Reusing glycoscience resources for GO curation — strategy

Decision record for the Glycobiology project: how to reuse the external
glycoscience resources catalogued in GLYCOBIOLOGY-resources.md as
inputs to GO curation, rather than just describing them.

Framing

Most glycoscience resources reach GO only indirectly today (via UniProt / EC / InterPro /
Reactome). There is no official cazy2go or glyconnect2go. So "reuse via mappings" means the
mapping is the deliverable — but the repo already has the exact machinery for this: curated
SSSOM mapping sets validated with linkml-term-validator, exactly as for
RHEA (rhea2go), InterPro (interpro2go), and NCBIFam.

Modes of reuse — decisions

cazy2go: the forward-propagation source

A CAZy family → GO molecular-function mapping, the glyco analogue of interpro2go. Seeded in
cazy2go.sssom.yaml from the project's five exemplar GT families, each row
backed by a completed gene review:

Predicate semantics (mirrors interpro2go.sssom.yaml): exactMatch = holds for ~all (animal)
family members → ready to propagate; narrowMatch = the GO term describes a subfamily only, so
the family is too coarse and subfamily/EC resolution is required first.

The trivial redundancy filter

For each row, the family's EC(s) usually already reach the same GO term via ec2go
(GO_REF:0000003) or an InterPro entry reaches it via interpro2go (GO_REF:0000002). Those rows
are redundant (no marginal annotation) and are dropped by a closure filter — the same
EC-masking computation RHEA already runs (ec2go is fetchable from
current.geneontology.org/ontology/external2go/ec2go). The genuine cazy2go contribution is the
non-redundant remainder: families with no EC, no ec2go row, or no InterPro coverage. Redundancy
status is noted per row in the seed; computing it live across all families is the scale-up step.

Beyond the seed — the generated full mapping

The 5-family hand-curated seed is now extended to all CAZy enzyme families derivable from public
data, reproducibly, by build_cazy2go.py →
cazy2go.generated.sssom.yaml. No row is hand-typed: each is the
live join

CAZy family --(reviewed Swiss-Prot CAZy↔EC xref, UniProt REST)--> EC --(ec2go)--> GO MF

Current build (2026-06): 3,183 reviewed UniProtKB CAZy+EC entries → 302 families → 702 mapping
rows over 283 families (19 families carry an EC with no ec2go term — gap candidates). The same
exact/narrow logic as the seed is applied automatically:

The trivial redundancy filter is built in: 80 generic-parent rows (glycosyltransferase activity, hexosyltransferase activity, hydrolase acting on glycosyl bonds, …) were dropped where
the family also resolved to a specific child — so the output already prefers the specific term. The
generated rows are cross-consistent with the hand-curated seed (e.g. GT13→GO:0003827,
GT65→GO:0046922 both resolve to the same exactMatch the seed asserts).

Caveat (the redundancy property): because every row is derived through an EC that already
reaches GO via ec2go, the rows are EC-reachable by construction. Their value is the same as
interpro2go's — annotating a protein from CAZy family membership when it lacks an EC/ec2go
annotation — but the marginal-vs-ec2go contribution still needs closure-filtering across organisms
(the RHEA EC-masking method). Finer subfamily resolution (for the narrowMatch
families) is the next refinement, seedable from the dbCAN-sub subfamily → EC table
(dbCAN3).

Did it yield anything new, or just confirm?

At the GO-term level it is confirmatory by construction — derived through ec2go, it cannot mint
a term ec2go lacks. But the join surfaced genuinely new, actionable signal:

• 19 GO-coverage-gap families — curated EC(s) but no ec2go term at all. Notably a whole class:
  the auxiliary-activity / LPMO redox CAZymes (AA9, 28 reviewed members; AA10–AA14; CBM104) on
  EC 1.14.99.54/56, plus GH130 (β-mannoside phosphorylases) and GT81. Systematic GO-MF gaps,
  not confirmation. (Some are partial ECs / ec2go lag; AA9 & GH130 carry full ECs — real candidates.)
• Independent rediscovery of the B3GALNT2 gap: EC 2.4.1.313 is absent from ec2go and is
  carried by GT31 — the same gap our hand B3GALNT2 review raised as a proposed_new_term, found
  here with zero manual input. Validates the gap-finding (not just echoing known annotations).
• A coarseness map: GT1 → 33 GO MF terms (214 members), GH13 → 27, GH1 → 25, GT4 → 21 — quantifies
  exactly which families are too heterogeneous to propagate from (the narrowMatch set) and should be
  prioritised for dbCAN-sub subfamily resolution.

cazy2go vs interpro2go (computed) — the marginal-knowledge test, via
compare_cazy_interpro.py (UniProt CAZy+EC+InterPro × ec2go ×
interpro2go). This is the opposite of the RHEA result: where RHEA was ~84% masked by EC2GO,
cazy2go is only 20% masked by interpro2go —

