NCBI RefSeq vs UniProt/GOA: prokaryotic GO pipeline comparison

Bottom line: since 2022, NCBI's Prokaryotic Genome Annotation Pipeline (PGAP)
has put GO terms on RefSeq prokaryotic proteins, and per-assembly GAF files now
sit on the NCBI FTP site. GO does not ingest them, and nobody had compared them
with GOA. The two pipelines differ in kind. NCBI is winner-take-all: GO is an
attribute of the single protein family model that names the protein, so every
protein gets terms from exactly one evidence source. UniProt/GOA is a union of
InterPro2GO, TreeGrafter, UniRule, ARBA, EC/Rhea2GO, SubCell and others. On
Pseudomonas putida KT2440, scored against 956 curated gene reviews, NCBI is
about as accurate and much less complete. Wrong terms (REMOVE) are rare on both
sides: 3.2% of the NCBI terms reviewers could judge, against 4.8% of GOA's
electronic terms, and NCBI has fewer true-but-too-general terms (5.9% against
11.6%). But NCBI covers only half of the reviewed core molecular functions (52%,
against 81%) and about a third of core processes (37%, against 67%), and it has almost no cellular-component
annotation. The biggest NCBI-specific source is CDD domain architectures
(SPARCLE), not the CDD cd models that InterPro integrates. They name 47% of
GO-bearing proteins and are the least concordant source. Each pipeline has its
own failure modes: GOA transfers functions between paralogs, NCBI's specific
equivalogs avoid that, and NCBI has wrong GO on a few named models (listed
below) that propagates to every genome.

This page also carries the history of GO–NCBI contact on prokaryotic GO, which
was reconstructed from email when the topic resurfaced in October 2026.

Relationship to the NCBIFam project

NCBIFam measures how NCBIFAM/CDD signatures reach GOA through
InterPro2GO, and proposes an ncbifam2go built from hmm_PGAP.tsv. This project
compares the outputs: the GO that NCBI actually assigns to proteins against the
GO that GOA assigns, gene by gene.

One correction to that project's framing: NCBIFam.md says that "CDD-proper has
no native GO". That is true of the cd models, but NCBI also keeps GO as an
attribute of CDD domain architectures (SPARCLE, ArchId: in the GAF
WITH/FROM column), and has propagated it to proteins since March 2023. This GO
is not in cddannot*.dat or hmm_PGAP.tsv, so the NCBIFam probe could not see
it. In the KT2440 GAF it is the largest single source.

How the two pipelines assign GO

The winner-take-all rule is the main structural difference. NCBI's design
favours precision: a protein named from a specific equivalog HMM inherits only
that HMM's curated terms, and broad domain terms from other hits are not added.
GOA favours recall, and stacks domain-level InterPro2GO terms with family- and
rule-level terms.

The winner-take-all rule applies to prokaryotes only. NCBI annotates eukaryotic
RefSeq with a separate pipeline (since 2023): InterProScan with InterPro2GO and
PANTHER/TreeGrafter GO, keyed to NCBI GeneIDs. That pipeline is much closer to
UniProt's and is out of scope here.

Pilot: P. putida KT2440 (PSEPK)

Script: compare_refseq_goa.py. Two runs are
kept:

• results/PSEPK/: NCBI GO from the
  live RefSeq genome record (annotation RS_2026_02_04, PGAP 6.10), with each
  protein's naming model taken from its WP_ record. This is the default.
• results/PSEPK_ftp_RS_2025_02_17/:
  NCBI GO from the downloadable GAF (annotation RS_2025_02_17, generated
  2025-06-12, GO version 2025-03-16). The detailed numbers further down this page
  come from this run; Live record vs FTP gives the 2026
  headline numbers.

Both runs use GOA for taxon 160488 from QuickGO and UniProt proteome
UP000000556, all fetched on 2026-10-18.

uv run python projects/REFSEQ_GO/compare_refseq_goa.py \
  --accession GCF_000007565 --taxon 160488 --proteome UP000000556 \
  --locus-regex 'PP_\d+' --out-dir projects/REFSEQ_GO/results/PSEPK
# add --ncbi-source ftp to use the per-assembly GAF instead of the live record

The script resolves the accession through the NCBI Datasets v2 REST API and
rejects a superseded accession. NCBI's Python client (ncbi-datasets-pylib) was
retired at the end of 2024, so the REST API is called directly. Datasets genome
packages have no GO option. GO therefore comes either from the GenBank record's
/GO_function, /GO_process and /GO_component qualifiers, fetched with
E-utilities, or from the FTP GAF. The genome record does not say which model
named each protein, so that is read from the Evidence-For-Name-Assignment block
of the WP_ protein records. The annotation release and fetch times are recorded
in summary.yaml.

