TCDB → GO Transporter-Activity Mapping & Gap Project

Bottom line: the Transporter Classification Database (TCDB) gives each
transport system a TC number, the transport analogue of an EC number, and GO
curators have already attached TC references to 170 molecular-function terms.
But GO has no tc2go annotation pipeline, so a protein's TC number never
becomes a GO annotation: half of the reviewed Swiss-Prot transporters with a TCDB
cross-reference (4,431 of 8,758) carry no transporter-activity term. We
extracted GO's 194 TC-to-GO leads and scored each against the reviewed proteins
that carry the TC id. 80 are safe to propagate, almost all at the level of a
single TC system; 18 would over-propagate at the family level; 67 point to
specific systems whose members lack the specific term; the other 29 are
whole-class leads (12), leads with no reviewed member (10) or uncertain (7). We
also filtered TCDB's
own noisy GO dump into 477 machine-derived candidate rows and hand-backed eight
exemplar mappings. All four SSSOM sets validate.

We did this because EC and RHEA each turn a classification into GO annotations
through a pipeline, and transporters have no equivalent. The evidence-scored
leads are a ready starter set for proposing a tc2go pipeline to GO.

Overview

This project examines the Transporter Classification Database (TCDB) as a
source of GO molecular-function annotations for membrane transport proteins, in
the same spirit as the RHEA (reactions), EC/ec2go, and
CAZy source-audit projects. TCDB is the only transporter
classification adopted by the IUBMB, organising transport systems into a
five-level TC number (class.subclass.family.subfamily.system, e.g.
2.A.69.1.1) that is the transport analogue of an EC number.

The defining fact of this project — and what makes it different from RHEA and EC
— is a missing annotation pipeline, not a missing mapping:

GO already curates TCDB cross-references on its terms, but there is no
tc2go external2go annotation pipeline. GO curators have attached TC
references to 170 non-obsolete molecular-function terms (via xref: TC:
clauses and definition dbxrefs) — a real, high-quality partial TC→GO mapping
that lives inside the ontology. What is missing is the propagation half: TCDB
is absent from GO's external2go directory (which ships ec2go, rhea2go,
interpro2go, … but no tc2go), so unlike EC (GO_REF:0000003) and RHEA
(GO_REF:0000116) a protein's TCDB classification is not turned into a GAF
annotation. The correspondence exists on the terms; it just isn't applied to
proteins. Both facts confirmed live (GO release 2026-05-19; go-basic.obo).

So the analysis runs in two directions:

1. Forward (candidate sources). GO's 170 TC xrefs are the best starting leads,
   but neglected and unreviewed. TCDB also publishes a per-protein GO dump
   (https://tcdb.org/cgi-bin/projectv/public/go.py), noisier but wider. Both are
   candidate pools; neither is safe to propagate without per-entry curation.
2. Reverse (gap). Where does a UniProtKB entry carry a TCDB cross-reference
   (DR TCDB;) but no GO transmembrane-transporter-activity term at all?
   Because GO's xrefs sit on terms and are never propagated to proteins, this
   "falls through the cracks" gap is structural, not incidental.

See TCDB-METHODOLOGY.md for queries, the reproducible
probe script, and the closure caveat.

Key Findings (scoping pass)