• 283 families have a cazy2go GO MF term; 56 (20%) are fully masked by interpro2go, but
  227 (80%) reach ≥1 GO MF term InterPro does not supply (544 marginal (family,GO) pairs,
  exact-match upper bound).
• Why the inverse of RHEA: interpro2go is deliberately conservative — a signature maps to GO
  only when the term holds for all matching proteins, so for heterogeneous enzyme families it
  withholds the specific catalytic term (or gives only a generic parent). cazy2go (via EC) supplies
  the reaction-level specificity InterPro withholds. The two routes are complementary by
  altitude, not redundant.
• The safe, actionable win = 110 mono-specific (exactMatch) families with a marginal term →
  safe to propagate at family level. Headliners: GT27 (polypeptide-GalNAc-T, mucin-type
  O-glycosylation initiators, 46 members), GT66 (OST glycotransferase, N-glycosylation), GT39
  (dolichyl-P-mannose-protein mannosyltransferase), GT13 (GnT-I), GT21 (ceramide
  glucosyltransferase), GH109 (α-N-acetylgalactosaminidase). (A handful only reach a generic
  hexosyltransferase activity and are weaker.)
• The other 117 marginal families are poly-specific (narrowMatch) — real specificity gain but
  unsafe at family level; route via subfamily/EC (dbCAN-sub).
  Closure-filtered result (closure_cazy_interpro.py, GO is_a
  closure via oaklib / cached sqlite:obo:go). I expected closure to shrink the 544 headline; it
  barely does (544 → 529) — instead it shows what kind of contribution it is. Partitioning the 544
  marginal pairs:

So only 3% of the marginal is closure-masked; 97% is real (529 pairs = altitude + gap). The
TRUE_GAP set spans 98 families and includes biologically central activities InterPro2GO does not
supply for the family at all: GT27/CBM13 → polypeptide-GalNAc-T (mucin O-glyco initiators),
GH19 → lysozyme, GH14 → α-amylase, GH33 → sialidase, PL1 → pectin lyase,
GT43 → β-1,4-xylan synthase, GH32 → fructan fructosyltransferase. (A few are still generic,
e.g. GH72 → hexosyltransferase activity, or noisy multi-activity families like GH23 → peptidase.)

Caveat: is_a-only closure; an ALTITUDE_GAIN whose only InterPro ancestor is near the
molecular_function root is a weak gain. And the partition is per (family, term): a TRUE_GAP in a
poly-specific family (e.g. GH1 90 members → one plant glucosyltransferase) is real coverage but
still unsafe to propagate at family level — route via subfamily.

Bottom line: cazy2go is not mostly redundant with interpro2go and closure does not
explain it away — 97% of the exact-match marginal survives closure as genuine altitude (74%) or
coverage (23%) gain. It adds reaction-level MF specificity for ~110 mono-specific families safely
(family-level) and more via subfamilies; 124 pairs across 98 families are activities InterPro2GO is
silent on entirely.

Status

• Seed: cazy2go.sssom.yaml — 5 exemplar GT-family rows, review-backed, GO ids QuickGO-verified.
• Safe propagation set: cazy2go.safe.sssom.yaml — 60 rows (select_cazy2go_safe.py):
  mono-specific families with a specific GO MF that is a closure-confirmed gain over interpro2go
  (35 true-gap + 25 altitude-gain), skos:exactMatch, safe to propagate at family level. Tightened
  after hand-review (cazy2go-truegap-review.md) to drop systematic
  artifacts: CBM families (non-catalytic binding modules), peptidase MF terms (multidomain
  bleed-through, e.g. GT44 toxin autoprotease), the generic GO:0016837, and the CAUTION families
  (subfamily-split / heterogeneous, e.g. AA5).
• TRUE_GAP hand-review: the 53 mono-specific true-gaps adjudicated →
  34 ENDORSE (actionable interpro2go/GO gaps, e.g. GT27 polypeptide-GalNAc-T, GH109/GH110/GH89
  glycosidases, GT117/GT105 PMT) · 13 CAUTION (need specific EC / subfamily) · 6 REJECT
  (artifacts, now auto-filtered). See cazy2go-truegap-review.md.
• Generated: cazy2go.generated.sssom.yaml — 702 rows / 283 families (build_cazy2go.py),
  full derivation incl. poly-specific narrowMatch families.
• Validation: all three sets pass just validate-cazy-mappings — SSSOM structural validation +
  GO term/label validation (MF-branch-bound) on the regenerated *.terms.yaml, via the new
  cazy_go_mapping.yaml schema (mirrors rhea_go_mapping). Object labels in the safe/generated sets
  come from the GO ontology, so labels match.
• Marginal-vs-interpro2go: computed (compare_cazy_interpro.py, 20% masked) and closure-filtered
  (closure_cazy_interpro.py, 97% of marginal survives: 74% altitude, 23% true-gap).
• TRUE_GAP hand-review: done — cazy2go-truegap-review.md
  (34 ENDORSE / 13 CAUTION / 6 REJECT; artifacts auto-filtered from the safe set).
• Subfamily resolution: done — cazy2go-subfamily.md. dbCAN-sub was
  inaccessible, so resolved via InterPro co-occurrence within reviewed members
  (subfamily_resolve.py, cazy2go.subfamily.tsv):
  66/90 poly-specific families resolved; 71 CONFIRMS (re-derives existing interpro2go — method
  validation) + 210 family-specific PROPOSED signature→GO candidates across 56 families (e.g.
  GH28→polygalacturonase, GT4→sucrose synthase, GH47→α-1,2-mannosidase). Fixes the poly-specific
  safety problem by attributing each activity to its InterPro-signature subfamily.
• Next: (1) human GO/InterPro2GO curator sign-off on the 34 ENDORSE true-gaps + 210 subfamily
  proposals before any submission; (2) GlycoCoO→GO alignment SSSOM; (3) GlyGen-join confirmatory probe.