TCDB -> GO Methodology Notes

TCDB → GO Methodology Notes

Parent project: TCDB.md

What exists, and what doesn't

Two separate things are easy to conflate:

  1. TC references on GO terms — these exist, but are neglected leads. GO
    carries TC references on molecular-function terms, as term-level xref: TC:
    clauses and as definition dbxrefs (def: "…" [TC:…]). These are largely
    neglected and the term-vs-definition distinction is not meaningful, so treat
    both alike, as sources. Extract them from go-basic.obo:

bash uv run python extract_go_tc_xrefs.py --stats # 202 TC references on 170 non-obsolete MF terms / 185 TC systems / 63 families uv run python extract_go_tc_xrefs.py -o tc2go.from_go.sssom.yaml

Every emitted row is skos:relatedMatch + semapv:UnspecifiedMatching — an
unreviewed lead. Whether it is safe to propagate (a protein with the TC id
inheriting the GO term) is scored next.

Propagation scoring (evidence, not assertion)

curate_propagation.py scores every one of the 194 leads
against UniProt member evidence — no hand-assertion:

uv run python curate_propagation.py        # ~185 UniProt calls, one per TC id
# -> data/propagation_evidence.tsv and tc2go.propagation.sssom.yaml

For each (TC, GO) lead it fetches the reviewed Swiss-Prot proteins carrying the
TC id and counts how many also carry the GO term (closure-expanded, computed
locally from the cached go-basic.obo is_a graph). TC matching is level-aware: a
5-level system is matched exactly; a coarser id by <id> OR <id>.* (the dot
stops 2.A.1 matching 2.A.10); placeholder segments (1.B.1.-.-) are stripped.
The fraction k/n drives the verdict, encoded as the SSSOM predicate:

These two are listed in evaluation order, and the order is load-bearing: k=0
also satisfies k/n<0.5, so GAP_CANDIDATE must be tested first or every specific
system with zero carrying members would be filed as narrow — inverting the reading
of the evidence and hiding the lead from the gap harvest.
- broadMatch CLASS_LEVEL — a whole TC class/subclass (level ≤2): broad by
construction, out of scope for substrate-level propagation. 12.
- relatedMatch NO_REVIEWED_MEMBER — n=0: only TrEMBL members, or the
reviewed protein lacks a DR TCDB xref (e.g. MelB); cannot assess. 10.
- relatedMatch UNCERTAIN — the ambiguous small-n middle. 7.

Two of these are direct feeds to the gene-review workflow: a JUSTIFIED lead
with k<n (e.g. 3.A.3.1.1 → GO:0005391, 4/5) and every GAP_CANDIDATE are
reverse-gap finds — the activity is real for the system but the member is missing
the (specific) term. Re-score without re-querying via
curate_propagation.py --reclassify (reads the cached evidence TSV). The
hand-curated tc2go.sssom.yaml is a deeper, individually read-through pass over a
handful of exemplars.

  1. An external2go annotation pipeline — this does not exist. Every other
    source-audit project here hangs off a public mapping file (ec2go, rhea2go,
    interpro2go, …); TCDB has none:

bash # ec2go, rhea2go, ... are present; tc2go is not curl -sL http://current.geneontology.org/ontology/external2go/ | grep -o 'external2go/[a-z0-9_]*' | sort -u curl -sIL http://current.geneontology.org/ontology/external2go/tc2go # no such file

So GO's curated term-xrefs are never propagated to the proteins that carry a
DR TCDB; cross-reference — there is no forward GO_REF volume (as there is for
RHEA's GO_REF:0000116). That reframes the project: the mapping partly exists,
but the pipeline is missing, so we (a) extract GO's existing xrefs, (b)
characterise the extra material TCDB publishes in go.py, and (c) measure the
structural protein-level gap that the missing pipeline leaves behind.

Direction 1 — What TCDB publishes: the go.py dump

TCDB serves its own per-protein GO associations as a flat TSV:

curl -sL https://tcdb.org/cgi-bin/projectv/public/go.py -o tcdb_go.tsv
# format:  GO-id <TAB> TC-number(5-level) <TAB> family-name
head -1 tcdb_go.tsv
# GO:0008076    1.A.1.1.1   The Voltage-gated Ion Channel (VIC) Superfamily

This is not a curated mapping — it is an aggregation of the GO terms found on
the member proteins of each TC system, so a single TC number carries a bag of GO
terms spanning all three aspects (a channel activity MF, its
plasma membrane/integral component of membrane CC, its transmembrane transport BP, and incidental protein binding). To turn it into a tc2go
candidate we keep only the transporter-activity molecular-function slice:

uv run python tcdb_go_probe.py --tcdb-go   # rows, distinct TC/GO, aspect skew, noise

The MF filter uses the is_a closure of the two transporter-activity roots,
fetched live from QuickGO (no local go-db needed):

curl -sL 'https://www.ebi.ac.uk/QuickGO/services/ontology/go/terms/GO:0022857/descendants?relations=is_a'
curl -sL 'https://www.ebi.ac.uk/QuickGO/services/ontology/go/terms/GO:0005215/descendants?relations=is_a'

Only ~13% of go.py rows survive this filter. build_tc2go.py
then aggregates the surviving 5-level TC numbers up to the TC family (3-level),
attaches the current QuickGO label, drops obsolete terms, and emits SSSOM:
exactMatch when a family has a single surviving MF term (mono-specific),
narrowMatch when it has several (poly-specific → the term is a subfamily
property). Mirrors GLYCOBIOLOGY/build_cazy2go.py.

uv run python build_tc2go.py --min-support 2 -o tc2go.generated.sssom.yaml

Direction 2 — Reverse gap ("falls through the cracks")

A UniProtKB entry can carry a TCDB cross-reference and yet have no GO
transporter-activity term. With no tc2go, this is the expected steady state,
and its size is the argument for building one. Measured directly from the UniProt
REST API — no go-db required:

# reviewed entries carrying any TCDB xref
curl -sIL 'https://rest.uniprot.org/uniprotkb/search?query=(database:tcdb)+AND+(reviewed:true)&size=0' | grep -i x-total-results
# ...of those, how many lack a transporter-activity MF term (GO:0005215 closure)
curl -sIL 'https://rest.uniprot.org/uniprotkb/search?query=(database:tcdb)+AND+(reviewed:true)+NOT+(go:0005215)&size=0' | grep -i x-total-results

tcdb_go_probe.py --gap automates this. UniProt's go:<id> query field expands
to descendants, so these counts are already ontology-closure-filtered on the GO
side (verified: go:0005215 root returns more entries than the mid-level child
go:0015075).

Why an entry falls through the cracks

  1. No pipeline. The dominant cause here, unlike RHEA. A transporter with no
    EC (most secondary transporters) and no informative InterPro2GO term has no
    automated route to a transport MF term.
  2. Specificity mismatch (confounder — not a true gap). The entry may carry a
    more specific transport MF child, or a transport BP term
    (transmembrane transport) / channel CC term, none of which the exact
    GO:0005215-MF-closure count credits.
  3. Family poly-specificity. Even a hypothetical family-level tc2go would
    often supply only a generic parent; the substrate-specific child needs the
    4/5-level TC id.

The closure caveat is mandatory

The gap headline in TCDB.md is exact-match against the
GO:0005215 MF branch, so pct_missing is an upper bound. The honest next
step is a cross-branch closure filter: subtract entries carrying any transport MF
descendant, any transport-process BP term, or any transporter/channel-complex CC
term before calling an annotation a true gap. This is the identical lesson the
RHEA, CAZy, and SPKW projects learned: naive counts overstate the signal.

Reproducibility & limitations