Contested Functions: a 2025-2026 read-list

Species: human

Contested Functions — a 2025-2026 read-list

A survey of literature published between January 2025 and August 2026, looking for
individual human genes whose molecular function is actively disputed, and turning each
dispute into a question a GO curator could adjudicate.

This is a companion to the parent project, which registers what
biology does not know. Here the problem is the opposite and, for curation, sharper: biology
knows two incompatible things, both published, both often already in GOA.

The contested-function gap

The parent page's taxonomy has three kinds of gap — biology (nobody knows), curation
(known but unannotated), ontology (known but inexpressible). Cases below mostly belong to a
fourth kind:

Contested-function gap — two or more lines of experimental work assign the gene product
mutually incompatible molecular functions, and no consensus has formed. Neither UNDECIDED
(which implies we could not access the evidence) nor ACCEPT (which implies the matter is
settled) is quite right, and the annotations frequently coexist in GOA with experimental
evidence codes on both sides.

Two properties make these unusually high-value curation targets:

  1. The dispute is already inside GOA. For TMEM175 the GOA carries proton channel activity
    (IDA, three references) and potassium ion leak channel activity (IDA, two references),
    all positively qualified. For TMEM65 it carries calcium:sodium antiporter activity
    twice — once as enables/IDA and once as NOT|enables/IDA — citing the two conflicting
    2025 papers, so the controversy is encoded literally as a positive and a negative annotation
    of the same term on the same gene. A reviewer does not have to go looking for the problem;
    it is sitting in the annotation set.

Read the qualifier, always. An early draft of this page asserted that GOA carried the
ion-channel claim for TMEM120A. It does not: the three experimental annotations of
monoatomic ion channel activity are all NOT|enables. Reading only the term and
evidence code inverts the meaning of the annotation, and the mistake is invisible unless the
qualifier field is fetched explicitly.
2. A downstream literature usually keeps building on the contested claim. The pattern is
consistent: a methods-level or biochemical challenge appears in one venue while a large
disease-biology literature continues to assume the original assignment. NAT10, FTO, NSUN2 and
TRMT6/TRMT61A are all in this state.

A caution that applies to every entry below. A live dispute is not a licence to REMOVE
an experimental annotation. Per the repo's standing rule, curators who made an IDA read the full
text and we have not. The right output for most of these is a MODIFY to a term the evidence
actually supports, a KEEP_AS_NON_CORE, or an ACCEPT whose reason states plainly that the
assignment is contested and names the challenge. The value here is in writing the dispute
down, not in picking a winner from an abstract.

How this list was built

Free-text PubMed searching for "controversy"-type phrasing is close to useless — the API's query
translator silently drops phrase clauses and ANDs everything else, so "remains controversial"
returns nothing while a five-term query returns zero rather than an error. The searches that
actually worked were:

Retraction status was not checked during the survey, and should have been. That gap surfaced
when reviewing TNFRSF21: the founding claim behind its entry had been retracted in 2024 and the
survey had no step that would have caught it. Checking afterwards, all 104 PMIDs cited on this
page were re-queried for Retracted Publication / Expression of Concern publication types and
for RetractionIn / ExpressionOfConcernIn / ErratumIn links. None is retracted. Three
carry errata, all already handled in the relevant reviews: PMID:35679869 (ErratumIn 35931039),
PMID:41679312 (ErratumIn 42001851), and PMID:36805701 — which has two, PMID:37626192 from 2023
and PMID:40890508 from 2025; only the second was known to this survey. A retraction/erratum sweep
belongs in the method from the start, not as an afterthought.

Every PMID below was verified against PubMed by re-fetching title, journal, year and DOI; the
DOIs were extracted from the article's own ELocationID/ArticleIdList, not from the
reference list, which is an easy way to attach a real-looking but wrong DOI to an entry. Current
GO molecular-function annotations were read live from QuickGO. Repo status was checked directly
against genes/human/.

What was not done: no full texts were read. Confidence ratings reflect journal, assay type,
independence of the labs, and whether the two sides actually engage each other — not a judgment
on the underlying experiments.

Shortlist

Ranked by curation value = (sharpness of the dispute) × (exposure in GOA) × (absence of an
existing review).