Identifiers and annotation releases

• No NCBI Gene records. All 6,027 KT2440 Gene records were discontinued on
  2020-04-02; one remains live. NCBI now provides Gene records only for
  prokaryotic reference genomes
  (RefSeq prokaryote FAQ),
  and in February 2020 it reduced Reference assemblies to the 15 with
  expert-provided annotation. E. coli K-12 (4,652 live records), B. subtilis
  168 (4,537) and P. aeruginosa PAO1 (5,698) keep theirs.
• What remains is the RefSeq genome record. Genes are features on
  NC_002947.4 with a locus tag (PP_RS00285) and the original TIGR tag as
  old_locus_tag (PP_0055). The protein is a non-redundant WP_ accession,
  shared by every genome that encodes the identical sequence. The earlier
  genome-specific NP_/YP_ proteins were suppressed in NCBI's one-time
  re-annotation project, which for KT2440 replaced the record on 2016-03-04
  (re-annotation project).
• Re-annotation continues. Since then the genome has been re-annotated
  repeatedly with newer PGAP versions. The live record carries
  GCF_000007565.2-RS_2026_02_04 (PGAP 6.10). Locus tags were kept:
  PP_RS00285/PP_0055 is unchanged.
• The downloads lag the live record. The Datasets API and the FTP files
  (GAF, feature table, GBFF; all dated 2025-06-12) still serve
  RS_2025_02_17 (PGAP 6.9). The live record does carry GO, as CDS
  qualifiers, so the script now reads GO from it by default
  (--ncbi-source live).

Live record vs FTP

Between the February 2025 annotation (FTP) and the February 2026 annotation
(live record), 742 of 5,483 shared genes (13.5%) changed GO: 934 terms were
added and 180 removed, 134 genes gained GO for the first time, and 316 product
names changed. Locus tags did not change. The conclusions hold on the newer
annotation:

Every model-level error listed under Failure modes is still
present in the live record.

Gene models differ, and the obvious join key is lossy

Genes are joined on the original PP_xxxx locus tag: from the RefSeq gene
old_locus_tag, and from UniProt's ordered locus name.

• 5,487 RefSeq protein-coding loci (plus 132 pseudogenes) against 5,563
  locus-tagged UniProt entries; 5,390 are shared.
• 173 loci are UniProt-only. Of these, 39 are RefSeq pseudogenes and 134
  are absent from the current RefSeq annotation altogether. These are original
  2002 gene calls that the PGAP re-annotation dropped or merged. 97 loci are
  RefSeq-only.
• 1,007 UniProt entries (18%) have no RefSeq WP_ cross-reference, even
  though most of their loci have a WP_ protein. A join on WP_ cross-references
  would silently lose almost a fifth of the proteome, so the locus-tag join is
  needed.
• 20 WP_ proteins are encoded at several loci (identical paralogs).

Coverage

GOA has about twice as many terms per gene. This is the expected result of a
union against a single source.

Agreement by NCBI evidence source

Each NCBI term was compared with the GOA terms for the same locus and aspect,
closure-aware over is_a and part_of. Only locus–aspect pairs annotated on
both sides are counted here; the rest appear in the TSV as
other_side_no_aspect.

Most NCBI-only terms from architectures are broad ligand-binding or
transport terms, such as metal ion binding (91), ATP binding (45),
ABC-type transporter activity (31) and monoatomic ion transport (30). This
fits domain-architecture-level GO. Those terms have not yet been checked against
curated reviews.

From the GOA side, the sources least represented in NCBI are UniRule (46% of
compared terms have no NCBI counterpart), TreeGrafter (42%) and ARBA (31%).
InterPro2GO is 17%. The top GOA-only terms are UniRule's DNA-templated transcription (155 loci) and generic DNA binding and transmembrane transport.

Experimental GOA annotations (EXP/IDA/IMP/IPI/IEP, 147 rows) are mostly
uncovered by NCBI. In aspects where both sides annotate, 16 match exactly, 2 are
more specific in NCBI, 17 are more general in NCBI, and 40 have no NCBI
counterpart.

Staleness

The KT2440 GAF was generated in June 2025 against GO 2025-03-16, and has not
been refreshed since. 49 of its 1,626 distinct terms (3%) are obsolete in the
current GO. Most belong to the 2025–2026 obsoletions of "via X" pathway
variants and amino-acid metabolism, e.g. GO:0009089, GO:0006542 and
GO:0051082. The comparison counts these as NCBI-only.

