Glycobiology Project

Bottom line: glycosylation is the most common and diverse protein
modification, built by large, sequence-similar families of glycosyltransferases
and read by lectins, and GO annotation of these glycogenes is easy to get wrong
at the level of specificity. We reviewed every existing GO annotation on 11
human glycogenes (seven exemplars spanning transferase, lectin and CDG axes,
then four mucin-type O-glycan initiation genes), indexed 17 glycobiology
pathway modules whose 100 cited gene reviews add 2,748 more annotations, and
built a CAZy-family-to-GO mapping (cazy2go) as the glyco analogue of
interpro2go. Across the 376 single-gene exemplar annotations only 4 were
removed, while 107 were kept as non-core, 40 marked over-annotated, 38 modified,
and 2 proposed as new: the problem is altitude and pleiotropy (generic parent MF
or membrane terms, downstream physiology on single-sugar enzymes), not wrong
functions, and the module cohort shows the same skew. The cazy2go work
produced a 60-row safe-to-propagate set and 34 hand-endorsed family-level gaps
in interpro2go, still awaiting curator sign-off. The GOA closure query that
would size the full animal glycogene set is tracked with the remaining
follow-ups in #4094.

Overview

Glycobiology studies the structure, biosynthesis, degradation, and biological
roles of glycans (sugars) and glycoconjugates (glycoproteins, glycolipids,
proteoglycans, GPI-anchored proteins). Glycosylation is the most common and
structurally diverse co-/post-translational modification: glycans decorate the
majority of the cell-surface and secreted proteome, mediate cell–cell and
host–pathogen recognition, and tune protein folding, trafficking, stability, and
signalling. Genetic defects in the glycosylation machinery cause the
congenital disorders of glycosylation (CDGs) —
a heterogeneous family of >130 Mendelian diseases — and aberrant glycosylation is
a hallmark of cancer, inflammation, and autoimmune disease.

This project has two complementary aims:

1. GO-usage audit (animals). Survey how the glycobiology-related GO term
   landscape — glycosyltransferase/glycosidase activities, carbohydrate binding
   (lectins), and the N-/O-linked/GPI glycosylation biological processes — is
   actually used to annotate animal gene products (human, mouse, rat, worm), and
   identify systematic over-/under-/mis-annotation patterns the way the
   SPKW, RHEA, and INTERPRO source-audit
   projects do for their pipelines.
2. External-resource landscape. Map the dedicated glycoscience databases,
   ontologies, and tools (the GlySpace Alliance — GlyGen, GlyCosmos,
   Glyco@Expasy/GlyConnect — plus GlyTouCan, UniCarbKB, CAZy, and the
   GlycoConjugate Ontology GlycoCoO) and assess how each could feed, cross-check,
   or extend GO-based curation of glycogenes. Reference dossier:
   GLYCOBIOLOGY-resources.md.

The two aims meet at one question: where does GO under-represent glycan biology
that the specialist glyco resources capture well, and where does GO
over-annotate glycogenes (e.g. propagating downstream "guilt-by-substrate"
metabolic processes onto an enzyme that only adds one sugar)?

Exemplar reviews (calibration set)

Seven human genes were reviewed to calibrate the over-/under-annotation
hypotheses across the main functional axes of animal glycobiology. Phase 1
covered one glycosyltransferase, one lectin, and one CDG gene;
Phase 2 broadened the GT axis across four more transferase sub-types
(O-fucosyl-, GlcNAc-branching-, sialyl-, galactosyl-). Each Phase-2 gene was
reviewed with a FutureHouse Falcon deep-research report integrated from the
start, and the Phase-1 three were retro-fitted with their Falcon findings.

All seven validate clean (status: DRAFT).

