:0000000 is an object property, not a class. Its label is as given, but runoak -i ols:mro info MRO:0000000 returns 404 — that route serves classes — and a property cannot be bound through a reachable_from dynamic enum the way the five class terms above can. The config change below covers the classes; expressing "this allele is a member of that haplotype" needs its own answer.

So adding MRO: ols:mro to conf/oak_config.yaml is a one-line change, after which HLA serotypes and alleles — the classes, at least — validate and cache exactly like every other bound term. This does not reopen the backfill §5 closed: the 41 unbound rows still name no HGNC gene, and binding them to gene_term would still be wrong. It changes which slot they are waiting for, not whether they are waiting.

MRO also separates the chain from the protein complex — the distinction §5 needs when it observes that hgnc:4948 (HLA-DRB1) cannot tell the DR15 haplotype of anti-GBM disease from the DRB1*04 shared epitope of rheumatoid arthritis. The complex term is what a peptide is presented by, so it is the natural target of the peptide-to-allele link §5 says is missing everywhere.

IEDB's schema answers the report's central question, and answers it asymmetrically

bcell_search and tcell_search are parallel tables over the same assay corpus. Both carry parent_source_antigen_iri (a UniProt accession), disease_iris, assay_iris (OBI), host_organism_iri (NCBITaxon), qualitative_measure and pubmed_id. They differ in one place, and the difference is not incidental:

Human-host records, queried 2026-09-07	B cell	T cell
with a disease assigned	1,352,189	268,067
with an MHC allele (`mhc_allele_iri`, an MRO term)	0	297,935

MHC restriction is a field only the T-cell table ever populates. That is the strongest available argument against modelling this in dismech as one generic "antigen" slot shared by both lineages: the recognising lineage determines which fields are even meaningful. A B-cell antigen link needs an antigen and an isotype; a T-cell antigen link needs an antigen, an MRO restriction and a presentation relation. A single slot would leave one of those two permanently half-empty, which is the failure mode §3a already documents in the genetic block.

The Celiac exemplar, tested

§6 asserts that in celiac disease B cells target tissue transglutaminase while T cells target deamidated gliadin peptides on HLA-DQ2/DQ8, and notes that the entry names the antigen on neither node. IEDB bears out the half of that claim which is a clean separation, and qualifies the other half:

Celiac disease, IEDB, queried 2026-09-07	B cell	T cell
assay records	824	2,748
against TGM2 (`UNIPROT:P21980`)	256–472 (see §9)	0
carrying an MRO restriction	0	1,876

Zero of 2,748 celiac T-cell records are against TGM2 — a result two independent counting methods agree on exactly, and the sharpest single finding here. The B-side claim needs more care than §6 gives it: TGM2 is heavily represented (256 of 824 by server-side filter, 472 by counting the returned column — §9 explains the spread), but it is not the top B-cell antigen. Gliadin UNIPROT:D2T2K3 is, at 318.

So the textbook framing holds asymmetrically rather than as a clean swap: gliadin is seen by both lineages, and TG2 is the antigen seen by only one. The T-cell antigens are the gliadins (UNIPROT:D2T2K3, UNIPROT:A0A060N479, UNIPROT:Q402I5), restricted dominantly by MRO:0001229 (HLA-DQA1*05:01/DQB1*02:01) and MRO:0001620 (HLA protein complex with DQ2 serotype). The lineage-specific part of the divergence is not a curatorial hypothesis — it is a zero in a public dataset, addressable by identifier.

Divergence is disease-specific, which is itself the argument for a slot

Celiac is not the general case, and a recommendation built only on it would be wrong:

Disease (IEDB, 2026-09-07)	B-cell records	T-cell records	Dominant antigen
celiac disease	824	2,748	gliadin leads both; TGM2 B-only (256 by server-side filter vs 0)
type 1 diabetes mellitus	550	4,407	overlapping, different leaders — GAD65 (`UNIPROT:Q05329`) leads the B side, insulin (`UNIPROT:P01308`) the T side
myasthenia gravis	77	726	convergent — AChR α (`UNIPROT:P02708`) leads both
pemphigus	251	64	convergent — DSG3 (`UNIPROT:P32926`) on both, 81 B / 60 T

What this is not is a clean two-way split into "divergent" and "convergent" diseases. Lineage skew is a property of an individual antigen, not of a disease, and every disease here mixes shared and skewed antigens:

Disease	dominant antigen	most lineage-skewed antigen
celiac	gliadin `D2T2K3`, both sides	TGM2 — 256 B, 0 T
myasthenia gravis	AChR α, both sides (26 B / 488 T)	AChR δ 0 B / 102 T; IFN-α2 22 B / 0 T
type 1 diabetes	differs by lineage — GAD65 (B), insulin (T)	none exclusive; all top antigens on both sides
pemphigus	DSG3, both sides (81 B / 60 T)	DSG1 88 B / 4 T

What makes celiac the striking case is not that it is the only disease with a one-sided antigen — myasthenia gravis has four, and pemphigus's DSG1 is nearly one-sided at 88 versus 4. It is that celiac's exclusion is on a major antigen against a large opposing corpus: TGM2 is the second most-studied B-cell antigen in the disease and draws zero hits across 2,748 T-cell records. In myasthenia gravis the exclusive antigens are minor ones while the dominant antigen is shared.

