Swiss-Prot Annotation Gain & Coverage

Parent project: RHEA.md - Mappings: rhea2go.sssom.yaml - Script: rhea_annotation_gain.py

Gain is measured from Swiss-Prot (reviewed) entries: how many reviewed entries carry the RHEA
reaction but do not already have the mapped GO term (go: is closure-aware, so a more specific
child counts as covered).

Headline

Swiss-Prot coverage status

We systematically scored all 256 remaining unmapped EC-bridge reactions (RHEA whose EC maps to
a single specific GO term) for Swiss-Prot gain. Only 4 had any reviewed gap; all 4 are now
mapped. So every cleanly-mappable RHEA reaction on a reviewed entry is now covered -- the
EC-bridge Swiss-Prot residual gain is 0. Reviewed enzymes carrying a reaction almost always
already have the term (manual curation / EC2GO), so the curated gain is intrinsically small.

What remains is a missing-GO-term problem, not a mapping problem

The only Swiss-Prot entries with a RHEA still lacking a reaction-specific term are those whose
reaction has no GO term to map to at all. Scoping the new-term-needed set (unmapped reactions
whose EC is absent from ec2go):

• ~2,331 such reactions exist; ~1,942 appear on at least one reviewed entry
  (150-reaction sample: 125/150 had a reviewed entry).
• These cannot be closed by adding rhea2go rows -- each needs a new GO molecular-function
  term (a GO term-request effort). We have seeded 7 such proposals (sssom:NoTermFound);
  the remainder is the backlog to full reviewed coverage.

So: full Swiss-Prot coverage via existing GO terms is achieved; full coverage period is
gated on creating ~1,900 new GO terms, not on more mappings.

Swiss-Prot gains (every reviewed annotation the mappings add)

Total: 42 reviewed annotations across 14 mappings.

Reproduce

uv run python rhea_annotation_gain.py

Litmus test: is there a missed-mapping gap? (no)

Per the project's value test — would a Swiss-Prot curated entry lose a good,
non-redundant GO annotation if we did not make the mapping? — and without
assuming EC2GO will be extended, I hunted the no-ec2go space for reactions
where a correct GO term already exists but isn't reachable by EC/RHEA xref.

First, a structural check: GO's own RHEA xrefs equal rhea2go exactly (7,746,
zero difference) and GO's EC xrefs equal ec2go exactly (4,830, zero
difference). So there is no hidden bridge — a reaction unmapped in rhea2go whose
EC is absent from ec2go has no xref path to GO at all.

Name-matching reviewed enzymes for those reactions to GO terms produced 30
Swiss-Prot-gain candidates — but verification against each GO term's definition
reaction rejected all of them: the matched term is for a different reaction
(different cofactor, electron acceptor, or stereochemistry), which is exactly why
GO keeps it separate and why it isn't xref-linked:

The remaining candidates matched generic class roots (hydro-lyase activity,
demethylase activity, 5'-nucleotidase activity) or were outright name
collisions (RHEA:16301, an IL-18-receptor protein carrying a spurious NADase
reaction, matched interleukin-18 receptor activity; another matched iron ion binding). None is a good non-redundant annotation.

Conclusion. There is no missed-mapping gap: every reviewed entry whose
reaction has a correct existing GO term is already covered (the EC-bridge work
above), and the reactions that remain genuinely lack a correct GO term —
GO distinguishes the cofactor/stereochemistry variants they represent. So closing
the rest is strictly a new-GO-term effort (the 7 seeded sssom:NoTermFound
rows are the start), not more mapping. This holds the line on the litmus test:
we do not add a mapping unless a Swiss-Prot entry would gain a correct,
non-redundant annotation.

Turning the litmus result into new-term proposals

The reactions the litmus hunt surfaced are precisely the new-term backlog: a
reviewed enzyme carries the reaction, Swiss-Prot lacks a correct term, and GO has
only a sibling term for a different variant. Batch 6 promotes 10 of these to
verified sssom:NoTermFound proposals (definition-checked against the sibling),
e.g. ferredoxin-dependent protochlorophyllide reductase (EC 1.3.7.7, ~327
reviewed entries; GO has only the NADPH POR GO:0016630), alcohol dehydrogenase
(quinone) (EC 1.1.5.5), 2-methylcitrate dehydratase (trans-forming) (vs the
cis GO:0047547), and UDP-N-acetylglucosamine 2-epimerase (hydrolysing) (human
GNE; vs the non-hydrolysing GO:0008761). The proposal set is now 17.