RHEA → GO Contribution & Gap Project

IN_PROGRESS PIPELINE

RHEA → GO Contribution & Gap Project

Bottom line: RHEA reactions reach GO molecular function through the
rhea2go mapping (GO_REF:0000116), but UniProt enzymes nearly always carry an
EC number too, and ec2go already supplies the same GO term for 88% of reactions
that have both. We measured where RHEA adds something EC does not (772 GO terms
reachable only through RHEA, 462 reactions whose EC has no ec2go line), where
reactions have no GO target at all (2,731 of 7,635 enzymatic reactions, 36%),
and where the mapped term fails to reach reviewed entries. We then curated 154
new RHEA-to-GO mappings, each backed by a reviewed enzyme: 109 ready-to-add exact
matches, 17 broad matches, and 28 new-term suggestions where GO has no specific
activity. Added to rhea2go, the GO-bearing rows would give 71 new annotations to Swiss-Prot
entries; after scoring all remaining EC-bridge reactions, no cleanly mappable
reviewed gap is left, so the remaining problem is missing GO terms rather than
missing mappings. A per-family scan of three promiscuous families sharpens
this: in human cytochrome P450s 68% of curated reactions have no mapping at all
(and 95% of those carry no EC either), while the closure-filtered reverse gap is
empty — so drop-out, not mis-propagation, is what is left to fix. The SSSOM set
validates; the forward cross-organism scan is still pending.

We did this because RHEA is a reaction-grounded, curated source of enzyme
function, and we needed to know whether its GO contribution is real or only
duplicates EC before relying on it in gene reviews.

Overview

This project examines the RHEA reaction database as a source of GO molecular-
function annotations, in the same spirit as the SPKW (UniProt
keywords) and UniPathway source-audit projects. RHEA reaches GO
through the public rhea2go mapping; in GOA these annotations appear as
GO_REF:0000116, assigned_by=RHEA, aspect molecular_function, evidence
IEA. GO_REF:0000116 is the RHEA counterpart of SPKW's GO_REF:0000043 and
EC2GO's GO_REF:0000003.

Unlike SPKW and UniPathway, RHEA is a molecular-function source (enzymatic
reactions), so the analysis runs in two complementary directions:

  1. Contribution (forward). Where is GO_REF:0000116 the unique
    (closure-filtered) source of an enzyme-activity annotation, and are those
    contributions useful or over-/mis-propagated? — the SPKW-style audit.
  2. Gaps (reverse). Where does a UniProtKB entry carry a catalytic activity
    with a RHEA cross-reference, but the GO molecular-function term that
    rhea2go maps that reaction to never propagates into GOA? These are
    annotations that "fall through the cracks" — curation gap-filling
    opportunities rather than over-annotations.

See RHEA-METHODOLOGY.md for the queries, the
reproducible probe script, and the all-important closure caveat. The deeper
EC-masking and specificity analysis lives in
RHEA-EC-SPECIFICITY.md.

Key Findings (scoping pass)

RHEA is mostly masked by EC

UniProt enzymes almost always carry both an EC number and a RHEA reaction,
and EC also reaches GO — via ec2go (GO_REF:0000003). So RHEA's GO_REF:0000116
contribution is only "real" where EC2GO did not already supply the same term.
Computed live from rhea2go, ec2go, and the RHEA REST API
(rhea_ec_specificity.py):

So most GO_REF:0000116 volume is redundant with EC2GO and would be dropped by
the closure-filtered uniqueness query. RHEA's genuine value is the 772
RHEA-only terms, the 462 EC-without-ec2go reactions, the 118
differing-term reactions, and the reverse-propagation gap below.

Specificity cuts both ways

Full tables and examples: RHEA-EC-SPECIFICITY.md.

Reverse-Gap Pilot (live, exact-match)

Reviewed-only UniProtKB entries carrying each reaction, and how many lack the
rhea2go-mapped GO molecular-function term. Generated with
rhea_go_gap_probe.py (--gap-sample).
pct_missing is an exact-match upper bound — see the closure caveat below.

RHEA Mapped GO MF term N reviewed N missing (exact) % missing Note
RHEA:21248 GO:0034062 5'-3' RNA polymerase activity ≥500 497 ~99% capped at 500; striking — generic parent rarely used directly
RHEA:10596 GO:0004713 protein tyrosine kinase activity ≥500 198 ~40% capped at 500; many entries carry specific RTK children
RHEA:13065 GO:0016887 ATP hydrolysis activity ≥500 33 ~7% capped at 500
RHEA:15421 GO:0004467 long-chain fatty acid-CoA ligase activity 90 2 ~2% clean propagation
RHEA:16505 GO:0008813 chorismate lyase activity 148 0 0% clean propagation
RHEA:11628 GO:0004737 pyruvate decarboxylase activity 33 0 0% clean propagation
RHEA:22636 GO:0047840 dCTP diphosphatase activity 49 0 0% clean propagation
RHEA:46608 (no rhea2go mapping) — — — "no GO term" gap class → candidate new term

Reading the pilot. The clean 0–2% reactions show the pipeline works well for
specific, single-product enzyme activities. The high-gap reactions
(RHEA:21248, RHEA:10596) are exactly the generic-parent reactions where
rhea2go points at a broad activity term that curated entries express via more
specific children — so most of that 40–99% is expected closure coverage, not a
genuine missing annotation. The job of the closure-filtered follow-up is to
separate the residual true gaps (entries with neither the parent nor any
child) from this expected altitude difference.