So the answer a curator needs is per-antigen and per-lineage, and it cannot be inferred from the disease. dismech currently has no field in which to record it either way, so the information is lost precisely where it is most informative. That is a stronger case for the link object in §7 than a uniform-divergence story would have been.

What IEDB does not solve

Diseases are DOID, not MONDO. Any join runs through a DOID→MONDO mapping and inherits its gaps. Matching on disease name is worse: pemphigus vulgaris returns nothing because IEDB files those records under pemphigus, which returns 251 B-cell and 64 T-cell records — so a naive name join silently under-reports rather than failing.
Antigens are UniProt, and nothing in dismech accepts a UniProt accession. GeneDescriptor binds hgnc:, and GeneProductDescriptor — the obvious candidate — cannot take one either: its required GeneProductTerm is reachable_from: NCIT:C26548, and the string UNIPROT appears nowhere in src/dismech/schema/dismech.yaml or conf/oak_config.yaml, so UNIPROT:P21980 fails term validation today. A protein-level antigen is also a different entity from the gene that encodes it, which is the distinction §3a shows the genetic block collapsing. Unlike the HLA gap this is not a configuration line: OLS and OAK do not serve UniProt, so the options are a new prefix plus fetcher, mapping antigens onto NCIT gene-product terms, or a UniProt-typed slot. That decision is out of scope here and is flagged, not taken.
An assay count is not a mechanism. These are counts of published experiments, weighted by what was studied and fundable. They say what has been measured, never what matters — and qualitative_measure includes Negative records, which a naive count silently treats as support.
Coverage is uneven and cannot be assumed. Prevalent, heavily-studied autoimmune diseases are deep; rarer entries in this cohort may have nothing.

IEDB data is public domain and its 2024 update is PMID:39558162.

8. What this audit deliberately does not decide

Whether to change the schema. The numbers say the information is absent from structure; they do not say the cost of adding a slot is worth paying. That is a design-decision call (docs/explanation/design-decisions.md), and the 105 unbound autoantibody strings could equally be addressed by binding them to existing HP terms with no schema change at all.
Which sense of "different antigens" matters most. All four are gaps, but they are not equally expensive. Note that the one that looked mechanical is not: §5 shows all 41 unbound HLA rows name a serotype, allele or haplotype rather than a gene, so there is no gene_term backfill to do. What the HLA work needs is somewhere to put an allele and a peptide-to-allele link — and §7a finds that MRO already supplies both, reachable through the ols: adapter dismech already uses. That makes it a smaller job than this report first judged, but still a slot decision rather than a sweep.
Whether an unbound antigen is a defect. Many are correctly unbound — Anti-Endomysial Antibodies names a tissue staining pattern, not a molecule, and forcing a CURIE onto it would be worse than leaving it. The dismech-terms rule stands: no term beats a bad one.

9. Method and caveats

scripts/immune_antigen_audit.py (just immune-antigen-audit), offline, over kb/disorders/ and kb/modules/. Runtime is hardware-bound: ~70 s on a CI runner, ~3 min 40 s on a throttled container.

The IEDB figures in §7a come from a different method and carry different caveats. They were obtained by ad-hoc HTTP queries against https://query-api.iedb.org on 2026-09-07, not by the committed script, so they do not regenerate with just immune-antigen-audit and are reproducible only by re-issuing the queries. Counts are exact (Prefer: count=exact) rather than sampled, but they are counts of assay records, so a protein studied often outranks a protein that matters; qualitative_measure includes Negative records, which these totals do not exclude.

Per-antigen counts are method-dependent, and the tables above disclose the spread rather than picking a number. Filtering server-side on parent_source_antigen_iri=eq.UNIPROT:P21980 returns 256 celiac B-cell records; paginating the whole 824-row result and counting that same column in the returned rows gives 472. The rows involved carry several curated accessions for one protein — SRC454731, P21980.2 and NP_004604.2 all appear — so the server-side equality filter and the projected column do not agree on what counts as that antigen, and this report does not claim to know which is canonical. Presence and absence are unaffected: the zero on the celiac T-cell side and the 0 versus 297,935 MHC split reproduce identically under both methods, and the whole-table totals (824, 2,748, 550, 4,407, 77, 726) are single-method counts with no such ambiguity. Ranking within a lineage is also unaffected, since one method is used throughout a given comparison — which is how the type 1 diabetes row was corrected: a 500-row sample had suggested GAD65 led the T-cell side, and full enumeration of all 4,407 rows shows insulin leads it, 1,057 to 692.

