Organelles

Bottom line: this project sets how reviews should treat cellular-component
(CC) annotations to organelles. The first question was whether an annotation to
a root organelle term (GO:0005739 mitochondrion, GO:0005783 endoplasmic reticulum, GO:0005777 peroxisome, ...) is ever informative for a
membrane-bounded organelle, given that every protein in one sits either in a
membrane or in a lumen. The decision (2026-10-16) is that the organelle is
informative, and the call depends on the evidence. A root term from an
experiment that could not resolve the sub-compartment is ACCEPTed. A root term
from synthesis (IBA, IC, IEA and the other propagation codes) is a red flag
that the review must examine. Either way, a review should boldly propose the
membrane, lumen or matrix term when the gene has none and several lines of
evidence (a membrane domain, a targeting signal, a known function) point to
one. The red-flag worklist is
root_organelle_red_flags.md. In the
2026-10-18 run it lists 1,081 synthesis rows for mitochondrion that were not
REMOVEd, 283 of them in reviews that give the gene no mitochondrial
sub-compartment.

The question

Is it ever informative to annotate at the root organelle term, e.g.
mitochondrion? Don't you always know if it's membrane or lumen, at least for
membrane-bounded organelles?

The topology argument is sound. A membrane-bounded organelle splits cleanly
into its membrane(s) and the aqueous space(s) they enclose, and GO models this:
mitochondrial outer membrane, mitochondrial intermembrane space,
mitochondrial inner membrane and mitochondrial matrix are all part_of
mitochondrion. A protein in the mitochondrion is in at least one of them.

Whether you know which one is a separate question, about evidence rather than
biology. Below are the cases where the root may be the honest answer, from most
to least defensible.

The assay cannot resolve the sub-compartment. Organelle proteomics
(HDA/HTP, e.g. purified-mitochondria mass spectrometry), light-microscopy
co-localisation with a whole-organelle marker, and GFP-fusion screens report
"in the organelle" and nothing more. GO practice is to annotate to the
precision of the evidence, so for that row the root is correct. What we
still have to decide is whether the review should leave it there when
other knowledge (a cleavable matrix-targeting sequence, a TM segment, the
known topology of the pathway) settles the question.
The protein is in more than one sub-compartment. Examples are proteins
spanning the inner and outer membranes (contact-site and import-supercomplex
components), proteins dual-targeted to matrix and intermembrane space, and
proteins that change compartment as they are imported. The root then
summarises a union. The alternative is to list each sub-compartment, which
loses nothing.
Peripheral association with the cytosolic face. A cytosolic protein
recruited to the outer membrane (mitophagy adaptors, innate-immunity
signalling proteins on stressed mitochondria) is often annotated to
mitochondrion. The precise place is the outer membrane (cytosolic side).
GO's intrinsic-component terms are obsolete, and the membrane terms do not
encode sidedness, so curators may avoid a membrane term to avoid implying
that the protein is integral to it.
The organelle is a series, not a single compartment. For endosome
(early, late, recycling, multivesicular body), vacuole and autophagosome
(phagophore, autophagosome, autolysosome) the unresolved dimension is often
the maturation stage, not membrane versus lumen.
The nucleus is a special case. It is membrane-bounded, but nearly all
nuclear proteins sit in the lumen, and what matters is the internal
structure (chromatin, nucleolus, speckles, nuclear pore), not membrane
versus lumen. For a transcription factor the useful refinement is chromatin
(or nucleoplasm as a fallback), not the envelope.

Cases 2 and 3 give no reason to keep the root term as a summary: there is a
precise answer and GO can say it. Case 1 is the real reason root terms exist.
It concerns evidence, so it belongs on evidence rows. Cases 4 and 5 are
separate questions that should get their own sections later.