• GO carries TC cross-references, but as neglected sources (the correction to
  the headline). GO terms hold curator TC references: 202 references on 170
  non-obsolete molecular-function terms, spanning 63 TC families / 185 TC
  systems (extracted live from go-basic.obo by
  extract_go_tc_xrefs.py). These are real leads,
  but they are largely neglected and are not a ready mapping: an xref does not
  by itself justify propagating the GO term to every protein with that TC id.
  That propagation judgment is curated by hand, per entry (see below).
• What is actually missing is the annotation pipeline. TCDB is not in GO's
  external2go directory, so those curated term-xrefs are never propagated to
  the proteins that carry the TC number. EC and RHEA turn a classification into
  a GAF annotation (GO_REF); TCDB does not. That, not the absence of a mapping,
  is the structural gap.
• TCDB also publishes a raw GO dump, but it is noisy and multi-aspect. The
  go.py file has 34,497 rows mapping 4,943 distinct TC systems
  (5-level) across 589 TC families (3-level) to 3,518 distinct GO terms —
  but only 13% of rows (4,557) and 12% of GO terms (424) are actually
  transmembrane-transporter-activity (molecular function). The dump is
  dominated by cellular-component and biological-process terms
  (GO:0016021 integral component of membrane ×3,876;
  GO:0005886 plasma membrane ×2,300; GO:0055085 transmembrane transport
  ×1,422; GO:0006810 transport ×882) plus outright noise
  (GO:0005515 protein binding ×452; GO:0005488 binding ×145;
  GO:0003824 catalytic activity ×56 — on transporters). It is an aggregation
  of member-protein annotations, not a curated mapping, so it cannot be
  adopted as tc2go wholesale.
• The reverse gap is large and structural. Of 8,758 reviewed (Swiss-Prot)
  entries carrying a TCDB cross-reference, 4,431 (50.6%) carry no GO
  transporter-activity term at all (GO:0005215 closure). This is the
  headline: half of all reviewed transporters with an IUBMB TC classification
  have no transport MF term propagated from it. (Exact-match upper bound on the
  gap; see caveat.) Across all of UniProtKB the TCDB cross-reference count is
  16,344.
• The gap is a specificity problem, mirroring RHEA. A TC family
  (3-level) is frequently poly-specific: the MFS superfamily 2.A.1 alone spreads
  across 49 distinct substrate-specific MF terms in go.py
  (D-glucose, fructose, nitrate, heme, biotin, spermine, …); the APC family
  2.A.3 across 11. The right GO target lives at the subfamily level, not
  the family — exactly the parent-vs-child specificity lesson RHEA and CAZy already
  learned. Some families are cleanly mono-specific (e.g. 2.A.69 auxin efflux
  carrier → GO:0010329).
• This repo already contains the exemplars. 353 gene folders carry a
  DR TCDB; cross-reference (e.g. PIN1, AUX1, SOS1, HKT1, CHL1/NRT1.1, AQP1),
  several with experimental transport-activity GO terms — ready-made backing for
  a curated tc2go seed and for closure-filtered gap review.

GO's TCDB cross-references: neglected sources, not a ready mapping

GO carries TC references on molecular-function terms. Extracted live from
go-basic.obo (extract_go_tc_xrefs.py, obsolete
terms excluded):

	Count
TC xref references on GO terms	202 (61 xref: TC: + 141 definition dbxref)
Distinct TC→GO source pairs (deduped)	194
Distinct GO terms (all molecular function)	170
Distinct TC systems (5-level)	185
Distinct TC families (3-level)	63

These are leads, not an asserted mapping. GO's TC xrefs are largely
neglected, and the term-xref: vs definition-dbxref distinction carries no
reliable signal, so they are treated alike. Crucially, an xref says "a curator
once linked this TC entry and this GO term" — not "every protein with this
TC number has this function". Whether a lead is safe to propagate (a protein
inheriting the GO term from its TC id) is a separate judgment that must be
curated by hand, per entry — see the propagation verdicts below.

They are shipped, inverted to TC→GO, as
tc2go.from_go.sssom.yaml — every row
skos:relatedMatch + semapv:UnspecifiedMatching (i.e. explicitly unreviewed
source). Many are keyed at the specific TC system (e.g. GO:0005335 ← TC:2.A.22.1.1, SERT), where propagation is usually safe; others sit on a whole
family, where it usually is not. They cover only 63 of TCDB's 2,235 families
(3%) and, being ontology annotations rather than an external2go mapping, are
not propagated to proteins — so they do not close the reverse gap below.

Propagation curation: every GO-xref lead scored by evidence

Whether a lead is safe to propagate — can a protein carrying the TC id inherit
the GO term? — is answered for all 194 source pairs with UniProt member
evidence, not assertion. For each pair, curate_propagation.py
fetches the reviewed Swiss-Prot proteins carrying that TC id (level-aware: a
5-level system matched exactly, a coarser id by dot-bounded prefix) and measures
what fraction also carry the GO term (ontology-closure expanded, so a member with
a more specific child still counts). The fraction is the propagation signal:

Verdict (SSSOM predicate)	Rule	Count
JUSTIFIED (exactMatch)	≥2 members, ≥70% carry the term (or 1/1)	80
GAP_CANDIDATE (relatedMatch)	specific system (≥4-level), member(s) exist but 0 carry it	67
NOT JUSTIFIED at this TC level (narrowMatch)	≥3 members, <50% carry it	18
CLASS_LEVEL (broadMatch)	whole TC class/subclass (≤2-level); broad by construction	12
NO_REVIEWED_MEMBER (relatedMatch)	no reviewed protein xref'd to the TC id	10
UNCERTAIN (relatedMatch)	ambiguous small-n middle	7

The rows are listed in evaluation order, which matters for the two middle rules:
k=0 also satisfies "<50% carry it", so GAP_CANDIDATE is tested first. A specific
system whose members all lack the term is a reverse-gap lead (the members are
under-annotated), not evidence that the term fails to propagate — filing it as
narrowMatch would both hide it from the gap harvest and overstate the narrow count.

Shipped as tc2go.propagation.sssom.yaml
with the per-pair evidence (k/n) in every comment; the full table is
data/propagation_evidence.tsv. Findings:

• Propagation is safe at the system level. 70 of the 80 JUSTIFIED pairs are
  5-level TC systems (e.g. 3.A.2.1.1 → GO:0046933 proton-motive-force-driven ATP synthase activity, 8/8; TC:2.A.19.2.- → GO:0015369 calcium:proton antiporter,
  8/8) — the evidence backs the "curate at the system, not the family" rule.
• narrowMatch is where an xref would over-propagate. e.g. the MFS
  sugar-porter subfamily 2.A.1.1 → GO:0005351 carbohydrate:proton symporter activity (only 35/88 members, 40%), or 1.A.9 → GO:0160039 serotonin-gated chloride channel activity (1/59): correct for a few members, wrong for the rest.
• GAP_CANDIDATE is a reverse-gap harvest. 67 leads are specific systems where
  GO curates a precise term but the reviewed member is annotated only to a
  parent/sibling — e.g. 2.A.1.1.10 → GO:0005364 maltose:proton symporter whose
  member P15685 carries the general carbohydrate:proton symporter activity but not
  the maltose-specific child. These are refinement leads for the gene-review
  workflow, not propagation bases.
• CLASS_LEVEL (12) and NO_REVIEWED_MEMBER (10) are honestly out of scope —
  class/subclass xrefs (TC:1 → channel activity) are broad by construction; the
  no-member cases have only TrEMBL entries or a reviewed protein that lacks the
  DR TCDB line (a UniProt xref gap, e.g. MelB). Neither is silently included.
• Several k<n JUSTIFIED cases are also gap finds — e.g. 3.A.3.1.1 → GO:0005391 (4/5): the activity is shared, so the 5th member is a missing
  annotation.

TCDB's own go.py dump: usable seed, but only after filtering

Slice	Count	% of dump
Rows (GO ↔ TC-system)	34,497	100%
Distinct TC systems (5-level)	4,943	—
Distinct TC families (3-level)	589	—
Distinct GO terms	3,518	—
Rows whose GO term is a transporter-activity MF	4,557	13%
Distinct transporter-activity MF terms	424	12% of GO terms
TC systems with ≥1 transporter-activity MF term	2,866	58%
TC families with ≥1 transporter-activity MF term	222 / 589	37%

Rows by TC class: 1 channels/pores 7,984 · 2 electrochemical-potential-driven
11,959 · 3 primary active 10,290 · 4 group translocators 890 · 5
transmembrane electron carriers 693 · 8 accessory 1,092 · 9 incompletely
characterised 1,589. Computed live by
tcdb_go_probe.py (--tcdb-go); the transporter-activity
MF closure (1,042 terms, is_a GO:0022857 ∪ GO:0005215) comes from QuickGO.

The take-home: the clean, adoptable slice of go.py is the 13% that is
transporter-activity MF. build_tc2go.py distils exactly
that slice into an SSSOM candidate set (below).