Every count must carry its disease filter, and one in an earlier draft did not. The pemphigus row originally reported DSG3 as 145 B / 134 T, which are corpus-wide counts with no disease_names filter; scoped to pemphigus the figures are 81 B / 60 T. The unfiltered number was larger than the whole denominator it sat next to — pemphigus has 64 T-cell records in total — which is the tell for this class of mistake, and the reason the query shapes are written out below rather than left implicit.

Disease selection is by IEDB's own disease name, which is DOID-derived and does not match dismech's MONDO labels: pemphigus vulgaris matches nothing while pemphigus matches 251 B-cell and 64 T-cell records, so a name join under-reports silently.

The queries behind §7a take these shapes:

BASE=https://query-api.iedb.org
# the Prefer/Range pair returns an exact total in Content-Range, not rows
COUNT="curl -s -o /dev/null -D - -H 'Prefer: count=exact' -H 'Range: 0-0'"

# whole-table count
eval $COUNT '"$BASE/tcell_search?disease_names=cs.%7Bceliac%20disease%7D"'

# a disease-scoped, antigen-scoped count — BOTH filters, always.
# Omitting disease_names here is what produced the erroneous pemphigus figures.
eval $COUNT '"$BASE/bcell_search?disease_names=cs.%7Bceliac%20disease%7D&parent_source_antigen_iri=eq.UNIPROT:P21980"'

# the B/T MHC asymmetry
eval $COUNT '"$BASE/tcell_search?host_organism_iri=eq.NCBITaxon:9606&mhc_allele_iri=not.is.null"'

Only three classes can answer the lineage question, and the headline is reported over those. cell_types is a slot on Pathophysiology, Biochemical and ExperimentalModel only (FunctionalEffect has affected_cell_types). An evidence item or a treatment cannot record a lineage, so counting its silence as a gap inflates the rate: 2,694 of the 3,122 antigen-naming objects (86.3%) are of such classes. The 64% figure is taken over the 428 that could carry the slot; the 95% figure over all 3,122 is reported alongside it and is a denominator artifact. Eligibility is decided by path, because the walker sees raw mappings with no class information: the path's last segment must be a member of a pathophysiology / biochemical / experimental_models list. That excludes objects nested inside an eligible one — pathophysiology[0].evidence[2] is an EvidenceItem — and treating the subtree as eligible would add 907 such objects (495 evidence items, 255 downstream links, 105 biological_processes descriptors) and report 88% instead of 64%. The test does follow nesting in the other direction, so stages[0].pathophysiology[1] counts; no antigen-naming object currently sits there, so it does not move the 428.
"No lineage" is not the same as "no cell types". Of the 2,967 objects with no B/T/APC lineage, 2,892 have an empty or absent cell-type slot and 75 have one that names a non-lymphoid cell. The script reports these separately.
The cohort is a keyword union, not an ontology closure. No MONDO descendant query was run (that needs the MONDO build). 263 of 573 entries are in the cohort because their prose says "autoimmune". False positives are certain — an entry mentioning autoimmunity in a differential diagnosis is counted.
The antigen patterns are deliberately generous, because the point is to find every place an antigen is named, and none of those places is a slot. §2a shows the noise this admits (anti-Müllerian hormone, anti-DNase B). Raw pattern counts are upper bounds; the structural counts (cell_types present or absent, biomarker_term bound or not) are exact.
Lineage is matched on curated CL labels by substring, not by CL closure, so a node bound to an unusual CL term whose label omits "B cell" / "T cell" is counted as unattributed. This biases both unattributed figures upward slightly; the correction is bounded by the 20 uses of the ambiguous CL:0000542 lymphocyte.
The HLA row classification is a name test, not a curator's judgement. A row counts as bindable only when its name is exactly an HGNC HLA gene symbol; serotypes, alleles, haplotypes and region descriptions are counted as not bindable. "Explained" means the row's notes say why it is unbound, matched on wording — a row could be deliberate without using those words. One row (Juvenile_Idiopathic_Arthritis — "Non-HLA immune-susceptibility genes…") matches the HLA substring while being about the opposite, and is counted.
Attribution is scored on the antigen-naming object itself, not on its parent or siblings. A node whose description names an antigen while a sibling node carries the B cell counts as unattributed — correctly, since no relation between them is recorded, but a human reader would often infer one.