First survey of the corpus

root_organelle_usage.md is generated by
ORGANELLES/scripts/root_organelle_usage.py. Its per-row data,
ORGANELLES/root_organelle_usage.yaml, is a local artifact (git-ignored, 1.6
MB); re-run the script to get it. The script scans every
*-ai-review.yaml, takes each non-negated existing annotation to one of ten
root organelle terms, and asks whether the same review also has a more specific
part of that organelle (an is_a/part_of descendant, from GO via OAK). The
descendant can be another kept annotation, a proposed replacement or a
core_functions location. Run of 2026-10-18 (after merging main; the
generated files are authoritative):

organelle	rows	genes	genes also given a sub-compartment	rows with only the root
mitochondrion	3037	1051	702 (67%)	720
nucleus	5657	2076	829 (40%)	2587
endoplasmic reticulum	1103	584	401 (69%)	282
Golgi apparatus	493	311	172 (55%)	194
peroxisome	317	121	53 (44%)	123
lysosome	280	118	75 (64%)	73
endosome	198	131	86 (66%)	71
chloroplast	144	98	21 (21%)	95
autophagosome	88	50	16 (32%)	48
vacuole	29	25	7 (28%)	20

What the survey showed (first run, 2026-10-16; counts in these bullets are
from that run, and the proportions are the same in the later one):

When a sub-compartment is known, the root is still ACCEPTed. For
mitochondrial genes that also have a sub-compartment, 1,576 of 2,134 root
rows are ACCEPT, 438 KEEP_AS_NON_CORE, and only 43 MODIFY. No review uses
TOO_GENERAL for a root organelle row, and under the policy below none should.
The pattern is the same for nucleus (2,576 ACCEPT) and ER (694 ACCEPT).
Membrane and lumen split as expected. The mitochondrial genes divide into
membrane (284), lumen (239) and both (108). ER genes are overwhelmingly
membrane (310 of 398), Golgi genes are membrane or "other" (mostly
sub-cisternae and vesicles), and nucleus genes are lumen or other (chromatin).
This fits the topology argument.
Rows with only the root term are mostly high-throughput or electronic.
Of the 664 mitochondrial rows with only the root term, 180 are HDA, 148 IEA
and 90 IBA, and 97 are IDA. Many look like case 1. Some look like plain
curation gaps, where a known matrix enzyme or tRNA synthetase has no matrix
term (candidates to check: AARS2, CLYBL, PYCR2, MMADHC). Others look like
cases 2 and 3: proteins with two locations, or proteins recruited to
mitochondria under stress (candidates: AMACR, IRF3, RIPK1, NLRP3, PARP1).
Chloroplast is mostly predictions. 102 of 144 rows are IEA or ISM, and
26 have already been REMOVEd. The plastid question is mostly about targeting
predictions, not sub-compartments.

These are counts, not judgments. The "candidates" above are genes to examine,
not findings.

Policy

Decided 2026-10-16, revised 2026-10-18 (add the sub-compartment alongside the
root rather than substituting it). It applies to the first-pass organelles (mitochondrion, ER,
Golgi apparatus, peroxisome, lysosome); the nucleus, the series organelles and
plastids come later.

