InterPro Mapping Review Project

Overview

This project reviews InterPro2GO mappings — the GO annotations a protein receives
because it matched an InterPro signature. In GOA these are the IEA annotations with
REFERENCE = GO_REF:0000002, WITH/FROM = InterPro:IPRxxxxxx, and ASSIGNED BY = InterPro. They are produced by a single curated mapping file (InterPro2GO)
that attaches GO terms to an InterPro entry; every protein matching that signature
then inherits all of the entry's terms, with no subfamily resolution.

Two complementary workstreams:

1. Review the flagged mappings. Every gene review already adjudicates its
   InterPro2GO annotations. We aggregate those verdicts to find which InterPro entries
   systematically over-annotate (and to surface individual suspect mappings for
   correction). This is the InterPro analogue of the IBA Annotation Review
   and Over-Annotation Patterns projects.
2. Deep-research the families themselves. For PANTHER, just fetch-gene already
   auto-caches the family into interpro/panther/<FAM>/, and gene deep research is a
   generated process (just deep-research-perplexity ...). There was no equivalent
   deep-research process for InterPro entries — the gap this project closes by adding a
   just deep-research-interpro-family recipe (see
   Family deep research below).

Why InterPro2GO over-annotates

InterPro2GO is a blanket rule per entry. The recurrent failure modes:

• Fold ≠ function. A signature of type: domain or homologous_superfamily
  identifies a module, not the whole-protein function. A whole-protein GO term mapped to
  that module propagates to every protein that merely contains the module.
• No subfamily resolution. A term that is true only for a subfamily is applied to
  the whole entry (the classic neo-functionalization / pseudo-enzyme problem, here
  without even the phylogenetic structure that IBA has).
• Generic terms. Entries routinely carry low-information terms (ATP binding,
  metal ion binding, membrane) that are true but uninformative as core functions.
• Process/component terms on broadly distributed domains. A process term can leak
  into taxa where the pathway is absent.

Current state of the evidence

Across all reviewed genes (regenerate the underlying TSVs — see
Reproducibility):

• 2732 reviewed gene files scanned
• 3652 InterPro2GO (GO_REF:0000002) annotations reviewed
• 1732 (47%) flagged suspect — i.e. the reviewer did not plain-ACCEPT them
• across 1826 distinct InterPro entries

Action breakdown of the reviewed InterPro2GO annotations:

Highest-impact InterPro entries

Top entries by number of suspect mappings (interpro_family_priorities.tsv). These are
the deep-research worklist:

Concrete example (kinase domain): plant BAK1 carried GO:0004672 protein kinase activity from IPR000719; reviewed as MODIFY → GO:0004675 (transmembrane receptor
ser/thr kinase activity), and its GO:0005524 ATP binding as KEEP_AS_NON_CORE.

Family-level deep research

Family deep research is a generated process, the same as gene deep research — not a
hand-written file. It judges whether each InterPro2GO term is appropriate for all
members of the entry, accounting for subfamily divergence, pseudo-enzymes, and taxonomic
scope. (The recipe and its wiring are in Reproducibility.)

A worked starting point (IPR000719, Protein kinase domain) is already cached. Its
InterPro2GO terms are GO:0005524 ATP binding and GO:0006468 protein phosphorylation
— both generic, which is exactly why genes matching it are repeatedly MODIFY'd to
specific kinase-activity children or marked non-core.

Proposed interpro2go edits (SSSOM)

The family verdicts are captured as proposed edits to the InterPro2GO mapping file in
SSSOM YAML — the same format the RHEA and
AMR projects use for proposing mappings —
INTERPRO/interpro2go.sssom.yaml. One row per
(InterPro entry, GO term) reviewed, with a 3-way predicate convention:

• skos:exactMatch — the interpro2go mapping is sound (holds for ~all members);
• skos:broadMatch — over-broad, retain but demote to a child entry / treat as non-core;
• skos:exactMatch + predicate_modifier: Not — remove (factually incorrect for the entry).