Verdict distributions (authoritative, from the YAMLs)

Only 1/303 REMOVE (ST6GAL1, a high-throughput-interactome protein binding),
against 102 NON_CORE + 34 OVER + 24 MODIFY — i.e. ~53% of annotations are
demoted, refined, or still unresolved but 99.7% are retained in some form. The
mis-annotation signal is overwhelmingly altitude/specificity and pleiotropy,
not wrong functions — exactly the project's prediction. (PMM2's three former
REMOVEs have since been softened to MARK_AS_OVER_ANNOTATED; the counts above
are recomputed from the YAMLs rather than carried forward.)

What the exemplars confirmed (audit hypotheses → evidence)

• Generic-MF-term collapse to the specific activity is the dominant GT fix.
  Every transferase carried over-general parent MF terms that were MODIFY'd to
  the specific child: B3GALNT2 (UDP-glycosyltransferase/hexosyltransferase →
  GO:0008376), MGAT1 (acetylglucosaminyltransferase → GO:0003827 GlcNAc-TI),
  ST6GAL1 (sialyltransferase → GO:0003835 α-2,6-sialyl-T), B4GALT1
  (galactosyltransferase/UDP-galactosyltransferase → GO:0003831).
• Substrate "guilt-by-association" over-annotation is real and gene-class-dependent.
  GTs and the precursor enzyme (PMM2) carry downstream whole-pathway terms beyond
  their molecular role. B3GALNT2's broad process terms were MODIFY'd to the precise
  GO:0035269 (O-mannosyl). PMM2 — which only supplies GDP-mannose and is not a
  glycosyltransferase — had protein N-linked glycosylation kept NON_CORE.
  MGAT1's many complex-N-glycan-dependent physiologies are likewise NON_CORE.
• Compartment errors. POFUT1 is a soluble ER-luminal enzyme; its membrane
  calls (TM-prediction/proteomics) were over-annotated and ER localization affirmed
  — a reminder that GT localization annotations need the right organelle.
• Lectin "protein binding" under-specificity is pervasive. Galectin-3 carried
  21 generic protein binding (GO:0005515) IPI annotations — all downgraded;
  many are glycan-mediated lectin contacts whose informative MF is carbohydrate binding (GO:0030246).
• Pleiotropy ≠ core. The dominant LGALS3 verdict was NON_CORE (62/106); its
  immune/apoptosis/fibrosis/cancer biology is downstream of one core activity
  (β-galactoside CRD binding + N-terminal LLPS lattice). A new GO:0062093 lysophagy
  annotation was added (the damaged-endomembrane glycan-sensing role).
• GO coverage gaps surfaced for curation — 8 proposed_new_terms across the set:
  protein-O-mannose β-1,3-GalNAc-T activity (B3GALNT2, the RHEA-flagged
  EC 2.4.1.313 gap); galectin-glycan-lattice-assembly + damaged-endomembrane-glycan-
  sensor (LGALS3); phosphomannomutase-in-GDP-mannose-biosynthesis (PMM2);
  O-fucosylation-coupled ER quality control (POFUT1); oligomannose→complex N-glycan
  conversion (MGAT1); protein-N-glycan α-2,6-sialyltransferase activity (ST6GAL1);
  and a lactose synthase complex CC term (B4GALT1).
• Citation hygiene (verify, don't trust). Reviews flagged real citation problems
  via reference_review: PMM2's FlyBase IMP MISCITED (PMID:9525984 is MPI/CDG-Ib,
  a different enzyme); POFUT1's MF TAS to a Drosophila survey marked MISCITED; and
  across all Falcon integrations, preprint-only / unresolvable-PMID claims and a
  stale EC number (ST6GAL1 "2.4.99.1" → current 2.4.3.1) were rejected and recorded
  in notes rather than cited.

Net: across all 303 Phase 1/2 exemplar annotations, the verdict skew (a single
REMOVE, a high-throughput-interactome protein binding; heavy NON_CORE + OVER
+ MODIFY) matches the project's prediction — glycogene mis-annotation is
dominated by altitude / specificity and pleiotropy, not outright wrong
functions. The independent 2,748-annotation module cohort
(pathway modules) reproduces this at
scale.