The root organelle term is informative. Never use TOO_GENERAL for it.
Knowing that a protein is in the mitochondrion is a real claim even when a
sub-compartment is also known. The term is not made redundant by its
descendants in the way that, say, catalytic activity is.
Direct evidence (IDA, EXP, IMP, HDA, HTP, ...): if the assay cannot
resolve the sub-compartment (organelle proteomics, whole-organelle
co-staining, GFP fusion), the row is ACCEPT (or KEEP_AS_NON_CORE if
the location is peripheral to the gene's function). It says what the
experiment showed. If the paper itself resolved a sub-compartment and
the curator stopped short, keep the row and add the sub-compartment as a
NEW row under rules 4-5.
Synthesis (IBA, IC, IEA, ISS, ISO, ISA, ISM, RCA, ...): a red flag.
Examine the row rather than accepting it by default. A propagation or an
inference has no assay limit to excuse the coarse term. Either the source
knew the sub-compartment and the term lost it, or the source did not know
the location at all. For IBA, check what the PAINT node's descendants
actually show. For IEA, check the mapping (UniProt subcellular-location
keyword, InterPro2GO, ARBA) and whether the gene really has the feature
the mapping assumes. If the organelle claim is right, the row is ACCEPT
(or KEEP_AS_NON_CORE) with a reason that records the check, even when
the source could not see the sub-compartment: the row's claim is the
organelle, and that claim is true. If the organelle itself is not
supported, the row is MARK_AS_OVER_ANNOTATED or REMOVE. The
sub-compartment never replaces the root on such a row; it is added under
rule 4.
Be bold about the sub-compartment. If the gene has no annotation to a
sub-compartment of the organelle and several lines of evidence agree, add
one as a NEW row (usually IC), alongside the root rows. Lines of evidence include
predicted TM segments or a membrane anchor, a cleavable matrix-targeting
sequence or a PTS1/PTS2, an ER signal peptide with or without a retention
motif, and a function that only makes sense on one side (a matrix TCA
enzyme, an inner-membrane carrier). Record the lines of evidence in
review.reason. If they disagree or there is only one, stay at the root.
Add, don't substitute. This follows GO's own practice for matrix
proteins with sub-mitochondrial evidence. DARS2 and MMUT carry
GO:0005759 mitochondrial matrix and keep their mitochondrion rows
(IBA, IDA, HTP, TAS) alongside it. It also keeps the review judging claims
rather than evidence codes: every row on the root term gets the same
kind of action, so the inconsistent-review-actions check needs no
exception. (A first version of the pilot MODIFYed the synthesis rows and
added a validator exemption for that split; both were dropped in review
of PR #5208.) For the IEA rows, the lasting fix is upstream: a UniProt
location keyword of SL-0170 (Mitochondrion matrix) instead of SL-0173
(Mitochondrion) maps to the matrix term directly.
core_functions.locations should name the sub-compartment where the
core activity is shown to happen, and may also carry the root if the
organelle-level claim is independently supported. A well-supported
sub-compartment location is not by itself where the function acts (MMADHC,
2026-10-18).
ER and secretory-pathway rows (settled in ER batches 1-4, 2026-11-04
to 2026-11-06):
Transit is not residence. Every secreted, plasma-membrane and
lysosomal protein passes through the ER. A biosynthetic ER pool stays
KEEP_AS_NON_CORE when there is direct evidence on the gene itself or
the endogenous protein is seen in its native cell; a synthesis row whose
source shows only transit, or retention of an overexpressed or mutant
protein, is MARK_AS_OVER_ANNOTATED.
Topology does not force a sub-compartment on a transit pool. A
membrane protein that only passes through the ER gets no ER membrane row.
Membrane and lumen lines. A tail anchor or TM without a signal
peptide, plus one ER-specific assay on the endogenous protein (a
marker-controlled fraction, an ER Ca2+ function, ER-membrane partners),
supports GO:0005789; a predicted TM with only an overexpression assay
does not. A soluble signal-peptide protein that is retained, by a
KDEL-type motif or shown by protease protection, supports GO:0005788.
The ER exit site (GO:0070971) needs a marker of transitional ER.
Stable tethering by an integral partner (added 2026-11-11, UFSP2) is
a third route to a sub-compartment row, for a soluble protein with no TM
or signal peptide. It needs (a) direct evidence that the endogenous
protein is at the organelle, (b) a named integral membrane protein of the
organelle that binds it, (c) dependence on that partner: losing it
(knockout or a binding mutant) releases the protein or destabilises it,
and (d) cytosolic synthesis with no signal peptide and no organelle
targeting or transit sequence, so the cytoplasmic leaflet is the only face
it can reach. Check (d) against the literature as well as UniProt
features: non-cleavable import signals (PTS1/PTS2, internal carrier
signals, twin-CX9C motifs) have no transit feature. (a) carries the location and (b) names what holds it
there; stability alone shows only an obligate partnership. A protein that
fails (d), such as a mitochondrial matrix protein held by an inner-membrane
partner, gets the side term for its own compartment instead (e.g.
GO:0099617 matrix side of mitochondrial inner membrane). Add the
cytoplasmic-side term as a NEW IC row beside the root: GO:0098554 for
the ER, whose definition includes peripherally associated proteins. GO:0098554 is part_of GO:0005789, so it is the
more specific statement, not an alternative to it. A NEW GO:0005789
row that we author still needs the protein's own TM or tail anchor; an
existing direct-evidence GO:0005789 row on a tethered protein (UFL1,
CDK5RAP3) stays ACCEPT and can be refined by adding the cytoplasmic-side
row. Brief or regulated recruitment (signalling, stress, autophagy,
Rab-dependent docking, a cycling complex such as SRP) is not tethering.
The same test will apply to the other organelles, but GO has no
cytoplasmic-side term for the Golgi stack or the peroxisome (only the
trans-Golgi network, GO:0160281, the lysosome, GO:0098574, and the
mitochondrial outer membrane, GO:0032473). A Golgi or peroxisome case
raises a proposed_new_terms entry rather than being forced onto a
narrower term.
Recruited cytosolic pools stay at the root, core only when the
function is shown at the ER. Membership of an ER-acting complex supports
the organelle root for an obligate subunit but never localises it to a
sub-compartment, and it is not core when the subunit's own data place it
elsewhere (TRAPPC4).
A new sub-compartment row refines generic rows: the gene's
GO:0016020 membrane rows become MODIFY to it, all rows on the term
alike.
Microsomes are not ER without an ER marker, and an isoform-restricted
UniProt ER location on a GO row with no isoform is a source_defects
entry.
Unreadable sources: a curated human ECO:0000269 line from a paper
we cannot read is kept on curator deference; a projection whose donor
paper is also unreadable is UNDECIDED.