Reproducibility

Supporting material under INTERPRO/ (extractor, generated TSVs,
and the SSSOM mapping set):

• INTERPRO/suspect_interpro_mappings.tsv — per-annotation verdicts.
• INTERPRO/interpro_family_priorities.tsv — per-entry worklist.
• INTERPRO/interpro2go.sssom.yaml — proposed edits.

# regenerate the per-annotation / per-entry worklist TSVs
uv run python projects/INTERPRO/extract_suspect_interpro_mappings.py

# validate the SSSOM mapping set (structure + GO label check on interpro2go.terms.yaml)
just validate-interpro-mappings

Family deep research (the InterPro analogue of just deep-research-perplexity) is wired as:

• templates/interpro_family_research.md — the deep-research prompt (judges whether each
  InterPro2GO term holds for all members; subfamily divergence, pseudo-enzymes, taxon scope).
• scripts/deep_research_interpro_family.py — wrapper that loads the cached
  interpro/<db>/<ID>/<ID>-metadata.yaml as context and runs the provider, writing
  interpro/<db>/<ID>/<ID>-deep-research-<provider>.md.

# provider defaults to falcon (Edison)
just deep-research-interpro-family IPR000719
just deep-research-interpro-family IPR001128 openai --fallback perplexity-lite
just deep-research-interpro-family PTHR10314 perplexity --database panther

STATUS

Workstream 1 — review flagged mappings

• [x] Extractor that aggregates InterPro2GO verdicts across all reviews
• [x] Per-entry priority worklist (1826 entries; 265 with ≥3 suspect mappings)
• [ ] Work down the top entries: confirm the suspect verdicts, propose canonical
  replacement terms, and feed corrections back to the gene reviews
• [ ] Summarize per-entry recommendations for InterPro2GO curators

Workstream 2 — family deep research

• [x] Confirmed the metadata fetcher supports interpro (IPR) entries
• [x] Generated deep-research process: templates/interpro_family_research.md,
  scripts/deep_research_interpro_family.py, and the
  just deep-research-interpro-family recipe
• [x] Seed example cached (interpro/interpro/IPR000719/)
• [x] First family deep research generated with falcon/Edison
  (IPR000719-deep-research-falcon.md): verdict that both InterPro2GO terms
  (ATP binding, protein phosphorylation) over-annotate the domain because it
  also matches pseudokinases — REMOVE at the domain level, restrict to catalytic
  children (GO:0004674 / GO:0004713)
• [x] Batch 1 of 5 more top entries researched with falcon/Edison (P450, Cu/Zn SOD,
  GPCR, NRAMP/SLC11, DnaJ) — see the table below
• [x] Began feeding verdicts back into gene reviews (DnaJ family): the reviews are
  strongly concordant with the family research — all 7 DnaJ genes with the
  InterPro2GO ATP binding annotation already flagged it (5 REMOVE, 1 MODIFY-to
  ATPase-activator, 1 over-annotated). Hardened the one soft outlier (yeast/YDJ1,
  MARK_AS_OVER_ANNOTATED → REMOVE) and attached the IPR012724 family report as
  corroborating evidence.
• [ ] Per-family, hunt for the higher-value case: a gene that currently ACCEPTs the
  flagged term but is actually one of the verdict's exception members (pseudokinase,
  copper chaperone, atypical chemokine/orphan receptor, non-catalytic P450) — a genuine
  missed over-annotation rather than a soft-vs-hard mismatch
• [x] Captured the family verdicts as proposed interpro2go edits in SSSOM YAML
  (INTERPRO/interpro2go.sssom.yaml, 17 mappings over the 6 entries) — the
  consortium-facing deliverable, validated via just validate-interpro-mappings
• [ ] Continue down the worklist (interpro_family_priorities.tsv)