Phase 3 — mucin-type O-glycan initiation

The O-glycosylation gap identified when the pathway modules were indexed is now
partly closed. Four human genes covering the initiation and first two elongation
steps of mucin-type O-glycan biosynthesis were reviewed de novo (73 annotations,
all validate clean):

Verdicts: 44 ACCEPT / 14 MODIFY / 5 NON_CORE / 6 OVER / 3 REMOVE / 1 NEW. Every
supporting_text quote was verified as a verbatim substring of the cached
publication (164 quotes, 0 failures) and every proposed GO term checked against
QuickGO for existence, aspect and obsolescence.

Note on provenance. Deep-research tooling was unavailable in this environment
(the configured OpenAI key is rejected), so these four carry no
-deep-research-PROVIDER.md file, unlike the Phase 1 and 2 exemplars. The
reviews are grounded instead in the cached publications, which were present for
every PMID cited by GOA. Per repository policy no self-authored file was named as
though it were provider deep research.

What Phase 3 adds to the audit

• The chaperone/enzyme boundary is already curated correctly, and that is the
  finding. Cosmc is a chaperone with no transferase activity, and GOA carries
  GO:0016263 on it as a NOT annotation — the curators had explicitly negated
  the client enzyme's activity on its chaperone. This was the batch's most
  instructive row: read without the negation flag it looks like a textbook
  necessity-mistaken-for-possession error, and it is the opposite. The seeding
  pipeline carries negated: true correctly; the lesson is for the reviewer, who
  must read the flag before reasoning about the row, since the negation inverts
  the meaning of an otherwise ordinary-looking IMP.
• Generic-MF collapse recurs, and extends to cellular components. Both
  GO:0016757 glycosyltransferase activity rows were MODIFY'd to the specific
  child, matching Phases 1-2. But the dominant MODIFY here was locational: six
  GO:0016020 membrane rows across three genes collapsed to GO:0000139 Golgi membrane. Bare membrane on a type II Golgi glycosyltransferase is the
  cellular-component analogue of the generic-MF problem.
• Guilt-by-substrate is strongest where the product has famous physiology.
  GCNT1 carries leukocyte tethering or rolling, its positive regulation, and cell
  adhesion molecule production — all because it builds the core 2 branch of a
  selectin ligand. The tethering assays behind them were run on GOLPH3-depleted
  cells, not GCNT1-depleted cells. GALNT1 carries GO:0019082 viral protein processing because Reactome models it glycosylating SARS-CoV-2 ORF3a; the
  enzyme's activity does not change with the provenance of its acceptor.
• A committed pathway enzyme was missing from its own pathway. C1GALT1 had
  no mucin-type O-glycosylation process annotation: its only two
  biological-process rows were angiogenesis and kidney development, both cited to
  a case-control SNP association study for IgA nephropathy. The comparator check
  confirms a gap rather than a convention — GALNT1 (upstream), GCNT1 (downstream)
  and even C1GALT1's own non-catalytic chaperone all carry the process term. Added
  as the batch's single NEW.
• A citation that format validation cannot catch. PMID:17228361 is correctly
  identified and correctly titled, and supports neither annotation drawn from it;
  it is recorded MISCITED in reference_review.
• Restraint where the cache is thin. PMID:16638743 is a GalNAc-T3/FGF23 paper
  cited as an IDA on GALNT1, abstract-only in our cache. Rather than allege
  mis-attribution, it is marked UNVERIFIED and the annotation ACCEPTed: a
  selectivity claim of that kind is normally established against a panel of
  isoforms, and the asserted activity is independently certain for this gene.

Still uncovered on this axis: the remaining GALNT paralogues (the family's
acceptor-specificity differences are what GO does not currently express), core 3
and core 4 branching, and the sialyl/fucosyl capping steps.