Survey under this policy

From the 2026-10-18 run, mitochondrion only, leaving out REMOVEd rows:

evidence class	rows, gene also has a sub-compartment	rows, root only
direct	1,433	365
synthesis	798	283

The 283 root-only synthesis rows (main codes: 149 IEA, 76 IBA, 37 ISS, 15 ISO) are the
first worklist. 143 are ACCEPT and 87 KEEP_AS_NON_CORE; under rule 3 each
needs a reason recording the check, and for many of these genes rule 4
should add a sub-compartment row.
The 798 synthesis rows in reviews that do have a sub-compartment (main
codes: 361 IEA, 333 IBA) are lower priority. The fine-grained location is already recorded,
so the question is only whether the propagated organelle claim stands.
The 365 root-only direct rows are mostly fine under rule 2, but rule 4
still applies to the gene: AARS2, CLYBL, PYCR2 and MMADHC are among the
candidates where function suggests the matrix.

Open questions

is_active_in on root rows. Some IBA rows carry is_active_in GO:0005739, which says the gene's activity happens there, not only that
the protein is present. In batch 2 five GO:0005739 rows carry it: the IBA
rows of ACACA, ADO, ARAF and BRAF, and ABCC9's Ensembl IEA. For ADO and
BRAF the review's own reason argues the activity is elsewhere (ADO's
N-degron oxidation is cytosolic; BRAF's mitochondrial pool has no assigned
activity). ARAF shares BRAF's PAINT node (PTN000584576) and is the arguable
case: its reason does give it a mitochondrial role, MST2 sequestration,
though that is not among its annotated functions. ABCC9's short SUR2
variants have a real mitochondrial function, and ACACA's row is already
over-annotated. The policy so far judges the term, not the qualifier. A
later pass should check is_active_in rows and propose located_in where
the activity is not shown in the organelle.
Fix the IEA rows upstream. The root-only IEA rows on AARS2, CLYBL and
MMADHC come from UniProt subcellular-location keyword SL-0173
(Mitochondrion). Genes whose entries use SL-0170 (Mitochondrion matrix),
such as DARS2, TARS2 and MMUT, get GO:0005759 automatically. A UniProt
feedback request to change the keyword would fix the IEA row at its source,
and the same applies to much of the 141 root-only IEA rows in the worklist.
Should validation warn about root-only synthesis rows that are ACCEPTed
with no reason that mentions the sub-compartment? Or should
core_functions.locations warn when it uses only a root term? Both checks
are mechanical, using the script's ancestor test.
How should we say "peripheral, cytosolic face" now that intrinsic-component
terms are obsolete? Is the bare membrane term enough, or is this a
proposed_new_terms or GO-ticket question? Answered 2026-11-11 for stably
tethered proteins: the "cytoplasmic side of X membrane" terms (e.g.
GO:0098554 for the ER) include peripherally associated proteins; see the
tethering bullet in rule 7. Briefly recruited proteins stay at the root.
For IBA red flags with a consistent pattern (a PAINT node at the root term
whose experimental descendants all show the same sub-compartment), should
we raise this with PAINT rather than fixing gene by gene?
Is there GO consortium guidance on CC annotation granularity to cite in
support of this policy?