Mapping Gaps Found

# Gap Size What it is
G1 EC-masking redundancy 4,324 reactions (88%) RHEA's GO term duplicates one EC2GO already supplies → low marginal value
G2 RHEA-only GO terms 772 terms Reachable via rhea2go but not ec2go — RHEA's unique forward contribution
G3 EC-without-ec2go 462 reactions EC has no ec2go line → RHEA is the only EC-adjacent GO route
G4 No-rhea2go enzymatic reactions 2,731 / 7,635 (36%) Genuine enzymatic reaction with no GO MF target (2,445 EC-level; 286 specific-reaction-only)
G5 Specificity-collapse 679 terms, up to 67:1 GO lacks substrate-specific MF children → many reactions flatten to one term
G6 Reverse-propagation gap pilot below UniProt RHEA annotations whose mapped GO term never reaches GOA

Worked case reviews of this gap on real Swiss-Prot enzymes — where a RHEA
reaction has no rhea2go mapping and the GO molecular function is missing or
only a class root — are in RHEA-GAP-CASES.md: PHYKPL
(MF = only lyase activity; propose new term), B3GALNT2 (class-level
GalNAc-T; propose new term; dystroglycanopathy gene), SAMD8/SMSr (a
lookalike-term case where GO:0002950 ceramide phosphoethanolamine synthase activity specifies CDP-ethanolamine rather than the RHEA reaction's
phosphatidylethanolamine donor; propose a PE-dependent new term), and
SULT6B1 (only sulfotransferase activity; cautious fill given
by-similarity evidence). Cases selected reproducibly by
rhea_gap_finder.py.

G1/G4 are mirror images: where EC and RHEA agree RHEA is redundant; where RHEA
has no GO term EC usually still carries the protein at coarser EC granularity —
so most gaps are specificity gaps (right activity, coarse GO representation),
not coverage gaps (no MF term at all). G2/G3/G6 are where RHEA genuinely adds
something EC does not. Detail and reproduction in
RHEA-EC-SPECIFICITY.md.

Promiscuous enzyme families: where the mapping drops out

The specificity-collapse result above (G5) was measured from the GO side — one
term absorbing many reactions. Measured from the protein side, in three
families picked for substrate promiscuity, the dominant failure is not collapse
but drop-out: of the 297 distinct reactions annotated to reviewed human
cytochrome P450s, 201 (68%) have no rhea2go target, and 191 of those
carry no EC number either — EC stops at "unspecific monooxygenase" where RHEA
continues to the substrate, so the EC bridge cannot rescue them. Human
carboxylesterases and microbial haloalkane dehalogenases show the same shape at
smaller scale.

Equally important, the reverse gap is zero in these families once is_a
closure is applied: none of the 98 entries lacks a GO term that a mapped
reaction would give it. That is direct confirmation of the closure caveat below
— the pilot's 40–99% exact-match figures were altitude difference, not missing
annotation — and it moves the remaining value from G6 to G4.

The haloalkane family supplies the rule that separates a mapping gap from a term
gap: GO once had the substrate-specific term GO:0018777 for LinB's
tetrachlorocyclohexadiene reaction and obsoleted it as "a specific substrate
of" GO:0018786. (Verified against QuickGO, GO release
releases/2026-09-27; the GO comment is itself truncated mid-sentence.) GO does not create a term per substrate, so an instance
reaction inside a class the term already names deserves a mapping, not a
new-term request.

Full analysis, per-family tables and the 18 resulting SSSOM rows (46 new
Swiss-Prot annotations — including the CYP1B1 estrogen 4-hydroxylase, CYP2J2
EPA/DHA epoxygenase and CYP11B2 aldosterone-synthase term proposals, and the
CYP4F22 ichthyosis-gene gap): RHEA-PROMISCUOUS-FAMILIES.md
(rhea_family_explorer.py). The P450 half of this
is followed up family-wide in the CYTOCHROME_P450 project,
which adds batch 8: the cascade-step gaps, where GO holds a reaction's chemistry
for one substrate (the leukotriene B4 ω-oxidation trio) and not for another on the
same protein.