Pathway modules

Alongside the single-gene exemplars, the repo holds 17 curated glycobiology
ModuleReview documents — pathway-level decompositions of N-glycan LLO
assembly and transfer, dolichol/nucleotide-sugar donor supply, the five GPI-anchor
stages, glycosphingolipid biosynthesis, and the lysosomal catabolism of
glycolipids, heparan sulfate and keratan/chondroitin sulfate. Index, with the
cohort analysis:
GLYCOBIOLOGY-modules.md.

These were curated independently as pathway-level documents, so the corpus
probes a wider glycobiology space than the exemplar set alone implies. The 17
modules cite 100 distinct human gene reviews (all present in
genes/human/, zero overlap with the 11 single-gene exemplars) carrying 2,748
adjudicated annotations. That cohort is a second, independent verdict baseline:

It corroborates the project's central claim on 10× the data — REMOVE stays
under 1%, retention ~99.2% — while showing the exemplar set is deliberately
NON_CORE-skewed (62 of its 102 NON_CORE verdicts are LGALS3 alone). The gap
between the two rows is a gene-class effect, narrow biosynthetic enzymes
versus a pleiotropic lectin, and should be read that way when the GOA closure
query yields a third number.

Two gaps the module axis exposes:

• O-glycosylation is uncovered. No module addresses mucin-type,
  GALNT-initiated O-glycan biosynthesis; the exemplars reach O-linked biology
  only via POFUT1 and B3GALNT2.
• The module cohort under-proposes new terms — 2 proposed_new_terms entries
  across 100 module genes, against 8 across the 7 Phase 1/2 exemplars. Almost certainly a
  curation-process artifact rather than a real coverage difference, and a
  concrete re-review target.

The glycobiology GO term landscape

Key entry points into the ontology (GO IDs/labels verified against QuickGO,
GO release current as of 2026-06). Closure (descendant) sets under these terms
define the working scope of the audit.

Molecular function (the enzymes and binders)

Biological process (where the sugars go)

These anchor terms give a reproducible way to pull the animal glycogene set from
GOA for the usage audit (filter by aspect + closure under these IDs, restricted
to the animal taxa we curate).

External glycoscience resources (landscape)

Full dossier with URLs, identifier schemes, licences, and programmatic-access
notes: GLYCOBIOLOGY-resources.md.
Headline resources:

The GlySpace Alliance (GlyGen + Glyco@Expasy + GlyCosmos) is the coordinating
umbrella; GlyTouCan is the shared structure-ID backbone they all link to.

Reusing these resources for curation

How the resources feed GO curation (forward cazy2go propagation, GlycoCoO→GO
alignment, and confirmatory GlyGen cross-checks; GO-CAM/pathway deferred) is a
decision record in
GLYCOBIOLOGY-resource-reuse.md,
with a seeded cazy2go.sssom.yaml (CAZy
family → GO molecular function, the glyco analogue of interpro2go) built from
the exemplar GT families.

Candidate animal genes already in the repo

The corpus already contains glyco-relevant gene reviews that can seed the audit
without new fetches — useful for calibrating annotation patterns before scaling:

• GPC6 (human glypican) — heparan-sulfate proteoglycan
• Notch1 / NOTCH1 (mouse, human) — O-fucosylation substrate; entry point to POFUT/POGLUT biology
• Uggt1 (rat) — UDP-glucose:glycoprotein glucosyltransferase (ER glycoprotein folding QC)
• Aga (rat) — aspartylglucosaminidase (N-glycan catabolism; aspartylglucosaminuria)
• Ugt2a1 / Ugt85A2 (rat, plant) — UDP-glucuronosyl/glucosyltransferases (small-molecule GTs — note paralog/specificity pitfalls)
• clec-60 (worm), lys-7 (worm) — C-type lectin / carbohydrate-recognition domain proteins
• ROT1, CNE1, CRH1, PDI1 (yeast) — ER glycoprotein folding/QC context (out-of-animal-scope but methodologically adjacent)

A reproducible GOA closure query (per the term table above) will produce the
fuller candidate list; these are the already-curated anchors.