STATUS

2026-10-16

[x] Create project page and survey script
[x] First corpus survey of root organelle terms (10 organelles)
[x] Decide policy: evidence-based; no TOO_GENERAL; synthesis is a red flag;
be bold about sub-compartments (see Policy)
[x] Red-flag worklist generated (root_organelle_red_flags.md)
[ ] Work through the root-only mitochondrial synthesis rows (285 in the
2026-10-20 run), starting with human; apply rules 3 and 4
[x] Batch 1: 20 human genes (2026-10-20)
[x] Batch 2: 20 human genes (2026-10-24)
[x] Batch 3: 20 human genes (2026-10-25)
[x] Batch 4: the last 24 human genes (2026-11-03); human root-only list done
[x] Pilot rule 4 on AARS2, CLYBL, PYCR2, MMADHC (3 matrix, 1 unresolved)
[x] Revise to "add, don't substitute" after PR #5208 review: synthesis
root rows ACCEPT, sub-compartment as a NEW row; validator exemption
dropped
[ ] Check the rest of the 365 root-only direct mitochondrial rows for rule 4
candidates
[ ] Look up GO consortium CC annotation guidance on granularity and sidedness
[ ] Decide on a validation warning
[x] Add the policy to CLAUDE.md (2026-10-25, after batches 1 and 2); the
annotation-reviewer guidance still to do
[ ] Repeat for ER, peroxisome, lysosome and Golgi
[x] ER batch 1: 20 human genes (2026-11-04); 60 human genes remain on the
root-only ER list (survey regenerated 2026-11-04)
[x] ER batch 2: 20 human genes (2026-11-05); 40 human genes remain (39
distinct, since GBA duplicates GBA1)
[x] ER batch 3: 20 human genes (2026-11-05); 20 human genes remain (19
distinct)
[x] ER batch 4: the last 19 human genes (2026-11-06); human root-only ER
list done except GBA, a duplicate folder of GBA1
[ ] Separate sections for nucleus, endosome and autophagy series, and plastids

NOTES

2026-11-11 (tethering route added to rule 7)

Settled the open question from ER batch 4: should a soluble protein stably held
on the organelle by an integral-membrane partner get a sub-compartment row? Yes,
the cytoplasmic-side term, under a new bullet in rule 7. GO:0098554 cytoplasmic
side of endoplasmic reticulum membrane is defined to include proteins
"peripherally associated with" the cytoplasmic leaflet (GO definition: "The leaflet
of the endoplasmic reticulum membrane that faces the cytoplasm, including any
protein embedded in, attached to, or peripherally associated with it."). It is
part_of GO:0005789,
so it says more than the bare membrane term (which side, and that the attachment
is peripheral) rather than avoiding it. The route needs direct evidence that the
protein is at the organelle, a named integral partner, dependence on it, and
cytosolic synthesis with no signal or targeting sequence (which fixes the side),
checked both in the UniProt features and in the literature (for UFSP2, worm ODR-8 is
diffusely cytosolic without its partner); a
stability result alone shows an obligate partnership, not a location.

UFSP2 gains a NEW GO:0098554 IC row. It has no TM or signal peptide; its
partner ODR4 is a transmembrane ER protein; endogenous UFSP2 is lost in ODR4
knockout cells, which phenocopy UFSP2 loss (PMID:35926457); and it deUFMylates
uL24 on membrane-bound 60S subunits (PMID:38383785). The ODR4-UFSP2 complex at
the ER membrane was first shown in worm and human proteins by Chen et al.
(PMID:24603482), including a partner-dependent release test: worm ODR-8 washes
out of digitonin-permeabilised HeLa cells alone but is partly retained with
ODR-4, which also places it on the cytosolic face. The 2022 and 2024 papers cite
that work for the anchoring. The
location itself rests on the curated IDA row (PMID:35926457). The three root
rows stay ACCEPT and the second core function names the cytoplasmic side.
What this means for the UFMylation machinery. UFL1 and CDK5RAP3 are also
soluble and held at the ER by DDRGK1, and both already carry direct-evidence
GO:0005789 rows ACCEPTed in their reviews. Those rows stay: they are true,
because the cytoplasmic side is part of the ER membrane, and an assay on the
endogenous protein cannot be second-guessed by topology. A cytoplasmic-side row
would refine them and is left for a later pass. DDRGK1 has its own TM segment,
so GO:0005789 is the right term for it.
Checked and left at the root. The other ER-batch reviews whose reasons
mention tethering or peripheral pools do not meet the test. UVRAG reaches the ER
through PtdIns(3)P and RINT-1, a peripheral complex, and leaves on autophagy;
USO1 is Rab1-recruited; the TRAPP subunits and SRP72 belong to complexes that
cycle on and off the membrane; UFM1 is a modifier whose ER pool is its
conjugates; KIFAP3, CAVIN1 and DLG4 have no named integral ER partner.