Family deep-research verdicts (falcon/Edison)

• [ ] Run just deep-research-interpro-family <IPR> (falcon/Edison default) for the next entries

Last updated: 2026-06-20

NOTES

2026-06-20

Project creation. Scoped the InterPro2GO (GO_REF:0000002) review. Built the
extractor and the per-entry priority worklist from all 2732 reviewed genes: 3652
InterPro2GO annotations, 47% flagged suspect across 1826 InterPro entries. Broad
domain/superfamily signatures dominate the suspect list (protein kinase domain, P450,
Cu/Zn SOD, GPCR), confirming the "fold ≠ function" failure mode as the main driver.

Closed the PANTHER-vs-InterPro deep-research gap. Gene deep research is a generated
process (just deep-research-<provider>); there was no equivalent for the InterPro
entries behind InterPro2GO annotations. Added the InterPro-family analogue —
templates/interpro_family_research.md, scripts/deep_research_interpro_family.py, and
the just deep-research-interpro-family <IPR> [provider] recipe (provider defaults to
falcon/Edison) — so families are researched by the same generated pipeline (output:
interpro/<db>/<ID>/<ID>-deep-research-<provider>.md), with IPR000719 cached as a
seed.

Batch 2 + first proposed new mapping. Researched 6 more families; 3 grounded cleanly
(Radical SAM, Ras-type small GTPase, MADS-box) and are in the SSSOM (now 25 mappings).
Notable findings:

• Genericity ≠ wrongness (Radical SAM, IPR007197). I predicted the MF root term
  catalytic activity would be a REMOVE. The research says ACCEPT: because the
  superfamily catalyzes >100 mechanistically different reactions, the only universally
  true MF really is "is an enzyme", so the maximally generic term is the correct
  family-level annotation — replacing it with anything more specific would over-annotate.
• First annotation-gain proposal (Ras-type, IPR020849). The entry maps GTP binding
  but not GTPase activity (GO:0003924), even though the GTP-hydrolysis machinery
  (P-loop, Switch II/Gln61, Mg²⁺) is universal — so we propose ADDING it (an
  exactMatch row, flagged for curator confirmation since intrinsic hydrolysis is
  GAP-accelerated).
• Domain-intrinsic vs whole-protein (MADS-box, IPR002100). I expected to add
  DNA-binding TF activity; the research argues against it — the MADS domain provides
  DNA binding + dimerization, but being a transcription factor is a whole-protein
  property (K/C domains, complex context), so adding it would over-annotate the domain.
• QC catch. 3 of the 6 runs (sigma-54, pseudouridine synthase, GAPDH) silently
  returned ungrounded reports (exit 0, file written, but "no contexts were retrieved
  … not grounded in evidence", zero real citations) — likely Edison retrieval throttling
  under 6-way parallel load. A sequential re-run also failed fast (~3 s each, no
  retrieval), so this is a transient Edison retrieval-backend outage, not a load issue —
  these 3 entries are deferred (metadata cached) for re-running when the backend
  recovers. Excluded from the SSSOM. (Detect with: grep for "no contexts were retrieved"
  or citation_count/zero real refs — a groundedness guard worth adding to the wrapper.)

Batch 1 of family deep research (falcon/Edison). Ran five more top entries: P450
(IPR001128), Cu/Zn SOD (IPR001424), GPCR Class A (IPR000276), NRAMP/SLC11
(IPR001046), and DnaJ (IPR012724). See the verdict table under Workstream 2. A
recurring, independently-reached pattern: cofactor/binding terms (heme binding, metal ion binding) and broad transport terms hold family-wide, but whole-protein activity
and process terms attached to a structural module over-annotate — most sharply for
IPR012724, where Edison flags ATP binding as factually wrong on DnaJ (the Hsp70 partner
binds ATP), and for IPR001424, where superoxide metabolic process mis-annotates
copper-chaperone members that do not dismutate.