Allergens Project

Bottom line: allergens are being knocked out (CRISPR Fel d 1 cats), neutralized with
antibodies and used in immunotherapy, yet for many of them nobody knows what the protein does
in its source organism. We use the allergen label only as a prioritization bucket: being an
allergen is an IgE property of a sensitized human, not an evolved function, so the reviews
stay on native function. We built an allergen-to-UniProt index (32 genes, 31 molecules), a
registry snapshot and worklist from UniProt's Allergen keyword (1,020 reviewed entries), and
an IEDB epitope ETL, then ranked targets by epitope load times function uncertainty. So far 24
genes are reviewed (the full cat set, dog, horse, cow, mouse, rat, house dust mite and birch
allergens, plus the uteroglobin comparator Scgb1a1): 219 existing annotations, none removed, 62 marked
over-annotated (40 of them, mostly ISS metabolic and insulin-signalling terms, on the mouse and rat MUPs) plus 18 new terms
proposed. The headline is that the top-priority allergens (Bet v 1, Fel d 1, Can f 1) are
heavily IgE-targeted and still have no known native function. Coverage of the registry is
small and the reviews are still marked DRAFT; the status list at the end has the most current
counts, and the "16 genes" figures in the middle sections predate the later cohorts.

Why an allergens cohort

This is not a new kind of annotation. It is a triage layer that selects genes for
ordinary function review. The motivation is concrete and clinical:

We increasingly intervene on allergens — so we should know what they do first.
Major allergens are being knocked out, neutralized, or engineered:

• CRISPR knockout of the Fel d 1 genes (CH1/CH2) to make hypoallergenic cats
  [PMID:35386981].
• Antibody neutralization (anti–Fel d 1 in cat food; therapeutic IgG4 in
  patients).
• Allergen-specific immunotherapy with recombinant/peptide allergens.

Each of these abolishes or suppresses a protein. The obvious safety question —
"what physiological function do we lose?" — is exactly a gene-function-review
question. For the cat allergen Fel d 1 the honest answer is largely unknown, and
that is the single most decision-relevant finding, captured in each review's
knowledge_gaps.

Scope and unit of analysis

The reviews are per gene / UniProt entry, but allergen databases are organized
per allergen molecule (e.g. WHO/IUIS Fel d 1, which spans two genes,
FELCA/CH1 + FELCA/CH2). The project therefore needs an index/mapping layer
that bridges that one-to-many relationship — the allergen molecule is the cohort
member; the genes are the review units.

Prioritization metric

Rank candidates within the bucket on intervention pressure × function
uncertainty:

• Intervention pressure — is the protein being drugged, knocked out, or
  engineered? (IgE prevalence, epitope load, "therapeutic target" flag.)
• Function uncertainty — how confidently is the evolved function known? This is
  the axis already modelled in the Function Knowledge Gaps
  project.

Allergens score high on both, which is why they are a natural first cohort. The
per-gene deliverable is "here is the evolved function, or here is the
sharply-bounded gap" — not "is it an allergen" (WHO/IUIS already says so).

External databases

These are triage inputs and cross-references only — they never change a GO
annotation. Build ETL only for sources with genuine APIs; for the rest, parse
downloads (do not write code that just points at a website).

Anchors: WHO/IUIS (defines the cohort) + IEDB (rich, API-accessible epitope
data). Layer SDAP / AllergenOnline later for cross-reactivity. Treat Allergome as a
cross-reference resolver since UniProt already links it.

These fit the repo's existing patterns: cache-per-record (like publications/,
reactome/), an index TSV (like gocams/index.tsv) keying allergen → UniProt →
review, and SSSOM mapping sets (like the ARO→GO / RHEA→GO projects) for
allergen→UniProt links. Allergen-specific metadata (route of exposure, IgE
prevalence, epitopes) would live in a schema side-block, not in GO — note that
UniProt keyword KW-0020 Allergen exists but there is no clean "allergen activity"
GO molecular function, which is precisely why the side-channel is needed.

Case study: the secretoglobin allergens

The first genes curated under this lens are all secretoglobins, which makes a
sharp point: the family shares a "PLA2 modulation + hydrophobic-ligand binding +
immunomodulation" theme, yet no family member has a confirmed endogenous
function — including the best-studied one.

The contrast is instructive for the "know before you knock out" thesis: even
uteroglobin, the archetype with decades of study, is annotated mainly through
PLA2 inhibition and Th2 suppression rather than a settled endogenous purpose;
the uteroglobin knockout mouse shows inflammation/cancer susceptibility, so
"harmless to remove" is not a safe default for Fel d 1 either.