Reverse gap: half of reviewed transporters lack a transport MF term

Reviewed Swiss-Prot entries with a DR TCDB; xref	Count	%
Total	8,758	100%
…with GO:0022857 transmembrane transporter activity (closure)	4,208	48%
…with GO:0005215 transporter activity (closure)	4,327	49%
…with NO transporter-activity term (gap upper bound)	4,431	50.6%

Computed live against the UniProtKB REST API (database:tcdb AND reviewed:true,
GO-closure-expanded on the GO side) by tcdb_go_probe.py
(--gap). GO's curated term-xrefs do not help here — they sit on GO terms,
not on the proteins — so the only routes by which these transporters could get a
transport MF term are EC/ec2go (many transporters have no EC),
InterPro/interpro2go, or manual curation. A structural half-coverage is
therefore unsurprising, and is precisely the gap a tc2go propagation pipeline
would close.

Closure caveat (mandatory). The 50.6% is an exact-match upper bound: some
entries carry a transport-related term outside the GO:0005215 MF branch (a
transport process BP term, or a channel-complex CC term) that a closure-aware,
cross-branch audit would credit. As in RHEA/CAZy, high gap = candidate for
closure-filtered review, not a confirmed missing annotation.

Specificity: the TC family is usually too general

• TC → GO is a specificity bottleneck. Poly-specific superfamilies collapse
  many substrates onto one family id: 2.A.1 (MFS) touches 49 distinct
  transporter-activity MF terms, 2.A.3 (APC) 11, 2.A.55 (NRAMP) 11. A single
  family→GO row is therefore usually a narrowMatch (the GO term is really a
  subfamily property), not an exactMatch.
• …but some families are clean. 2.A.69 (auxin efflux carrier) →
  GO:0010329 auxin efflux transmembrane transporter activity is mono-specific —
  one GO term fits the whole family — a ready-to-add exactMatch row.
• A family name is not a specificity claim. 1.A.8 (MIP) looks like a clean
  water-channel family, but it splits into classical aquaporins (water) and
  aquaglyceroporins (glycerol), so GO:0015250 water channel activity is a
  subfamily property and the mapping is a narrowMatch. Mono-specificity has to
  be checked against the member set, not inferred from the family label.

The curated seed encodes this distinction directly (below).

Curated mappings (SSSOM)

Filling the gap is a curation deliverable, not just an audit. Four SSSOM files,
same format as RHEA and
CAZy:

• tc2go.from_go.sssom.yaml — GO's TC xrefs
  as unreviewed sources/leads (194 TC→GO pairs, 170 MF terms, 63 families;
  all relatedMatch + UnspecifiedMatching). The candidate pool to curate from
  — not a mapping to adopt as-is.
• tc2go.propagation.sssom.yaml — those
  same 194 leads scored for propagation by UniProt member evidence (80
  JUSTIFIED exactMatch, 67 GAP_CANDIDATE, 18 narrowMatch, 12 CLASS_LEVEL, 17
  other; see above). Evidence-derived (semapv:CompositeMatching), the systematic
  first pass over every lead.
• tc2go.generated.sssom.yaml — a second,
  noisier candidate pool: a machine-derived coverage extension (154 families,
  477 rows: 46 exactMatch, 431 narrowMatch) distilled by
  build_tc2go.py from the live join of TCDB's go.py
  with QuickGO, keeping only non-obsolete transporter-activity MF terms and
  aggregating to the TC family (min 2 supporting members). 110 of its 154
  families are not in GO's xrefs — the genuine coverage it adds. No row is
  hand-typed; this is TCDB's own assertion filtered for aspect + obsolescence, and
  it is family-level, so it must be reviewed before adoption.
• tc2go.sssom.yaml — the hand-curated exemplar
  layer: 8 mappings, each with an explicit propagation verdict and backed by a
  reviewed Swiss-Prot transporter read individually (deeper than the automated
  scoring). This is where a source becomes a fully-backed mapping:
• exactMatch — propagation justified (member set is mono-specific at the
  cited TC level): the 5-level systems 2.A.22.1.1 SERT → GO:0005335 (backing
  SERT/SLC6A4 P31645) and 3.A.3.1.1 Na⁺/K⁺-ATPase → GO:0005391 (ATP1A1
  P05023) — both curated up from GO sources — plus the mono-specific family
  2.A.69 AEC → GO:0010329 (PIN1 Q9C6B8, IDA/IMP).
• narrowMatch — propagation NOT justified at that TC level (the GO term is
  a subfamily property): 1.A.8 MIP → GO:0015250 water (aquaglyceroporins move
  glycerol, not water); 2.A.17 POT/PTR → GO:0015112 nitrate (CHL1/NRT1.1
  Q05085, IMP — a classic too-general-family case: a peptide transporter family
  whose plant NPF subfamily moves nitrate); 2.A.18 AAAP → GO:0010328 auxin
  influx (AUX1 Q96247, IDA); 2.A.36 CPA1 → GO:0015385 Na⁺/H⁺ (SOS1 Q9LKW9);
  2.A.38 Trk → GO:0015079 K⁺ (HKT1 Q84TI7 — see the file comment; HKT1 is the
  counter-example, not the backing).
• Corroboration by GO's own xrefs is uneven. Six of the eight sit in a family
  GO already xrefs (1.A.8 and 2.A.69 do not). For 2.A.22.1.1 → GO:0005335,
  3.A.3.1.1 → GO:0005391 and 2.A.36 → GO:0015385, GO's xref is the same
  TC→GO pair — genuine independent corroboration. For 2.A.17, 2.A.18 and
  2.A.38 only the family co-occurs: GO xrefs 2.A.17 to GO:0015333 (peptide,
  not nitrate), 2.A.38 to GO:0015387/GO:0009674 (not GO:0015079), and
  2.A.18 only at 5-level ids for other substrates. Family co-occurrence alone is
  not corroboration of the mapped term.