Curation considerations (hypotheses to test in the audit)

• Substrate "guilt-by-association" over-annotation. A glycosyltransferase
  that adds one sugar may be annotated to the whole downstream pathway/process of
  the mature glycan (cf. the RHEA/EC altitude problem). Expect MARK_AS_OVER_ANNOTATED
  / MODIFY candidates.
• Paralog/specificity errors in large GT families. UGT, GALNT, ST3GAL/ST6GAL,
  B3GALNT, fucosyltransferase families are big and sequence-similar; IEA/IBA
  propagation can land the wrong substrate specificity on the wrong paralog
  (cf. the de Crécy-Lagard PLI/PARALOG_OVERANNOTATION pattern).
• Lectin "protein binding" under-specificity. Carbohydrate-binding proteins
  are often annotated only with generic protein binding; the informative
  carbohydrate binding (GO:0030246) child is the better MF (per repo guidance
  to avoid protein binding).
• GO vs specialist-resource coverage gaps. Site-level and structure-level
  glycosylation evidence in GlyConnect/UniCarbKB may have no GO counterpart →
  proposed_new_terms / gap-filling candidates, mirroring the RHEA reverse-gap
  analysis.

Open questions

• How large is the animal glycogene set under the anchor terms, and what is the
  ACCEPT/MODIFY/OVER/REMOVE verdict distribution relative to the corpus baseline?
  (Partly answered: 111 human glycogenes are already reviewed — 11 single-gene
  exemplars plus the 100 genes in
  pathway modules — but the two cohorts
  give materially different distributions, so "the" baseline depends on
  gene-class composition. The closure query needs to report that composition, not
  just a single percentage.)
• Can CAZy GT/GH family membership be used as an independent check on
  glycosyltransferase/glycosidase MF annotations (right family → right activity
  class)?
• Which GlyConnect/UniCarbKB enzyme→glycan→site facts have no GO representation,
  and are they new-term candidates?
• Is GlycoCoO alignable to GO well enough to import/cross-validate annotations?

Status

• Started: 2026-06-21
• Maturity: IN_PROGRESS — scope, GO anchor-term set (QuickGO-verified), and
  external resource landscape defined; 11 human exemplar reviews complete
  (the 7-gene Phase 1/2 set plus 4 mucin-type O-glycan initiation genes; 376
  annotations), and all validate clean. The first seven are backed by a
  FutureHouse Falcon deep-research report integrated into the YAML, while
  Phase 3 was curated directly from cached GOA publications. Together they
  confirm the altitude/specificity + pleiotropy over-annotation hypotheses (184
  ACCEPT / 107 NON_CORE / 40 OVER / 38 MODIFY / 4 REMOVE / 2 NEW / 1 UNDECIDED).
  17 pathway modules and their 100-gene / 2,748-annotation
  cohort are now indexed under
  GLYCOBIOLOGY-modules.md and reproduce
  the same skew independently.
• Next steps: (1) run the GOA closure query for the anchor terms across the
  animal taxa to enumerate the glycogene set and its verdict baseline — the
  111-gene reviewed human corpus (11 exemplars + 100 module genes) is the
  denominator to measure it against; (2) promote the 8 exemplar
  proposed_new_terms as GO new-term requests; (3) get curator sign-off on the
  34 hand-endorsed cazy2go true gaps and 210 family-specific subfamily
  proposals; (4) build the GlycoCoO/GlycoEnzOnto→GO SSSOM alignment and GlyGen
  cross-check probe; (5) add a mucin-type GALNT-initiated O-glycan module; (6)
  re-review the module cohort for proposed_new_terms, which it currently
  under-proposes (2 across 100 module genes); (7) extend Phase 3 across the
  remaining GALNT paralogues and the core 3/core 4 and capping steps. See
  #4094.