The cat Fel d 1 reviews additionally drew on FutureHouse Falcon deep research,
which surfaced the experimentally-grounded LPS-binding / TLR4-enhancement activity
(Herre et al. 2013, [PMID:23878318]) and ligand-binding data
([PMID:34026578]); all such claims were verified against the cited primary
literature before annotation.

Allergen → UniProt index

The cohort membership and the molecule→gene bridge are maintained as a generated
TSV, ALLERGENS/allergen_index.tsv, built by
ALLERGENS/build_allergen_index.py:

uv run python projects/ALLERGENS/build_allergen_index.py \
    -o projects/ALLERGENS/allergen_index.tsv

The builder is deliberately download-honest: it derives membership from the
already-cached UniProt records (the Allergen= name, the Allergen keyword, and
Allergome cross-references) and joins them to the local reviews. It does not
call or fake a WHO/IUIS API; to fold in the official registry, drop its downloaded
table into the folder and extend the merge step. Re-running picks up every allergen
gene present under genes/, so the index grows automatically as the cohort expands.

Columns: allergen_molecule (the WHO/IUIS unit), allergome_id, source_taxon_id,
species_code, gene_symbol, uniprot, uniprot_allergen_name, review_path,
review_status, n_core_functions, n_knowledge_gaps, function_gap_flagged,
iedb_epitopes, iedb_has_ige (the last two merged from the IEDB ETL below).

Membership is detected from the cached UniProt records either by the reviewed
Allergen keyword/Allergen= name or by an Allergome cross-reference (so
unreviewed TrEMBL allergens such as Fel d 7 and Fel d 8 are included). The index
currently holds 32 genes across 31 allergen molecules — the cat, dog, horse,
cow, mouse and rat danders, the headline mite (Der p 1/2/23) and birch (Bet v 1/2)
allergens, and allergens already in the repo for other reasons (e.g. human GBA1,
INS, GLA; yeast SOD2).

Cross-cohort priority (function gap × IEDB load)

With both axes populated, the index ranks the whole curated cohort. The highest-value
review targets are allergens that are both heavily IgE-targeted and of
uncertain evolved function:

The deprioritized rows (Der p 1, Der p 2) carry the largest epitope loads yet have
well-defined functions — exactly the "high data, low uncertainty" quadrant the metric
is meant to filter out. Conversely Bet v 1 rises to the top: the single
most-IgE-targeted allergen in the set whose physiological function is still unresolved.

The index now carries both axes of the prioritization metric: function_gap_flagged
(uncertainty) and the IEDB epitope counts (iedb_epitopes, iedb_has_ige —
intervention pressure; see below). The complete domestic-cat set, ranked by the two
axes together:

(Fel d 5/6 are cat immunoglobulins, out of scope.) The ranking falls out cleanly:
Fel d 1 tops it — heavily IgE-targeted (127 epitopes) and of unknown native
function — the textbook "know before you knock out" case. The two least-characterized
members (Fel d 7, Fel d 8, both unreviewed TrEMBL) also carry gaps, while the
well-understood Fel d 2/3/4 families are deprioritized despite real epitope load.
mouse/Scgb1a1 is intentionally absent — it is the secretoglobin comparator, not a
registered allergen, so it does not appear in the membership-derived index.

Registry coverage and fetch worklist

WHO/IUIS publishes no stable API, but UniProt's Allergen keyword (KW-0020) is
a curated, API-accessible proxy: each reviewed allergen entry carries its WHO/IUIS
designation inline in its protein names as (allergen <name>)
(e.g. (allergen Fel d 1-A)). ALLERGENS/fetch_uniprot_allergens.py
snapshots that registry and cross-references it against genes/ to produce a
prioritizable backlog:

uv run python projects/ALLERGENS/fetch_uniprot_allergens.py

Outputs (UniProt release 2026_02):

• ALLERGENS/uniprot_allergens.tsv — the registry
  snapshot: 1020 reviewed allergen entries spanning 624 allergen molecules,
  with accession, source organism/taxon, gene, WHO/IUIS name, molecule and Allergome id.
• ALLERGENS/allergen_worklist.tsv — the 1014
  registry members not yet fetched, each with a ready-to-run fetch-gene command.