Comparison results (KT2440)

Script: analyze_refseq_goa.py. The numbers in
this section are from the February 2025 (FTP) run; the same tables for the live
run are in results/PSEPK/. Generated tables:
analysis.yaml,
ncbi_review_cases.tsv,
goa_errors_vs_ncbi.tsv and
core_coverage.tsv. Cases checked by hand, and
re-checked against the live record:
failure_cases.yaml.

uv run python projects/REFSEQ_GO/analyze_refseq_goa.py \
  --results-dir projects/REFSEQ_GO/results/PSEPK --species PSEPK

Level 1: gross differences across the genome

• Cellular component is the largest gap. At goslim_prokaryote level, GOA
  places 981 loci in cytoplasm and 835 in plasma membrane; NCBI places 99 and
• GOA gets these from SubCell2GO, ARBA and the combined pipeline, which have
  no NCBI equivalent.
• Generic transcription terms. DNA-templated transcription is on 348 loci
  in GOA and 28 in NCBI, mostly UniRule adding it to transcription factors.
  Only 4 UniRule rows of this term fall in reviewed genes (2 marked
  over-annotated, 1 removed, 1 modified), which is too few to judge.
• NCBI has more ATP-dependent activity (274 loci against 214). This is the
  only slim class where NCBI leads.
• 322 genes have NCBI GO and no GOA GO. 213 of them are named by an
  architecture: transposases, Csu fimbrial proteins, Hcp T6SS effectors,
  MvaT/MvaU nucleoid proteins, GSTs, glycosyltransferases. This is NCBI's real
  coverage gain.
• 780 genes have GOA GO and no NCBI GO. Their RefSeq names show why. The top
  name is "LysR substrate-binding domain-containing protein" (22). When the
  naming model is a domain model with no GO, the protein gets nothing, even if
  its other hits (here a LysR HTH domain, which InterPro2GO maps to
  DNA-binding transcription factor activity) carry GO. This recall failure is
  built into winner-take-all.

Level 2: by source

From the NCBI side, agreement falls in this order: BlastRules (100% exact),
NCBIFam (90%), TIGRFAM (79%), then architectures (69%; 18% NCBI-only); see
Agreement by NCBI evidence source. From the
GOA side (with combined GO_REF:0000120 rows credited to each contributing
pipeline), the terms NCBI least often supports are GO inter-ontology links (55%
have no NCBI counterpart), TreeGrafter BP (35%) and UniRule BP (33%). InterPro2GO
MF terms more often have a more specific NCBI counterpart (19%) than none
(14%).

Level 3: against 956 curated reviews

Reviewers judged GOA rows, not NCBI rows. So an NCBI term shared with GOA
inherits the reviewer's verdict, and an NCBI term absent from GOA is judged only
by its relation to the accepted and removed terms. Terms with no such relation
are unassessed, and these were checked by hand (below).

"Correct" adds terms that are proper ancestors or descendants of an accepted
term. "Too general" is a term the reviewer marked over-annotated that is an
ancestor of a term the reviewer accepted: true, just less specific. Rows do not
sum to 100% because undecided (1–3%) is omitted. GOA terms agreed by several
pipelines (GO_REF:0000120) count for each of them. Among NCBI's HMM sources,
NCBI's own NCBIfam models have more wrong (5.1% against 2.1%) and more
too-general terms (6.5% against 2.9%) than TIGRFAM, mostly broad cofactor and
category terms (carbohydrate metabolic process, membrane, metal ion binding, ATP binding) attached to otherwise specific models.

Core-function coverage (terms in the reviews' core_functions):

NCBI covers a core function that GOA's electronic pipelines miss in only 7 MF
and 31 BP cases. The reverse happens in 211 MF and 266 BP cases.

GOA errors and NCBI. Of the 115 InterPro2GO terms that reviewers removed, NCBI
asserts the same term in 66. That much overlap suggests the two share
provenance for part of their family→GO mappings. Of the 70 TreeGrafter terms
removed, NCBI asserts one.

Failure modes

Every case below is recorded in
failure_cases.yaml.

Errors only NCBI makes. These come from the 224 unassessed NCBI terms. 64 of
those are obsolete terms (GO-version lag), not mapping errors. Among the rest,
the following are clear errors. Where marked (model), the GO is listed on the
HMM itself in hmm_PGAP.tsv, so one fix corrects every genome.