Gene The dispute Already in GOA? Repo Conf.
TMEM175 H⁺-selective channel vs K⁺ channel with incidental H⁺ permeability Both, IDA each way none High
TMEM120A Mechanosensitive channel (TACAN) vs ER acyl-CoA/lipid-synthesis protein Channel term is NOT×3; one stray positive ISS none High
TMEM65 / SLC8B1 Which protein is the mitochondrial Na⁺/Ca²⁺ exchanger — and is it Na⁺ at all? enables + NOT\|enables, same term, both IDA none High
FTO mRNA m⁶A demethylase vs m⁶Am/snRNA enzyme vs hydroxylase (no demethylation) Yes, IDA+IMP none High
NAT10 mRNA ac4C writer vs 18S rRNA/tRNA only Yes, IDA none High
TRMT61A mRNA m1A methyltransferase vs tRNA-A58 only (mapped sites are inosine) Yes, IDA none High
GSDMC Plasma-membrane pore vs Rab7-vesicle permeabilizer vs nuclear scaffold Only IBA lipid binding none High
CASP4 Binds LPS molecules vs LPS membranes of positive curvature; substrate identity Yes, IDA none High
MEFV Indirect Rho/PKN sensing vs direct CDC42 binding by B30.2 No GTPase binding term 937 lines, no CDC42 High
SLC45A4 Neuronal plasma-membrane polyamine vs peroxisomal putrescine transporter Both, IDA (+ legacy sucrose IBA) none High
ALKBH1 Genomic 6mA demethylase vs mt-tRNA f5C oxidase (is mammalian 6mA real?) Yes, EXP none High
NSUN7 Catalytically inactive pseudoenzyme vs active m5C writer methyltransferase activity IEA none High
GRID1 Non-ionotropic scaffold vs acetylcholine receptor vs glutamate/glycine-gated Yes — and a GABA receptor activity IDA none High
TMEM63B Mechanosensitive cation channel vs mechanically activated lipid scramblase Both, IDA none High
ZBP1 Can the Zα domains actually convert dsRNA to Z-form? Human vs mouse signalling Z-RNA immune receptor ISS none High
PRSS23 Serine protease vs serine pseudoprotease serine-type endopeptidase IEA none High
GPR50 / GPR37 / GPR158 / GPR75 Contested or absent deorphanizations Varies; GPR75 IBA-only none Med-High
ALDH4A1 Non-catalytic subunit of the MPC complex, or is MPC an MPC1/MPC2 dimer? No MPC term 452 lines, no MPC Med-High
LRRC8A cGAMP transport is real; is it core? (see correction below) Yes, IDA none Med-High
SLC7A11 Lysosomal proton channel, or H⁺ leak secondary to antiport? — 1443 lines; overstated Med
PNPLA3 Own lipase activity vs ABHD5 sequestration vs neomorph Yes, many EXP/IDA none Med
MTCH2 / MTCH1 Insertase vs BAX/BAK pore factor vs CPT1 regulator vs SLC25 carrier insertase IDA (MTCH1 too) 341 lines, thin Med

Worked entries

TMEM175 — potassium channel or proton channel? (top pick)

TMEM120A — mechanosensitive channel, or an ER lipid enzyme's activator?

TMEM65 and SLC8B1 — who is the mitochondrial Na⁺/Ca²⁺ exchanger, and is it sodium?

This is the cleanest three-way conflict found, and GOA encodes it directly.

FTO, NAT10, TRMT61A, ALKBH1, NSUN2, NSUN7 — the epitranscriptomic writer/eraser cluster

Six genes with the same structural problem, and they are best reviewed together because the
methodological argument is shared: the enzyme is real, but the claimed substrate class may be
a mapping artifact, while a large disease literature continues to build on it.

Curator question for the cluster: where the writer/eraser activity on rRNA or tRNA is
solid but the mRNA claim rests on a contested mapping method, is the right action MODIFY to
the well-supported substrate, or ACCEPT with the dispute recorded? A shared position across
the six would be more valuable than six independent calls.

GSDMC — three incompatible molecular functions, none of them in GOA

MEFV — the highest-value update to an existing review

PRSS23 — a clean IEA-versus-experiment case

SLC7A11 — a correction our own review needs

Not a literature controversy so much as an internal inconsistency the controversy exposes.

Second tier — verified, worth reading, lower priority

Systematic defects found while reviewing

Three findings from this batch are not about one gene. They are mechanical faults that
mis-annotate whole sets of proteins, and they are worth fixing upstream rather than gene by gene.