Every GO id/label is validated non-obsolete and molecular-function; the GO object
is bound to the MF branch. Validate all four with just validate-tcdb-mappings
(SSSOM structural validation + GO term/label validation; generated nested views
tc2go.terms.yaml,
tc2go.generated.terms.yaml,
tc2go.propagation.terms.yaml,
tc2go.from_go.terms.yaml).

Across all sources, 173 of TCDB's 2,235 families (8%) have at least one
candidate GO transporter-activity mapping.

How this differs from EC, RHEA, and CAZy

	EC (ec2go)	RHEA (GO_REF:0000116)	CAZy	TCDB
GO aspect	MF (enzyme activity)	MF (enzyme activity)	MF (glycoenzyme)	MF (transporter activity)
external2go annotation pipeline	yes	yes	no (built here from EC)	no
TC/ID references in the ontology (term xrefs)	some	some	some	yes but neglected — 170 MF terms / 63 families, unreviewed
Source GO material	curated mapping	curated mapping	via EC bridge	GO's neglected term xrefs + noisy go.py dump (both = leads)
Dominant failure mode	too-general term	parent-vs-child specificity	poly-specific family	no propagation pipeline + over-general family + go.py aspect noise
Emphasis	over-annotation audit	both	gap-filling	gap-filling (structural half-coverage)

TCDB is the most under-connected of these sources: it has a real, curated
term-level correspondence to GO, but no pipeline to propagate it to proteins, so
half of TC-classified transporters still lack a transport MF term. The expected
verdict skew is heavily toward NEW / gap-filling, tempered by the family
over-generality problem (prefer the subfamily-specific child term).

Curation Recommendations (preliminary)

1. Treat GO's TC xrefs (and go.py) as sources, then curate propagation
   per entry. An xref is a lead, not a licence to annotate: decide explicitly
   whether every protein at that TC level truly has the GO function before
   propagating. Record the verdict (exactMatch = propagate, narrowMatch =
   subfamily-only) as in tc2go.sssom.yaml.
2. Propagation is safe at the level that is mono-specific. A 5-level TC system
   is usually mono-specific (2.A.22.1.1 SERT → GO:0005335) → propagate; a
   3-level family is usually poly-specific (1.A.8 MIP mixes water channels and
   glycerol-permeable aquaglyceroporins; 2.A.22 NSS mixes serotonin/dopamine/GABA)
   → do not propagate the substrate-specific term to the whole family.
3. Filter go.py to transporter-activity MF before use. 87% of the dump is
   CC/BP/generic/noise; only the 13% MF-activity slice is adoptable, and even
   that needs specificity review.
4. Always closure-filter (cross-branch) before calling a gap. The 50.6% is an
   upper bound; subtract entries carrying any transport-related descendant/BP/CC
   term first.
5. Prefer EC-bridge support where it exists. Some transporters (e.g. primary
   active 3.A ATPases) carry an EC and reach GO via ec2go; use that as
   corroboration and to avoid duplicating an existing route.