2026-11-06 (ER batch 4: the last 19 human genes)

The last 19 human genes on the root-only ER list (SMPD2 to ZDHHC23); with GBA
left to the duplicate-folder cleanup, this finishes the human ER list. Two gained
a sub-compartment term (SMPD2 ER membrane, TRAPPC3 ER exit site) and three
changed action (STAU2, TP53, TRAPPC4, UFM1). The same PR carries the optional SCP2 and
PGRMC2 wording fixes from the ER batch 3 review and corrects NRXN1's mouse donor
claim on three non-ER rows.

gene	outcome	evidence
SMPD2	root; NEW GO:0005789 ER membrane (IC)	two predicted TM helices whose removal abolishes activity; overexpressed nSMase1 with ER and Golgi markers (PMID:10681556); endogenous phospho-nSMase1 co-stains with calnexin (PMID:25168245)
SPG11	root	endogenous spatacsin fractionates with REEP5 and STIM1 and resists extraction (PMID:37871017), but no TM is predicted, so no membrane row
SPHK2	root	serum starvation raises the ER share of overexpressed SphK2 (PMID:16118219); soluble kinase, recruited pool
SQSTM1	root	curated from an abstract-only TRIM13 paper (PMID:22178386), curator deference; consistent with its mitochondrion rows
SRP68	root; rows stay ACCEPT	tagged SRP68 co-stains with an ER marker and a point mutant loses it (PMID:28369529); core because SRP hands over at the ER
SRP72	root; rows stay ACCEPT	GFP-SRP72 co-localises with ER-Tracker (PMID:28369529); soluble SRP subunit on the cytosolic face
STAU2	root; IEA KEEP_AS_NON_CORE to UNDECIDED	by similarity only; mouse donor's abstract calls wild-type Stau2 mostly cytosolic (PMID:15166236)
TMF1	root; rows stay MARK_AS_OVER_ANNOTATED	the sole mouse source calls TMF/ARA160 "a Golgi-associated protein" (PMID:23000399); human TMF1 at Golgi cisternal tips (PMID:17698061)
TP53	root; IEA ACCEPT to KEEP_AS_NON_CORE	Synoviolin sequesters p53 at the ER (PMID:17170702); wild-type p53 at ER and MAM binding SERCA (PMID:25624484); recruited pool of a nuclear factor
TRAPPC1	root; row stays ACCEPT	TRAPP acts at ER exit sites or ERGIC (PMID:21525244); support is at complex level
TRAPPC3	root; NEW GO:0070971 ER exit site (IC)	endogenous mammalian Bet3 on Sec31-marked transitional ER, 7.6-22-fold enriched there by immuno-EM (PMID:16880271); binds SEC23
TRAPPC4	root; IEA and ISS ACCEPT to KEEP_AS_NON_CORE	mouse ER line is a curator inference; its own synbindin immuno-EM shows Golgi and vesicles, not ER (PMID:11018053); the ER is plausible only through TRAPP membership
TRAPPC5	root	by similarity with no named source; holds through TRAPP membership (PMID:16880271)
UFM1	root; IEA and ISS KEEP_AS_NON_CORE to ACCEPT	endogenous UFM1 partly in the BiP-marked ER fraction with UFBP1 (PMID:21494687); UFMylated RPL26 on ER-bound ribosomes (PMID:30626644)
UFSP2	root	ER localization depends on the TM protein ODR4, and ODR4 loss phenocopies UFSP2 loss for RPL26 deUFMylation (PMID:35926457)
USO1	root	fly p115 on tERs, Golgi and ER cisternae (PMID:12876273); soluble peripheral protein
UVRAG	root	"depends on PtdIns(3)P for its ER localization" in the RINT-1 tether (PMID:24056303); recruited C2-domain protein
VEGFA	root	ER line is for the L-VEGF189 isoform (PMID:11731620), curator deference; secreted, no retention signal
human/ZDHHC23	root; IBA stays MARK_AS_OVER_ANNOTATED	ZDHHC23 itself is at the Golgi and TGN (PMID:22399288); the ER evidence comes from paralogs