InterPro2GO maps "calcium permeable" to "calcium activated"

InterPro IPR045122 is named "Calcium permeable stress-gated cation channel 1-like".
InterPro2GO maps it to GO:0005227 calcium-activated cation channel activity, whose
definition requires a channel "that opens when a calcium cation has been bound by the channel
complex or one of its constituent parts".

Permeable-to-calcium and gated-by-calcium are opposite directions of causation. TMEM63/OSCA
channels are stretch- and osmolarity-gated and merely Ca²⁺-permeable, so the term is wrong for
every one of them. The signature matches 23,725 proteins across species; in human it lands on
TMEM63A, TMEM63B and TMEM63C. (The other human proteins carrying this term by InterPro IEA —
KCNN1-4 — are genuine SK channels, so the term itself is fine; only this mapping is wrong.)
Verified live against the InterPro and QuickGO APIs.

InterPro2GO gives ZBP1 a deaminase activity it has no domain for

InterPro IPR042371 is the generic "Z-binding domain" entry, and it carries both
GO:0003723 RNA binding and GO:0003726 double-stranded RNA adenosine deaminase activity.
The deaminase activity belongs to ADAR1, which shares the Zα domain. ZBP1 has a Zα domain and no
deaminase domain, so it inherits an enzymatic activity it cannot perform.

InterPro2GO gives every RIG-I-like receptor a DNA-binding activity

InterPro IPR006935 is "Helicase/UvrB, N-terminal", and it maps to GO:0003677 DNA
binding, GO:0005524 ATP binding and GO:0016787 hydrolase activity. The family is dominated
by DNA-acting enzymes — UvrB, type III restriction endonucleases — but the same N-terminal fold is
present in the RNA-sensing RIG-I-like receptors, so IFIH1 (MDA5), a cytosolic double-stranded
RNA sensor, carries a DNA-binding annotation. MDA5 has no reported direct DNA binding, and
activation by DNA viruses is accepted to be indirect via RNA. The signature matches 73,506
proteins. Verified against the InterPro API.

A permeation property written into a gating term, twice more

Two further cases, same mistake, found independently in different families — a property of what a
channel conducts recorded as a statement about what opens it:

The shape these three share

All three are the same failure, and it is worth naming because a per-gene fix will not stop it:
a domain-level signature carries an activity that belongs to only some of the proteins bearing
that domain. A Zα domain does not make you a deaminase; a helicase N-terminal fold does not make
you DNA-binding; being calcium-permeable does not make you calcium-activated. The mapping is
made once at the family level and then inherited by every member, including the ones for which it
is false. Three were found in a sample of ~55 genes, which suggests the rate is not negligible.

A frameshifted ORF's activity annotated onto the parent protein

ALKBH1 carries GO:0042056 chemoattractant activity and GO:0050918 positive chemotaxis —
incoherent for a matrix-targeted mitochondrial 2-oxoglutarate dioxygenase with no signal peptide.
The IEA derives from mouse ALKBH1 via PMID:16860792, whose own abstract explains the problem:
"The Nrp gene is encoded as a forward frameshift to the hypothetical alkylated DNA repair protein
AlkB." NRP is a secreted, SDF-1-like factor translated from a different reading frame of the
same locus. Its chemoattractant activity was annotated onto the dioxygenase. The mouse
annotation likely warrants MGI's attention as well.

The ontology cannot express protein nitration at all

GLOD4 is reported to catalyse selective peroxynitrite-mediated protein tyrosine nitration
(PMID:41628334, PNAS 2026). GO has no term for protein nitration, in any aspect — verified:
QuickGO searches for "protein nitration", "tyrosine nitration" and "nitrase" return nothing
relevant. The near misses are all wrong in a way that matters:

So the only truthful annotation available is the near-vacuous GO:0140096 catalytic activity,
acting on a protein. This is a clean ontology gap rather than a curation failure, and it is the
mirror image of the FTO m6Am problem noted above: where GO lacks a term, a real activity gets
recorded worse than a contested one that happens to have precise vocabulary.

UniProt cites a review as function evidence for PEX39

The FUNCTION block on Q5I0X4 (PEX39) cites two references: PMID:40739340, the primary
Nat Cell Biol 2025 paper, and PMID:37160800 — which is "Peroxisomes: novel findings and
future directions", a Histochem Cell Biol 2023 review/meeting article that predates the
deorphanization and is not primary evidence for anything about PEX39's function. Verified against
the UniProt REST API and PubMed. Minor, but it is the kind of citation that later propagates as
though it were independent support.

