NSUN2 (Q08J23) — review notes

The question this review has to settle

GOA carries two catalytic molecular functions for NSUN2, each with experimental support:

The enzyme is real and nobody disputes it. What is disputed is one of its substrate classes.
The position taken here: tRNA/ncRNA m5C is core; mRNA m5C is real but demoted to non-core,
with the methodological dispute recorded in reason and the unresolved question moved to
suggested_questions. Nothing is removed: GO:0062152 carries an EXP code, and CLAUDE.md
forbids overruling an experimental annotation whose full text I have not read.

Why the tRNA activity is core

NSUN2 is the human orthologue of yeast Trm4, and the founding paper shows position-specific
catalysis on a defined substrate:

Human loss-of-function confirms the substrate in vivo:

And the mitochondrial tRNA activity was reconstituted with purified protein — the gold standard,
a defined enzyme plus SAM plus substrate giving a mapped product:

A third, independent ncRNA substrate — vault RNA VTRNA1.1 C69 — is likewise attributed
exclusively to NSUN2:

The mRNA m5C claim: what is solid and what is not

Evidence that NSUN2 does methylate mRNA

The strongest orthogonal evidence is not bisulfite sequencing at all. The RNABPP chemoproteomics
study used mechanism-based covalent crosslinking (5-fluorocytidine traps the enzyme on its
substrate through the catalytic cysteine) combined with oligo-dT capture, and separately measured
nucleoside m5C by LC-MS/MS in NSUN2 knockout cells:

A catalysis-dependent rescue exists too — the mRNA export phenotype is restored by wild-type but
not catalytically dead NSUN2, which is the right control and is rarely done in this literature:

And in vitro methylation of a defined mRNA fragment was reported for p16:

So the activity itself should not be removed, and indeed the 2026 review that catalogues the
problems in this field still accepts the enzyme assignment:

Evidence that the mRNA m5C maps and their biological weight are unreliable

The same RNABPP authors are candid about the main confound in their own headline number:

The 2026 Cells review sets out the mapping problems:

and the stoichiometry problem, which is the deeper one because it undermines the reader-based
mechanisms on which most of the downstream literature rests:

The downstream cancer/inflammation literature

A large and still-growing set of papers assigns a phenotype to NSUN2 m5C on one named transcript,
typically read out by YBX1 or ALYREF, e.g.

These studies essentially never provide orthogonal, site-resolved validation of the single
methylation site (no purified-enzyme in vitro methylation of that site, no site-mutant knock-in,
rarely a catalytically dead rescue). Given the base rate of ~1 m5C per several thousand cytosines
in mRNA and the poor agreement between mapping methods, individual site-level claims of this kind
should be treated as provisional. This does not make the enzyme's mRNA activity false; it makes
the specific biology attributed to it weak.

Decision and the reusable principle

Principle: separate the enzyme from the substrate class, and demote rather than delete.

When a well-established enzyme is claimed to act on a second substrate class whose detection
depends on a contested assay:

  1. Ask whether the catalytic assignment has any evidence that does not depend on the contested
    mapping assay — purified-enzyme in vitro reaction, mechanism-based covalent capture,
    catalytically dead rescue, orthogonal mass spectrometry. For NSUN2/mRNA, several of these
    exist, so the MF term stays.
  2. Ask whether the site maps and the downstream biology have independent support. For
    NSUN2/mRNA they largely do not, so the term is marked non-core and the dispute is recorded.
  3. Never REMOVE on this basis when the annotation carries an experimental code, because removal
    asserts the activity does not occur, which is a stronger claim than the evidence supports in
    either direction.

The mirror-image case is NSUN7 (reviewed alongside this gene): there the enzyme itself fails —
no cofactor binding, catalytic residue substituted, no change in product on knockout — and the
only methyltransferase annotation is an IEA. That is a REMOVE. The contrast is the point: the
action is driven by whether the challenge lands on the catalysis or on the substrate mapping.

Paralog leakage in the IBA block

Three IBAs on NSUN2 look like NSUN4 functions propagated across a family-wide node
(PANTHER:PTN000516076). The WITH/FROM column of NSUN2-goa.tsv names MGI:MGI:1919431,
which is mouse Nsun4 — the mitochondrial 12S rRNA methyltransferase that partners MTERF4 in
mitoribosome large-subunit assembly:

IBA term WITH/FROM Comment
GO:0009383 rRNA (cytosine-C5-)-methyltransferase activity MGI:MGI:1919431, PTN000516076 Nsun4 substrate; NSUN2 has no reported rRNA substrate
GO:0006364 rRNA processing MGI:MGI:1919431, PTN000516076 ditto
GO:1902775 mitochondrial large ribosomal subunit assembly MGI:MGI:1919431, PTN000516076 NSUN4/MTERF4 function

NSUN2 is in the mitochondrial matrix, but its job there is mt-tRNA C48-50, and the same paper
that put it there reports that its loss does not much perturb mitochondrial translation:

I have marked these three MARK_AS_OVER_ANNOTATED with a propagation_review recording only what
I actually checked: the donor gene identity from the GOA WITH/FROM column. I did not inspect
the PAINT tree, so the PANTHER node is recorded as UNRESOLVED rather than being blamed. By
contrast the tRNA-related IBAs (GO:0016428, GO:0002127) list donors from five clades plus
NSUN2's own experimental annotation, which is the expected and correct pattern for a deeply
conserved Trm4 function, and they are accepted.

Term-choice problem: GO:0006397 mRNA processing

Four annotations (one ARBA IEA plus three IDAs from PMID:22395603, PMID:28418038, PMID:31358969)
use GO:0006397 mRNA processing, whose definition covers conversion of a primary transcript into
mature mRNA — splicing, capping, polyadenylation. None of the three cited papers reports NSUN2
doing any of those: they report mRNA stabilisation (p16, HDGF) and nuclear export
(ALYREF). This is a term-choice problem, not an evidence problem, so MODIFY with replacements
GO:0048255 mRNA stabilization and GO:0010793 regulation of mRNA export from nucleus (the
latter is already annotated to this gene from PMID:28418038).

Other decisions