• Wrong function on the model: SecG → sialic acid transport (model, TIGR00810);
  QueF → a tRNA-ribosyltransferase activity (model, TIGR03138); Lgt → protein
  lipoylation (model, TIGR00544); TatC → type II secretion (model, TIGR00945);
  MlaC-family Ttg2D → ABC-type transporter activity.
• Model name and GO disagree: TIGR01162 is named the N5-CAIR mutase but carries
  the carboxylase activity (model).
• A substrate given an activity: precursor peptide PqqA → catalytic activity
  (model, TIGR02107).
• Pathway context ignored: Aacs → mevalonate pathway (model, TIGR01217), a
  pathway KT2440 lacks.
• Complex activity given to one subunit: SdhC and SdhD → succinate dehydrogenase
  activity (model); YajC → protein-transporting ATPase (model); PP_3384
  (gluconate 2-dehydrogenase gamma subunit) → an NADP+ gluconate dehydrogenase
  activity.
• Architecture-level transfer (the CDD-specific mode): ImuB → DNA polymerase
  activity, though it is catalytically inactive; GlgE → oligosaccharide
  catabolic process, though it is biosynthetic; a saccharopine-dehydrogenase-family
  carboxyspermidine dehydrogenase → lysine biosynthesis; MoaA paralogs →
  molybdopterin synthase complex; OprD's basic-amino-acid substrate copied to
  OprE, NicP and GalP.

Errors only GOA makes, which NCBI's specific equivalogs avoided. Paralog
transfers dominate. TreeGrafter: Mqo1–3 → (S)-2-hydroxyglutarate dehydrogenase,
DapE → acetylornithine deacetylase, MetX → the acetyl- rather than
succinyltransferase activity, KdsC → a cytidylyltransferase, PhnX →
phosphoglycolate phosphatase, Gcl → valine biosynthesis. InterPro2GO: GlyS →
arginine-tRNA ligase, ThiI → CCA-adding activity, PurC → cobalamin
biosynthesis.

Errors both make (66 reviewer-removed terms that NCBI also asserts):
- Complex activity given to one subunit: the DNA polymerase III β, ε and τ subunits
given polymerase activity; EutC, CarA, NirD and SoxG.
- Wrong cofactor: NAD binding on the NADP-specific Icd; 2Fe-2S binding on PrpD,
which has no cluster; ATP binding on DnaJ and GroES.
- Paralog substrate: PssA given PgsA's reaction; NH3-dependent NadE given glutamine
activities.
- Reversed direction: GpsA as catabolic, Dfp as pantothenate catabolism.

Naming. Sometimes the name is the error. NCBI names PycB a sodium-extruding
oxaloacetate decarboxylase, where the review assigns pyruvate carboxylase. In the
other direction, UniProt still carries the 2002 submitter name for PP_3099 (puuD,
"uricase"), while NCBI correctly names it the T6SS sheath protein TssC.

Worked case: the ParAB chromosome-partition proteins

ParA and ParB were reviewed in P. putida KT2440 and in P. aeruginosa PAO1,
where both proteins have been characterized experimentally. Each review has a
companion PGAP prediction review that scores NCBI's GO term by term. The
orthologs are 84% (ParA) and 79% (ParB) identical, with identical Walker A
and CTP-binding motifs
(alignment), so any
difference between the two species comes from the pipelines, not the
biology.

Prediction reviews:
PP_0002,
PA5563,
PP_0001,
PA5562.

What the case shows:

• An architecture can fill a gap that GOA leaves completely open. GOA has no
  annotation of any kind for either ParA. The Pfam domain (AAA_31) has no
  InterPro2GO mapping and the PANTHER subfamily brings no IBA. In P. putida,
  UniProt still calls the Swiss-Prot entry "Uncharacterized protein PP_0002".
  NCBI's architecture GO for PP_0002 is correct, and it is the only GO this
  protein has anywhere.
• The same architecture is not used for the ortholog. The P. aeruginosa
  ParA is named by a BlastRule instead, which gives it a better name and no GO.
  Under winner-take-all the architecture's terms are then not passed on, so a
  well-studied ATPase ends up with no GO from either pipeline. Adding ATP
  hydrolysis activity (GO:0016887) and chromosome segregation (GO:0007059) to
  NBR012933 would reach the 53,716 assemblies that carry WP_003097243.1
  (NCBI identical-protein record, 2026-10-24).
• The naming model is not the most informative model. For both ParBs, the
  architecture passes on only DNA binding. TIGR00180 also matches, has the
  same product name, and carries chromosome segregation (plus a cytokinesis
  term that is wrong), but it does not pass on its GO because it did not name
  the protein. The obvious fix, adding GO:0007059 to architecture 11445026,
  would be wrong as things stand. The architecture is not specific to
  chromosomal ParB: besides WP_003100240.1 (39,835 assemblies, 2026-10-24) it
  also names plasmid partition proteins, among them F-plasmid YubM
  (WP_000117173.1), a Thermus thermophilus plasmid ParB (WP_011174478.1) and
  a Xylella fastidiosa plasmid ParB (WP_010895251.1). Chromosome segregation
  is wrong for those, which is why the NCBIFam
  seed maps it to TIGR00180 only as a sub-clade term. Some relatives are
  protected by more specific naming models (Noc by TIGR04285, SopB by NF010252,
  KorB by architecture 11488667), but not all of them. The process can only be
  added safely if NCBI first separates chromosomal from plasmid ParBs, for
  example with a chromosomal ParB equivalog that outranks the architecture.
• Neither pipeline annotates these ParBs as CTPases. The CDD region note
  still describes the N-terminal domain as a "ParB-like nuclease domain". The
  ParB CTPase activity (GO:0043273), measured directly on P. aeruginosa ParB
  (PMID:38843873), is absent for both proteins in NCBI and in GOA, and no
  matched family model carries it. GOA does have the term, by experiment, on
  Myxococcus xanthus ParB (Q1CVJ4; PMID:31835030, PMID:34562373), the only
  one of the 28,687 TIGR00180 proteins that carries it. B. subtilis Spo0J
  has it as a core function in this repo's review.
  These are the obvious seeds for a PAINT inference at the ParB family node.
• NCBI avoids two GOA errors. GOA gives both ParBs positive regulation of sporulation from the B. subtilis Spo0J-seeded PANTHER node (IBA in PAO1,
  TreeGrafter in KT2440), although neither species sporulates. It also gives
  the PAO1 ParB DNA-templated transcription from a UniRule. The reviews
  remove all three of these rows, and NCBI asserts none of them.

Second genome: E. coli K-12, via WP_ records

E. coli K-12 MG1655 is one of the 15 curated references, so PGAP never runs on
its assembly. PGAP's GO is attached to non-redundant WP_ proteins, though, and
almost every MG1655 protein is identical to a WP_ protein that PGAP annotates in
other genomes. The compare script's --ncbi-source wp mode follows each UniProt
entry's WP_ cross-reference and reads GO and the naming model from the WP_
GenPept record. It compares proteins, not gene models.

uv run python projects/REFSEQ_GO/compare_refseq_goa.py \
  --accession GCF_000005845 --taxon 83333 --proteome UP000000625 \
  --locus-regex 'b\d{4}' --ncbi-source wp --out-dir projects/REFSEQ_GO/results/ECOLI_wp
uv run python projects/REFSEQ_GO/analyze_refseq_goa.py \
  --results-dir projects/REFSEQ_GO/results/ECOLI_wp --species ECOLI

The WP_ records are the freshest PGAP source. Run in wp mode on KT2440
(results/PSEPK_wp/), the 4,504
proteins reachable both ways have the same naming model in all cases and the
same GO in 4,247 (94%) when compared with the live February 2026 genome record.
Of the 257 differences, 184 are on WP_ records modified during 2026, and the
examples are obsolete terms the genome record still carries but the WP_ record
has replaced (GO:0019557 → GO:0006548, GO:0008785 → GO:0102039). WP_ records are
dated up to 2026-10-16. The cost is coverage: WP_ cross-references are missing
from 18% of KT2440 UniProt entries but only 7% of K-12 entries (4,112 of 4,403
have one).

Results against experimental GO. K-12 has experimental GO on 3,328 of the
4,123 compared genes (results/ECOLI_wp/).
Experimental annotation is incomplete, so "no experimental support" means
unconfirmed rather than wrong.

This repeats the KT2440 pattern: similar precision, much lower recall for NCBI.
The comparison favours GOA, whose rules and mappings are partly built from these
same experimental annotations (IBA is excluded from "GOA electronic" for that
reason). On KT2440 the experimental set is small (57 genes) but gives the same
picture: 13.6% recall for NCBI against 45.6% for GOA electronic.