Coverage of this reviewed registry: 6 / 1020 entries (Fel d 1 chains
P30438/P30440; cat Fel d 2/3/4; human thioredoxin). Note this differs from the
local index count (16) above: the registry/worklist tracks only reviewed
UniProt entries, whereas the local index also counts unreviewed TrEMBL allergens
(Fel d 7, Fel d 8) and Allergome-listed entries that lack the UniProt Allergen
keyword. This calls a real API (UniProt REST) and records the release for
provenance — it does not fabricate or fake-fetch a WHO/IUIS table.

The worklist is currently ordered by organism then allergen name; true
intervention-pressure ranking (IgE prevalence, epitope load) awaits the IEDB
epitope step. It is the backlog from which the cohort is grown by running the listed
fetch-gene commands and then reviewing each gene.

IEDB epitopes (intervention-pressure axis)

ALLERGENS/fetch_iedb_epitopes.py populates the
second axis of the metric from the IEDB IQ-API (query-api.iedb.org, a real
PostgREST API), writing ALLERGENS/iedb_epitopes.tsv
and merging epitope counts into the main index:

uv run python projects/ALLERGENS/fetch_iedb_epitopes.py
uv run python projects/ALLERGENS/build_allergen_index.py   # re-merge into the index

Per allergen molecule it records distinct epitope, B-cell-assay, T-cell-assay and
reference counts and an IgE flag (has_ige — the most allergy-relevant signal).

Join caveat (handled honestly): IEDB keys allergens under its own UniProt
accessions (Fel d 1 = UNIPROT:A0ABI7XLA3), which differ from the Swiss-Prot
accessions used here (P30438). IEDB does label them by WHO/IUIS allergen name,
so the ETL joins by allergen-molecule name within source taxon rather than by
accession. This works for WHO/IUIS-style names (Fel d 1); it does not match
allergens that IEDB labels by ordinary protein name (e.g. human self-allergens
catalogued here as Hom s …), which therefore show 0 — meaning not matched, not
necessarily no epitopes. The cat cohort matches cleanly:

The numbers track clinical reality (Fel d 1 dominates) and complete the metric:
crossing IEDB epitope load with the function-gap flag yields the priority column in
the cat table above.

Worklist in action — the dog cohort (Can f 1, 2, 3, 6). Working the worklist by
intervention pressure (rather than alphabetically) picked the dog allergens next.
The same two-axis ranking applies, and Can f 1 is the standout:

Can f 1, the dominant dog allergen, mirrors Fel d 1 exactly: heavily IgE-targeted yet
of unknown specific ligand/function — the highest-value review target. (*Can f 3 = 0
is the same name-join limitation: IEDB labels dog albumin epitopes under other names.)

Two robustness fixes were needed to cover dog: the ETL now paginates via the
PostgREST Range header (the API rejects offset; some taxa exceed 1000 antigens),
and matches the allergen designation by regex so embedded IEDB names
("Major allergen Can f 1", "Lipocalin Can f 6.0101") reduce correctly to the molecule.

Status

• SCOPING. Architecture and first secretoglobin cohort drafted.
• Curated: FELCA/CH1, FELCA/CH2, mouse/Scgb1a1.
• Done: allergen→UniProt index (molecule↔gene bridge, now 16 genes / 15 molecules)
  and a UniProt-KW-0020 registry snapshot + fetch worklist (6/1020 reviewed-registry
  covered). The full domestic-cat allergen set (Fel d 1, 2, 3, 4, 7, 8) is curated.
• Done: IEDB epitope ETL — both axes of the prioritization metric are now live
  (function-gap flag × IEDB epitope/IgE load), realized in the cat priority ranking.
• Done: worked the worklist by priority — dog cohort (Can f 1/2/3/6), mammalian
  inhalants (horse Equ c 1/2/3/4, cow Bos d 2, mouse Mus m 1, rat Rat n 1), and the
  headline environmental allergens (mite Der p 1/2/23, birch Bet v 1/2). 32 genes /
  31 molecules now in the index with full IEDB load.
• Notable findings: Der p 2 is an MD-2-mimic TLR4 auto-adjuvant (direct Fel d 1
  parallel); MUP (Mus m 1 / Rat n 1) metabolic annotations are ISS over-propagation;
  Bet v 1 ABA-receptor annotations are fold-based over-propagation; Der p 23 chitin
  binding is correctly a negated GOA annotation.
• Next: extend the IEDB name-join to protein-name-labelled allergens (human Hom s …),
  then continue the backlog (other pollens, foods, molds, insects) by priority.

Slides

• Slides (Marp source: ALLERGENS-slides.md) — AI generated