NCBIFAM and CDD as GO sources

What the NCBI family models add, what InterPro hides, and a proposed ncbifam2go

AI Gene Review · projects/NCBIFam · 2026

Bottom line

  • NCBIFAM/CDD reach GO only via InterPro2GO; 60% of NCBIFAM models are unintegrated and contribute nothing, and GOA never names the member that fired.
  • NCBIFAM backs 705 (13%) of this repo's 5,549 InterPro2GO rows (sole signature for 250). NCBI's own metadata gives GO to 11,228 models that GO ignores.
  • We built a validated 250-row ncbifam2go seed and 2,455 EC-bridge candidates. The gain is large but mostly TrEMBL: 19 reviewed vs 26,578 entries in a 60-model sample.

The route today, and the proposed one

Why this resource

  • NCBIFAM is prokaryote-heavy, curated per family: 13,253 equivalog models where all members share one function.
  • Newer microbial biology (mobile elements, secretion, anti-phage defense, encapsulins) is where InterPro integration lags.
  • CDD-proper has no native GO (FTP files and Entrez); GO seen in CDD is borrowed from bundled NCBIFAM models.

Where the net-new annotations are

Orange rows: we proposed a specific term over NCBI's broad one ("refine"). Green: the one reviewed gap-fill that survived checking.

Checking the reviewed "gaps"

Row Reviewed gain Prokaryotes What the entries are
VirB5 → GO:0043684 4 4 genuine Brucella VirB5
dGTPase → GO:0008832 13 13 "-like", no EC: curators withheld it
IS630 → GO:0004803 2 2 uncharacterized ORFs; weak
LL-DAP → GO:0010285 2 0 plant ALD1 paralogs
spermidine synthase → GO:0004766 1 0 tobacco PMT1 paralog

Lesson: scope the gain to the model's lineage (PGAP applies to prokaryotes only), and read "-like" as a curator caution.

Structural checks (OpenScientist)

  • NF042963 DUF1156: an intact SAM-dependent DNA amino-methyltransferase, no nuclease. Adds MF GO:0009008; target base (m6A vs m4C) left undecided.
  • NF002326 dGTPase: SAMHD1-like broad dNTPase. Family split: GO:0016793 for all, GO:0008832 and GO:0106375 as clades.
  • NF033545 IS630: DDE transposase with intact catalytic triad; confirms GO:0004803.
  • NF041162 encapsulin: HK97 shell, not a membrane protein; confirms GO:0140737 (weakly blinded).

Status and next steps

  • ✅ Member-DB attribution re-join; CDD-own-GO question resolved
  • ✅ 250-row SSSOM seed validates (just validate-ncbifam-mappings)
  • ⬜ Promote the 2,455 EC-bridge candidates; build the 843 "refine" class
  • ⬜ Full-collection gain run; non-EC families need literature checks
  • ⬜ Exemplar gene reviews whose only support is an NCBIFAM equivalog

Read more: projects/NCBIFam.md · projects/NCBIFam/ncbifam2go.sssom.yaml · projects/NCBIFam/ncbifam2go.candidates.tsv