BlastRules name well-studied proteins and give them no GO. Only 53% of the
K-12 WP_ records carry any GO, against 68% for KT2440. The difference is
BlastRules: they name 771 K-12 proteins (19%), and only 13 of those (2%) carry
GO. Examples are diacylglycerol kinase (NBR011105), 5'-deoxynucleotidase
(NBR007780) and the As(III)-sensing ArsR regulator (NBR009015). BlastRules are
anchored on characterised proteins, so they win the naming for exactly the
best-studied genes, and winner-take-all then discards the GO the HMMs would have
given. In KT2440, BlastRules name 92 proteins, 22% with GO.

A clean case is EntE (P10378, WP_000026812.1), named by BlastRule NBR012093
with no GO. InterPro shows it also matches two equivalog HMMs with the same name,
NF008192 (residues 2–534) and TIGR02275 (3–527), both carrying GO:0008668 and
GO:0019290. GOA has GO:0008668 from five EXP and one IDA annotation. In a sample
of 60 BlastRule-named K-12 proteins without GO, 25 matched a GO-bearing
NCBIfam/TIGRFAM model (InterPro's matches; PGAP does not publish its own hit
lists).

Third genome: Synechocystis PCC 6803

A cyanobacterium, so a third phylum. UniProt's reference proteome (UP000001425)
and RefSeq GCF_000009725.1 are the same assembly, so this is a same-genome
comparison like KT2440. There are no reviews in this repo, so experimental GO
(209 genes) is the only ground truth.

uv run python projects/REFSEQ_GO/compare_refseq_goa.py \
  --accession GCF_000009725 --taxon 1111708 --proteome UP000001425 \
  --locus-regex 's(?:ml|[lms][lr])[0-9]{4}[a-z]?' --out-dir projects/REFSEQ_GO/results/SYNY3

Joining needed a second key. 345 UniProt entries carry only a submitter
gene name (e.g. two different entries named bcp) and no ordered locus name,
and almost no entries have a WP_ cross-reference. The script now joins such
entries by identical sequence to a RefSeq translation, which recovered 282 of
them; 63 have no identical RefSeq protein, which suggests different gene calls.
The live record (RS_2026_01_07, PGAP 6.10) is again a release ahead of
Datasets (RS_2025_06_29).

The gap is wider than in KT2440 or K-12. Partly this is because the
experimental set leans on location (255 of 658 rows are CC, mostly thylakoid
membrane), where NCBI assigns little. But NCBI also recovers only 10% of the
experimental MF terms, against 54% for GOA. Examples:

• The naming rule again. Gap2 (sll1342) is named by NF005641, an
  equivalog with no GO; it also matches TIGR01534, with the same name and
  GAPDH GO, which is discarded.
• BlastRule names, no GO. sll0418 (2-methyl-6-phytyl-1,4-hydroquinone
  methyltransferase, NBR012634) and sll1253 (A-adding tRNA
  nucleotidyltransferase, NBR015769) get specific names and no MF.
• No naming model at all. sll0554, ferredoxin:thioredoxin reductase, has no
  naming evidence and no GO; GOA gets the activity from EC2GO/Rhea2GO.
• Specific name, general GO. sll0228 is named N(1)-aminopropylagmatine
  ureohydrolase (BlastRule) but carries only "agmatinase activity"; GOA has the
  exact activity from EC2GO/Rhea2GO.
• Lineage-specific biology. The carboxysome shell proteins (sll1028–1031)
  lack "structural constituent of carboxysome shell" in NCBI, which GOA gets
  from InterPro2GO and UniRule.

The genome record lags the WP_ records. CcmL (sll1030) is named by
NF053795, which carries carboxysome GO, and its WP_ record (updated 2026-04-14)
has that GO; the January 2026 genome record does not. Across all proteins, the
genome record and the WP_ records agree on GO for 92.9% (Synechocystis) and
94.5% (KT2440), with the WP_ records generally the newer.

Counterfactual: NCBI without winner-take-all

What would NCBI's GO look like if a protein kept the GO of every NCBI model it
matches, not only the naming model? compare_refseq_goa.py --union-models
adds, to NCBI's actual GO, the GO of every matched HMM in hmm_PGAP.tsv:

• family: NCBIfam and TIGRFAM family models;
• all: family models plus NCBI's copies of Pfam models.

Matches are InterPro's, from each UniProt entry's NCBIfam and Pfam
cross-references, because PGAP does not publish its hit lists. CDD/SPARCLE
architectures cannot be added: only the winning architecture is visible.

Scored with the corrected verdict split (see the next section), on KT2440:

In E. coli K-12, experimental-term recall goes from 16% to 21% (family) and
23% (all). Results: results/{PSEPK,ECOLI_wp}_union_{family,all}/.

