Does pfam2go add precision?

Testing Pfam's GO mapping against InterPro2GO, and where new mappings could help

AI Gene Review · projects/PFAM · 2026

Bottom line

  • No hidden precision: of 9,871 pfam2go assertions, 9,844 match the parent InterPro entry and 0 are more specific. pfam2go is generated from InterPro2GO.
  • The real gap is coverage: 24,888 of 30,134 Pfam-A families (82.6%) get no GO term through InterPro, including SH2 and EGF.
  • We curated 9 Pfam families as entries: 5 proposed mappings, 4 rejected on same-family counter-examples.

Why ask

  • Gene reviews lean on InterPro2GO (GO_REF:0000002) for IEA rows.
  • An InterPro entry can lump several Pfams, so its term must cover all of them.
  • If pfam2go kept sharper per-family terms, it would be a free source of precision.
  • Method: map each Pfam to its parent entry via interpro.xml <member_list>, compare terms over the GO is_a/part_of DAG.

Hypothesis versus files

Part 1 result

Category Assertions Families
SAME as InterPro entry 9,844 —
MORE_SPECIFIC 0 0
MORE_GENERAL 1 1
DISJOINT, genuine 1 1
DISJOINT, release skew 25 13

The one specificity difference (MORE_GENERAL) runs the other way: PF08214 has GO:0004402 while its entry IPR016849 has the more specific GO:0010484. The genuine DISJOINT row is PF08214 → GO:0006355 (regulation of transcription). Skew: InterPro 109.0 membership vs a 28 Apr 2026 GO mapping snapshot.

Part 2: headroom is in coverage

Only 76 GO-bearing InterPro entries have ≥2 Pfam members, mostly redundant HMMs of one domain. Splitting lumped entries yields almost nothing.

Nine Pfam entry reviews

Proposed (5) Rejected (4)
PF27512 LeuD: GO:0003861, GO:0009098 PF14681 UPRTase: uridine kinases in family
PF02431 Chalcone: GO:0045430 PF16363 GDP_Man_Dehyd: GALE, UXS1 in family
PF07228 SpoIIE: GO:0030435 PF13360 PQQ_2: RqkA kinase, PedH in family
PF09043 Lys-AminoMut_A: GO:0047826 PF13561 adh_short_C2: FabG beside FabI
PF16552 OAM_alpha: GO:0047831

interpro/pfam/<PFAM>/<PFAM>-review.yaml, schema pfam_entry_review.yaml; all nine pass just validate-pfam-reviews. Proposals are candidates for curator validation.

Status and next steps

  • ✅ Part 1 and Part 2 scripts, committed results (RESULTS.md, HEADROOM.md)
  • ✅ 9 curated Pfam entry reviews, validated
  • ⬜ Author mappings for the ~18k named unmapped families (curation or grounded prediction)
  • ⬜ Compare subfamily-grained sources (NCBIfam, PANTHER) against InterPro2GO

Read more: projects/PFAM.md · projects/PFAM/HEADROOM.md · projects/PFAM/PROPOSED_MAPPINGS.md