RHEA vs EC: Masking, Specificity, and Mapping Gaps

Parent project: RHEA.md · Methodology: RHEA-METHODOLOGY.md

This page answers three questions that the first scoping pass deferred: how much
of RHEA's GO contribution is masked by EC, what happens to reaction
specificity on the way to GO, and what concrete mapping gaps exist. All
numbers are computed live by rhea_ec_specificity.py
from three public sources — rhea2go and ec2go (GOA external2go) and RHEA→EC
(RHEA REST API) — at GO release 2026-05-19. Nothing here is hardcoded.

The single most important caveat from the methodology carries over: GO-term
identity and GO-term set membership are exact-match. Where it matters
(specificity, true gaps) a closure-aware follow-up is still required; the
exact-match figures are bounds, and I say so each time.

1. RHEA is mostly masked by EC

UniProt enzymes almost always carry both an EC number and a RHEA reaction.
EC reaches GO via ec2go (GO_REF:0000003); RHEA reaches GO via rhea2go
(GO_REF:0000116). So before crediting RHEA with a contribution, we must ask
whether EC2GO already delivered the same term.

At the GO-term level:

At the reaction level (the 4,904 RHEA reactions that carry both a rhea2go
term and an EC number):

Takeaway. For ~88% of enzymatic reactions, the GO_REF:0000116 annotation
RHEA contributes is the same term EC2GO already contributes — RHEA is
redundant at the GO layer and its row would be dropped by the closure-filtered
uniqueness query (it is masked by an EC sibling). RHEA's real, unmasked value
is concentrated in four pockets:

1. the 772 RHEA-only GO terms EC2GO cannot produce,
2. the 462 reactions whose EC has no ec2go line (RHEA fills an EC2GO gap),
3. the 118 differing-term reactions where RHEA resolves a distinction EC's
   own GO term blurs, and
4. the reverse-propagation gap (UniProt RHEA annotations that reach neither
   EC2GO nor RHEA2GO in GOA — see RHEA.md pilot).

This is the concrete form of "RHEA contributions may be masked by EC": at the
term level most are, so a RHEA audit should be driven by the four unmasked
pockets, not by raw GO_REF:0000116 volume.

2. Specificity cuts both ways

RHEA is finer-grained than EC: one EC number routinely spans many RHEA reactions
(different substrates, cofactors, or directions). The question is whether that
fineness survives the mapping to GO.

2a. EC → RHEA → GO: specificity is usually flattened, sometimes kept

Of the 4,235 ECs that map to ≥1 rhea2go reaction, 435 map to more than one:

The 112 "spread" cases are where RHEA is most valuable: a single EC hides a
distinction that RHEA, and the GO terms it maps to, preserve. The cleanest
example is a cofactor split:

EC:1.1.1.256 → two RHEA reactions →
GO:0004090 carbonyl reductase (NADPH) activity and
GO:0004022 alcohol dehydrogenase (NAD+) activity

Here the single EC number is ambiguous about cofactor/substrate, but the two
RHEA reactions land on two distinct, more informative GO terms. For a protein
where UniProt asserts the specific RHEA, RHEA→GO yields a better MF term than
EC→GO would. (Other examples the script prints: EC:1.1.1.94 →
NADP+-specific glycerol-3-phosphate dehydrogenase; EC:3.1.3.56 → a specific
inositol-polyphosphate phosphatase.)

2b. RHEA → GO: many reactions collapse onto one generic term

The flip side. 679 GO terms are each backed by more than one RHEA reaction,
and the top "absorbers" swallow dozens of distinct reactions:

These are reactions that differ by substrate (which alcohol, which acyl-chain
length, which acceptor sugar) but for which GO has no substrate-specific
child — so 67 distinct RHEA reactions all flatten to one
alcohol dehydrogenase (NAD+) activity term. The specificity RHEA encodes is
discarded at the GO layer not because of a mapping error but because the GO MF
branch lacks the granularity. These are the natural candidates for
proposed_new_terms if substrate-specific MF resolution is ever wanted.

Net specificity picture. Going into RHEA you gain resolution over EC;
going out to GO you usually lose it again (74% collapse), except for the 26%
of multi-reaction ECs where GO happens to carry the cofactor/substrate split.
RHEA→GO is therefore a specificity bottleneck, occasionally a specificity
rescue.

3. Mapping gaps found

How the gaps interact

G1 (masking) and G4 (no-target) are mirror images: where EC and RHEA agree, RHEA
is redundant; where RHEA has no GO term, EC often still carries the protein via
its coarser EC-level term — so the enzyme is rarely left with no GO MF
annotation, it is left with a less specific one. The curation-relevant gaps are
therefore mostly specificity gaps (G4/G5 — the right activity exists in
biochemistry but GO represents it only coarsely) rather than coverage gaps
(the enzyme has no MF term at all), with G2/G3/G6 being the places RHEA genuinely
adds something EC does not.

Reproduce

uv run python rhea_ec_specificity.py     # all section-1/2/3 numbers
uv run python rhea_go_gap_probe.py --gap-sample   # the reverse-propagation pilot

ftp.expasy.org (the usual RHEA bulk-download host) is blocked by the web
container's network policy; the script uses the RHEA REST API
(https://www.rhea-db.org/rhea?columns=rhea-id,ec) instead, which is reachable.