• Adding family models is close to free. Recall rises on both genomes with
  almost no change in wrong or too-general terms.
• Pfam copies mostly add true but general terms. Of the 661 Pfam-derived
  terms KT2440 reviewers judged, 35% were marked over-annotated, nearly all for
  being less specific than an accepted term ("catalytic activity",
  "oxidoreductase activity", "membrane"); only 4% were REMOVEd. The REMOVEd ones
  are real errors shared with InterPro2GO, e.g. arginine-tRNA ligase activity
  on glycyl-tRNA synthetase. (An earlier version of this page reported "38%
  rejected", lumping the two.)
• It is not mainly CDD/SPARCLE. With family, the KT2440 genes that gain
  terms were named by NCBIfam HMMs (199), architectures (183), TIGRFAM (110),
  or had no NCBI GO at all (130). In K-12, most were genes with no NCBI GO
  (496), largely BlastRule-named, or NCBIfam-named (378); architectures
  accounted for 23.
• Most of the gap to GOA remains; the next section shows where it comes
  from.

Pipeline variants scored the same way

analyze_refseq_goa.py now scores every annotation set identically
(pipeline_variants in analysis.yaml), and splits the reviewer verdicts:

• wrong: REMOVE.
• too general: marked over-annotated, but an ancestor of a term the
  reviewer accepted. True, just less specific, and not an error.
• over-annotated (other): the rest of MARK_AS_OVER_ANNOTATED.

GOA's combined rows (GO_REF:0000120) are credited to each pipeline in their
WITH/FROM, so pipeline subsets can be scored.

KT2440 (956 reviewed genes, 57 with experimental GO):

E. coli K-12 (3,328 genes with experimental GO; only 83 reviewed genes, so
the review columns are noisy):

What this changes:

• Wrong terms are rare on every side (3–5% REMOVE). The earlier "12% vs
  21% rejected" was mostly true-but-general terms, concentrated in InterPro2GO
  (17% too general).
• NCBI's coverage is about InterPro2GO's. Core MF 52% vs 54%, core BP 37%
  vs 37%, at a lower too-general rate.
• ARBA adds little. Removing it moves core MF from 81% to 80%.
• GOA's lead comes from enzyme mappings and rules. Of 268 core terms that
  only GOA's other pipelines reach (beyond NCBI and InterPro2GO + TreeGrafter +
  IBA), the supplying pipelines are EC2GO (98), UniRule (95), Rhea2GO (71),
  UniPathway2GO (58), inter-ontology links (36) and ARBA (32); a term can have
  several. Examples: BetC choline-sulfatase activity (EC2GO), DavD
  glutarate-semialdehyde dehydrogenase (NADP+) (Rhea2GO), HisH glutaminase
  activity (EC2GO, Rhea2GO).
• NCBI already has EC numbers it does not turn into GO. 1,592 KT2440 CDS
  carry /EC_number, 148 of them with no GO molecular function. An ec2go step
  on NCBI's own EC numbers is the obvious next counterfactual.
• In K-12, IBA is circular. PAINT uses E. coli experimental annotations
  as evidence, so IBA's high experimental recall there is expected.

Examples of core functions NCBI misses and the family union recovers: RpoD
(named by NF004208, no DNA binding; TIGRFAM adds DNA binding and transcription
initiation), CbcX (named by a CDD architecture; TIGRFAM adds choline binding and
transport), DegP (named by NF008189; TIGRFAM adds serine-type endopeptidase
activity), FtsI (named by a BlastRule; NCBIfam adds penicillin binding).

Where NCBI's prokaryotic GO can be downloaded

NCBI's GO for KT2440 is in two places: the CDS features of the RefSeq genome
record, and the per-assembly GAF on the FTP site (a release behind). It is not in
gene2go, the GO file most tools take from NCBI. That file is keyed by NCBI Gene
ID, and KT2440 has no live Gene records.

Counts are from gene2go.gz dated 2026-10-18.

What "reference" means for GO. The reference genomes' RefSeq records take
their annotation from an outside group, not from PGAP, and carry no GO
qualifiers at all (0 in NC_000913.3 and NC_002516.2):

• MG1655: "annotation updates are derived from EcoCyc", from a collaboration
  of EcoCyc, the University of Wisconsin, UniProtKB/Swiss-Prot and NCBI.
• PAO1: "the Sept 08, 2010 version" of the PseudoCAP annotation from
  pseudomonas.com.

