ProtNLM2 "novel" hits are mostly InterPro2GO coverage gaps

Why the "correct novel" ProtNLM2 hits are missed by standard pipelines

← back to ProtNLM2 Evaluation

Bottom line

The predictions we called "correct novel" for uncharacterized proteins are likely
correct (they have no experimental support — the assessment is domain/orthology-based)
and, more importantly, they are not novel biology. They are functions that follow
directly from the protein's domains but that the standard InterPro2GO pipeline
(GO_REF:0000002) does not emit. In each case the gap is an InterPro/InterPro2GO
coverage problem, not a discovery. Three distinct mechanisms account for the three
flagship examples — verified against the live InterPro API and the current InterPro2GO
release (!version date: 2025/09/01).

The three mechanisms

1. No InterPro2GO mapping exists for the domain — olfactomedin (A0A8C9H4D2 / OLFML2A)

The protein carries the olfactomedin domain (InterPro:IPR003112, Pfam PF02191/OLF)
and sits in family IPR050605 (Olfactomedin-like domain-containing protein). Neither
entry has any GO term (InterPro API go_terms: null for both; zero lines in
InterPro2GO). So the ECM-organization function ProtNLM2 predicts is simply not derivable
from InterPro2GO.

Why no mapping? The olfactomedin domain is functionally promiscuous — it occurs in
myocilin, the olfactomedins/noelins, gliomedin, and the latrophilin adhesion GPCRs, whose
functions diverge widely — so a single reliable domain→GO mapping cannot be assigned.
(Note: the 11 "olfactomedin"-keyword lines in InterPro2GO are all olfactory
receptor/marker entries — IPR000725, IPR004117, IPR009103, IPR036727 — homonyms, not the
olfactomedin domain.)

2. Mapping exists only on a superfamily entry the protein was not assigned — KilA-N (A2FPI7)

The Trichomonas protein is assigned the domain-level entries IPR017880 (KilA, N-terminal)
and IPR018004 (KilA/APSES_HTH) — both have no GO term. The DNA-binding function does
exist in InterPro2GO, but only on the superfamily entry IPR036887 (HTH APSES-type
DNA-binding domain superfamily → GO:0003677 DNA binding), which this protein was not
assigned. So the DNA-binding annotation that would apply to APSES/KilA-N transcription
factors never propagates to it. (A2FPI7 currently has zero GO annotations of any kind.)

This is the one case where the GO annotation tracker shows active curator curation of the
family: geneontology/go-annotation#1188
(closed 2015) records that S. pombe bqt4, which carries the APSES domain (IPR003163), is
not a transcription factor — it anchors DNA to the nuclear envelope. The InterPro curator
deleted the transcription-factor GO terms from the entry (keeping DNA binding) and alerted
PROSITE/UniRule, because PROSITE PS51299 was auto-assigning DNA-binding GO/keywords from this
domain. The lesson generalises: the APSES/KilA-N domain is functionally heterogeneous (fungal
transcription factors, DNA-anchoring proteins, phage/eukaryotic KilA-N proteins), so curators keep
its GO deliberately sparse and conservative — which is why the domain-level entries A2FPI7 carries
have no GO at all.

3. Protein matched only an unintegrated Pfam; the InterPro entry that carries the mapping was not assigned — MCM-4 (A0A061AL94)

Here the InterPro2GO mapping is rich and present: IPR001208 (MCM domain) →
DNA binding + ATP binding, and IPR008047 (Mini-chromosome maintenance complex protein 4)
→ DNA binding, DNA helicase activity, ATP binding, DNA replication initiation, MCM complex.
But the C. elegans protein was assigned only Pfam PF21128 (WHD_MCM4), which is
unintegrated into InterPro and carries no GO term (InterPro API: integrated: null,
go_terms: null). Because it never matched the InterPro MCM entries, InterPro2GO had
nothing to key on — even though fully-assigned MCM4 orthologs receive the whole nucleus /
MCM-complex / replication annotation set.

Was a mapping removed?

Curator-driven changes to InterPro2GO are reported on the GO annotation tracker
(geneontology/go-annotation), so that is the
authoritative place to check. Searching it:

More broadly, InterPro2GO mappings are unstable and are removed in bulk from time to time — e.g.
a PANTHER 10.0 update left 332 InterPro entries signature-less, removing 522 InterPro2GO mappings
and ~178,926 protein GO annotations (Sangrador-Vegas et al., GO annotation in InterPro: why
stability does not indicate accuracy in a sea of changing annotations, Database 2016,
PMC4799721). But for our three examples the
tracker shows the dominant cause is conservative non-coverage (never mapped, or TF terms
actively pruned), not accidental loss.

Systematic view: this is the PFAM project finding at gene level

The Pfam → GO mapping project quantifies exactly this gap across all of Pfam:
InterPro2GO is deliberately conservative, and 82.6% of Pfam-A families (24,888) have zero GO via
InterPro — including canonical, well-understood domains (SH2, EGF, Kringle, Actin). ~18k of the
uncovered families are named and tractable. The three ProtNLM2 "novel" hits here are concrete,
gene-level instances of that gap: ML function prediction is filling the ~18k-family space where
InterPro has abstained. The PFAM project's conclusion — that new value comes from annotating
where InterPro abstained, not from splitting lumped entries — is the flip side of the same coin.

Curation implication

Verification

Sources: InterPro2GO (EBI) ·
GO_REF:0000002 ·
Sangrador-Vegas et al. 2016 (PMC4799721)