Follow-Up Targets

Target	Rationale
Closure-aware (cross-branch) reverse gap	Separate true gaps from transport-BP/channel-CC coverage in the 4,431 reviewed no-MF entries; promote real gaps to gene reviews.
Subfamily-level tc2go from go.py	Rebuild the generated set at 4/5-level TC ids so poly-specific families resolve to substrate-specific MF children.
Exemplar gene reviews	Run the full review workflow on 2–3 confirmed-gap transporters already in this repo (353 candidates), mirroring the RHEA/UniPathway exemplar pattern.
"No transporter-activity MF at all" family set	The 367 TC families with no MF-activity term in go.py → candidates for proposed_new_terms or accessory (class 8/9) exclusion.
Propose a tc2go propagation pipeline to GO	The strategic deliverable: the 80 evidence-JUSTIFIED leads (mostly 5-level systems) are a ready external2go starter set; package them (plus reviewed extensions) so a protein's TC number yields a GAF annotation.
Promote the 67 GAP_CANDIDATE leads	Specific systems whose reviewed member carries only a parent term — run them through the gene-review workflow to add the specific GO term.
Chase the 10 NO_REVIEWED_MEMBER cases	Reviewed protein exists but lacks the DR TCDB xref (e.g. MelB) — a UniProt cross-reference gap to report upstream.
Extend GO's 63 xref'd families	Promote the 110 novel go.py-generated families (and the 16 GO-xref families go.py misses) toward curated term-xrefs.

Project Status

• Started: 2026-07-18
• Maturity: IN_PROGRESS — GO's neglected TC term-xrefs extracted (194 leads) and
  every one scored for propagation against UniProt evidence (80 JUSTIFIED / 67
  gap-candidate / 18 narrow / 12 class-level / 17 other); the missing piece identified as the annotation
  pipeline, not the mapping; TCDB's go.py dump characterised as a noisy second
  candidate pool; the structural reverse gap quantified; four SSSOM sets built and
  validated.
• Computed live (2026-07-18): GO curates 202 TC references on 170
  non-obsolete MF terms / 63 families (go-basic.obo); no tc2go in
  external2go (GO release 2026-05-19); go.py = 34,497 rows → 4,943 TC
  systems / 589 families / 3,518 GO terms, of which 13% rows / 12% terms are
  transporter-activity MF; reverse gap 4,431 / 8,758 (50.6%) reviewed TCDB entries
  with no transport MF term (16,344 all-UniProtKB TCDB xrefs); 353 local gene
  folders carry a TCDB xref; 173 / 2,235 TCDB families (8%) mapped across all
  three sources.
• Reproducible scripts: TCDB/extract_go_tc_xrefs.py
  (GO's own TC xrefs → SSSOM), TCDB/curate_propagation.py
  (propagation verdicts from UniProt evidence),
  TCDB/tcdb_go_probe.py (dump characterisation + reverse
  gap), TCDB/build_tc2go.py (generated SSSOM).
• Mapping sets (all pass just validate-tcdb-mappings):
  tc2go.from_go.sssom.yaml — 194 GO-xref source
  leads; tc2go.propagation.sssom.yaml — the
  same 194 scored for propagation (80 JUSTIFIED / 67 gap-candidate / 18 narrow / 12
  class-level / 17 other);
  tc2go.generated.sssom.yaml — 477
  machine-derived candidate rows; tc2go.sssom.yaml — 8
  hand-backed exemplars.
• Current conclusion: GO does reference TCDB — 170 molecular-function terms
  carry TC xrefs — but these are neglected, unreviewed sources: they cover only
  3% of TCDB families and, lacking an external2go pipeline, are never propagated
  to proteins, leaving half of reviewed TC-classified transporters with no
  transport MF term. The high-value deliverable is a tc2go pipeline built on
  per-entry propagation curation (which TC level safely inherits which GO term),
  seeded by GO's xrefs plus reviewed extensions.

Slides

• Slides (Marp source: TCDB-slides.md) — AI generated