Their gene2go rows instead match the GO Consortium per-species files, which
UniProt generates (ECOLI-uniprot, PSEAE-uniprot, July 2026). Collapsed to
one row per protein, term and evidence code, these give 43,493 (E. coli) and
26,077 (PAO1) rows, against 41,101 and 22,779 in gene2go. The gap is about the
size expected from proteins without an NCBI Gene ID (3,863 vs 4,482 and 3,890
vs 4,388). So for reference genomes, gene2go carries UniProt-GOA's mix:
experimental, IBA and UniProt's electronic pipelines. For PAO1, PseudoCAP's own
curation is one contributor inside that file (2,066 of 32,020 rows,
assigned_by=PseudoCAP).

NCBI Gene pages show the same data, under a misleading label. The GO
section of a reference-genome Gene record matches the GO Consortium file
exactly: dnaA has 36 of 36 term–evidence pairs in common for E. coli
(GeneID 948217) and 11 of 11 for PAO1 (GeneID 878417). Each section carries one
attribution, "Provided by EcoCyc" or "Provided by PseudoCAP". For PAO1 dnaA,
none of the 11 rows were made by PseudoCAP (they are assigned_by GO_Central,
InterPro and UniProt), so the label names the organism's community resource,
not the source of the annotations. Only one gene per genome was checked.

gene2go has no source column, and its IEA rows have no PubMed ID (all 14,184
K-12 and 13,192 PAO1 IEA rows show -), so provenance cannot be read off it
directly.

PGAP's GO does not reach Gene or gene2go, even where Gene records exist.
"Reference genome" now means two sets. Since September 2024 it is the best
genome per species, about 23,700 prokaryotes (Datasets, 2026-10-18; the
May 2025 collection
had 21,794), nearly all PGAP-annotated
(terminology change).
The 15 "manually curated" references
(list) are the
submitter-annotated subset. Gene IDs go to reference genomes of species with at
least ten RefSeq assemblies, so PGAP genomes can have Gene records. The P. putida
species reference, NBRC 14164 (GCF_000412675.1, PGAP), has 5,573 live Gene
records, yet 0 gene2go rows and no GO on its Gene pages (checked for dnaA,
GeneID 45521526). gene2go and Gene's GO sections appear to carry GO Consortium
data only. Tools that take GO from NCBI through gene2go therefore never see
NCBI's prokaryotic pipeline annotation.

Unexplained: B. subtilis 168 has live Gene records, and the GO Consortium
has a BACSU-uniprot file with 21,806 rows for taxon 224308, but none reach
gene2go.

History of GO–NCBI contact on prokaryotic GO

Reconstructed from email and calendar records (C. Mungall's mailbox) and the two
public NCBI Insights posts.

Next steps

• [x] Per-review audit against the 956 PSEPK reviews (see
  Level 3).
• [ ] Send NCBI the model-level GO fixes marked (model) above, with the review
  evidence for each.
• [ ] Add to that list the ParAB gaps from the worked case:
  ATP hydrolysis activity and chromosome segregation on BlastRule NBR012933;
  separating chromosomal from plasmid ParBs before any chromosome-segregation
  GO is added to architecture 11445026; CTPase activity for ParB models; and
  the "nuclease" wording of the COG1475 region note. Separately, ask PAINT to
  consider a CTPase IBD at the ParB node, seeded by M. xanthus ParB.
• [ ] Score a random sample of the 160 non-obsolete unassessed NCBI terms, so the
  NCBI-only error rate can be estimated rather than just shown by example.
• [ ] Recompute obsolete NCBI terms with their GO replacements, to separate
  version lag from real disagreement.
• [ ] Find out why B. subtilis 168 has no gene2go rows despite live Gene
  records.
• [ ] Ask NCBI why the FTP files and Datasets lag the live record by an
  annotation release.
• [x] Second genome: E. coli K-12 via WP_ records, scored against
  experimental GO.
• [x] A same-assembly genome from another phylum: Synechocystis PCC 6803.
• [ ] Take up NCBI's offer of a one-off dump of all GO-bearing models
  (BlastRules and architectures as well as HMMs). Without it, architecture
  GO can only be reverse-engineered from per-genome GAFs.
• [ ] Draft text for GO's download documentation (go-site) that points to the
  NCBI per-assembly GAFs and states their scope: IEA only, one source per
  protein, no GAF for the curated reference genomes, possible GO-version lag.
• [ ] Topics for a renewed NCBI conversation: GAFs for the curated reference
  genomes; an obsoletion-notification route for NCBI model→GO mappings; a
  GO_REF per NCBI source type instead of a single PMID; and SPARCLE GO
  provenance.