Curated new mappings (SSSOM)

Filling the gaps is a curation deliverable, not just an audit. The curated
RHEA→GO mappings — reactions absent from rhea2go but with (or needing) a GO MF
term — are recorded in rhea2go.sssom.yaml, the same
SSSOM YAML format used by the ANTIMICROBIAL_RESISTANCE
aro2go mapping set. 154 mappings so far, each backed by a review of a reviewed
(Swiss-Prot) enzyme that carries the reaction
(RHEA-MAPPING-REVIEWS.md). The predicate encodes
the specificity finding:

Every GO id/label is verified non-obsolete against QuickGO; every RHEA id +
equation comes from the UniProt catalytic-activity line; every backing enzyme is
reviewed (mostly PE1, with catalytic-activity PMIDs). Validate with
just validate-rhea-mappings — SSSOM structural validation plus GO term/label
validation (GO objects bound to the molecular-function branch; generated nested
view rhea2go.terms.yaml).

Propagation gain. If these mappings were added to rhea2go, they would
add 71 new GO molecular-function annotations to Swiss-Prot (reviewed) entries, of
which 46 come from the promiscuous-family batch 7 — the
curation-relevant gain we track — filling real reviewed-entry gaps. The all-UniProtKB
figure is secondary and should be recomputed whenever the GO-bearing mapping subset
changes: reviewed enzymes carrying these reactions usually already have the term,
echoing the EC-masking result at the annotation level. See
RHEA-ANNOTATION-GAIN.md
(rhea_annotation_gain.py).

Methods

GO_REF:0000116 identification and the rhea2go mapping statistics were
computed live (GO release 2026-05-19). The reverse-gap pilot was computed live
against the UniProtKB REST API (rhea:<id> query field, reviewed:true).

The forward contribution audit reuses the closure-filtered uniqueness query
from UniPathway/SPKW, swapping the
reference id to GO_REF:0000116. That cross-organism scan needs the local
~/repos/go-db/db/*.ddb DuckDBs, which are not present in the web container,
so the UNIPATHWAY-style contribution table is staged, not yet populated (see
Follow-Up Targets; tracked in
#4013). The
reverse-gap probe needs no go-db and runs anywhere with network access.

Full queries, the probe script, and the mandatory closure caveat are in
RHEA-METHODOLOGY.md.

How this differs from SPKW and UniPathway

SPKW (GO_REF:0000043) UniPathway (GO_REF:0000041) RHEA (GO_REF:0000116)
GO aspect mostly BP (+ some MF/CC) mostly BP (pathways) MF (enzyme activity)
Dominant failure mode process conflation / "expression ≠ function" broad pathway-bucket propagation parent-vs-child altitude; wrong-substrate paralog
Direction of interest over-annotation (remove/modify) mostly contribution audit both: contribution and reverse gap-filling
Source status retired by GOA (~Apr 2026) for cellular organisms archived/legacy vocabulary active, curated, reaction-grounded

The RHEA source is better grounded than SPKW (a reaction is a precise
biochemical claim), so the expected verdict skew is toward ACCEPT /
KEEP_AS_NON_CORE on the forward side and toward NEW / gap-filling on the
reverse side — the opposite emphasis from the SPKW over-annotation hunt.

Curation Recommendations (preliminary)

  1. Treat the gap direction as the high-value half. Unlike SPKW, RHEA's main
    contribution is finding correct enzyme activities that GO is missing, not
    removing over-annotations. Prioritise reactions with a high closure-filtered
    gap.
  2. Always closure-filter before calling a gap. Exact-match gap percentages
    are upper bounds; subtract entries carrying any descendant of the mapped term
    (the SPKW/UniPathway closure pattern) before proposing additions.
  3. Generic-parent reactions deserve altitude review. Where rhea2go maps a
    reaction to a broad activity term (e.g. protein tyrosine kinase activity),
    prefer the specific child the curated entry already supports; the parent is
    often redundant.
  4. Reactions with no rhea2go line are proposed_new_terms candidates — a
    real UniProt-annotated activity with no GO MF target to land on.
  5. Forward-unique RHEA rows are usually keepers. A closure-unique
    GO_REF:0000116 activity term is a precise biochemical assertion; the review
    risk is wrong-substrate/wrong-paralog specificity, not process conflation.

Follow-Up Targets

Target Rationale
Forward closure-filtered cross-organism scan (GO_REF:0000116) The UNIPATHWAY-style contribution table; needs go-db DuckDBs. Quantifies where RHEA is the unique MF source.
Closure-aware reverse gap on the high-gap pilot reactions Separate true gaps from expected parent/child altitude for RHEA:21248, RHEA:10596; promote real gaps to gene reviews.
"No rhea2go mapping" reaction set Enumerate UniProt-used RHEA reactions with no GO MF target → batch proposed_new_terms.
RHEA directional-quartet join audit Test whether master-vs-directional id mismatch causes systematic non-propagation.
Exemplar gene reviews Pick 2–3 confirmed (closure-filtered) gap genes and run the full review workflow, mirroring the UniPathway exemplar pattern. CYP4F22 (ichthyosis; no specific MF term) and CYP11B2 (no aldosterone-synthase term) are the strongest candidates from the family analysis.
Extend the family scan rhea_family_explorer.py takes any UniProt query; other promiscuous families (UGTs, SULTs, FMOs, ABC transporters' ATPase reactions) should show the same drop-out signature.

Open follow-up is tracked in
#4013.

Project Status