Paraphrase presented as quotation, repo-wide, in an unvalidated field

Found while updating MTCH2: five of its supporting_text entries attributed to a deep-research
file began "Falcon synthesis supports..." and appear nowhere in that file. They are the
reviewer's own summary narration sitting in a field that, by schema, means a verbatim quote.

They survived because the reference validator checks supporting_text only for PMID:
references. Quotes attributed to file: references — deep-research reports, UniProt records,
GOA tables — have never been checked at all.

A repo-wide scan (scripts/check_file_supporting_text.py, output in
reports/file_supporting_text_mismatches.json) checked 59,253 file: supporting_text entries.
No cited file: path failed to resolve, and none that resolved failed to open.

This total is a dated snapshot and will drift every time main is merged in — earlier runs over
smaller corpora reported 55,433 and 59,154. The denominator is the unstable part; re-running after
the most recent sync moved it by 99 entries and changed the mismatch list not at all, so the 99
file: quotes main brought with it are all clean. Quote the narration subset below rather than
this figure if you want the number that means something.
Raw mismatch counts overstate the problem and should not be quoted as a fabrication count: UniProt
.txt records carry two-letter line-prefix codes that break substring matching (stripped in the
script), and genuine benign cases remain — smart quotes, ellipses, quotes stitched across
non-contiguous lines.

The unambiguous subset is the one worth acting on: 970 entries whose "quote" begins with a
narration word such as "Falcon report summarizes..." or "Falcon deep research supports...".
These cannot be verbatim source text by construction. Spot-checked: the ARATH/BAK1 deep-research
file contains no occurrence of "Falcon" anywhere in its prose — its only match is its own
provider: falcon frontmatter line — yet BAK1's review carries 52 supporting_text entries
opening with "Falcon report summarizes", and 74 such rows appear across the whole report. (A plain
grep of that file counts 53; the extra hit is a findings[].statement, not a supporting_text,
and the scan only looks at the latter.)

(An earlier version of this paragraph used ARATH/ABI1 as the example. That example went stale in the
main merge: ABI1's review no longer has any such entry and it no longer appears in the report at
all — it is one of the reviews already corrected on main, which is part of why the subset fell from
1,042 to 970. The old wording was also loose, claiming "zero occurrences of the string Falcon" when
three case-insensitive matches were present in provider metadata and artifact paths.)

Two things follow. First, extending the reference validator to cover file: references would stop
this recurring — it is the same check that already works for PMIDs. Second, the existing entries
need either requoting or demotion to a non-quote field; they are not necessarily wrong as
summaries, but they are not quotations and should not be readable as evidence that a source said
something in those words.

A confirmed broken citation in GOA, live now

NLRP3 carries GO:0060090 molecular adaptor activity and GO:0030674
protein-macromolecule adaptor activity as IDA from PMID:1189953. That PMID resolves to
"[Profanities and the profane person]", Acta Psiquiatr Psicol Am Lat, 1975. The intended
reference is plainly PMID:31189953 — "Structural mechanism for NEK7-licensed activation of
NLRP3 inflammasome", Nature 2019 — a digit-dropped identifier. Other NLRP3 annotations from that
same paper (GO:0051604, GO:0140608) use the correct ID, so the error is isolated to these two
rows. Confirmed against both the cached GOA file and a live QuickGO query, so it is present in the
current data rather than a stale local copy.

Unlike the two below, this one is not a judgement call: a well-formed identifier that resolves to
the wrong paper is exactly the failure mode the project's own reference-review guidance exists to
catch, and it is invisible to any check that only asks whether a PMID resolves.

Two suspected data errors, flagged not acted on

Checked and set aside

Recorded so the same ground is not re-covered.

Suggested first picks

If three genes are to be reviewed from this list:

  1. TMEM175 — sharpest dispute, both sides in GOA with IDA, a journal commentary framing it as
    open, no existing review. The single best test of whether our review format can represent a
    live controversy honestly.
  2. TMEM65 + SLC8B1 together — forces a two-gene review and a decision about a transport MF
    that GOA currently duplicates across two proteins.
  3. MEFV — highest-value update to work already done; a 937-line review with a documented,
    verified hole (no CDC42) that three simultaneous 2026 papers fill.

NSUN7 and PRSS23 are the cheapest wins if a quick demonstration is wanted: both are
pseudoenzyme reclassifications where the contested GO annotation is IEA-only.

Provenance and caveats

Three separate verification failures were caught while assembling this page. They are recorded
because each is a reusable trap, not because they are interesting individually.