What ER batch 4 showed:

A new sub-compartment can be the exit site. TRAPPC3's immuno-EM puts it
on Sec31-marked transitional ER, so GO:0070971 is the right sub-compartment,
not membrane or lumen. TRAPPC1, TRAPPC4 and TRAPPC5 stay at the root: complex
membership alone does not localise a subunit.
Tethered without a TM. UFSP2 has no anchor of its own but is held at the
ER by ODR4; that is a location, not a sub-compartment, so it stays at the
root.
Paralog evidence behind an IBA is not the gene's own. ZDHHC23's ER
claim rests on other ZDHHC enzymes; its own data say Golgi.
is_active_in: UFSP2 (IBA and IDA) and SPG11 IEA are justified; SMPD2
and USO1 IBA are soft; TMF1 and ZDHHC23 IBA are not supported.

2026-11-05 (ER batch 3: 20 human genes)

The next 20 human genes on the root-only ER list (LOXL2 to SLC22A5). Three
gained an ER lumen row (POFUT1, PRDX4) or ER membrane row (PGRMC2); four
changed action (LRP2, NENF, PARK7, RIPK2). Four of the genes (NENF, NLRP3,
PYCARD, SCP2) had their mitochondrion rows reviewed in mitochondrion batch 4,
and the ER calls were kept consistent with them.

gene	outcome	evidence
LOXL2	root	endogenous LOXL2 mainly in calnexin-positive ER, read by the authors as precursor awaiting export (PMID:28332555); soluble and secreted
LRP2	root; IEA KEEP_AS_NON_CORE to UNDECIDED	sole mouse donor is a cubilin paper, abstract-only, with no ER result (PMID:15616221)
LRP5	root	curated from an abstract-only variant paper (PMID:25920554); MESD is its ER chaperone (PMID:12581525); receptor in transit
LRP6	root	GRP94-dependent ER exit (PMID:23572575) and MESD chaperoning show a wild-type transit pool; the curated line comes from a variant paper (PMID:26387593)
human/MTMR9	root	"Partially localizes to the endoplasmic reticulum" (PMID:19038970), curator deference; soluble pseudophosphatase
NENF	root; IBA, IEA and IDA ACCEPT to KEEP_AS_NON_CORE	endogenous protein in a marker-controlled ER fraction (PMID:31536960); secreted, no retention motif, so a transit pool
NLRP3	root	mouse NLRP3 on the ER with ASC brought from mitochondria (PMID:23502856); soluble NLR bound to membranes by a polybasic linker (PMID:34861190)
NPHS2	root; isoform `source_defects`	ER location is for the Pod-short isoform retained in the ER (PMID:23648087); canonical podocin is at the plasma membrane
NRXN1	root; ISS stays MARK_AS_OVER_ANNOTATED	mouse source shows PDZ-dependent ER/Golgi exit and inactive retention signals (PMID:19036990)
NT5C3A	root	interferon-induced p36 co-stains with PDI (PMID:8557639); one line toward membrane, so no membrane row
PARK7	root; IEA and IDA ACCEPT to KEEP_AS_NON_CORE	endogenous DJ-1 in a marker-controlled ER fraction (PMID:31536960); soluble, surface-associated pool
human/PCK1	root	AKT-phosphorylated PCK1 moves to the ER (PMID:32322062); recruited cytosolic enzyme
PGRMC2	root; NEW GO:0005789 ER membrane (IC)	single predicted TM and no signal peptide; endogenous staining of central ER and nuclear envelope in three cell types (PMID:27754849), no named ER marker
PINK1	root	endogenous PINK1 in marker-controlled ER fractions (PMID:31536960); the IEA's ARBA source is no longer in UniProt
POFUT1	root; NEW GO:0005788 ER lumen (IC)	signal peptide, no TM, C-terminal RDEF retention motif; "soluble protein ... retained in the ER by a KDEL-like sequence" (PMID:15653671)
PRDX4	root; NEW GO:0005788 ER lumen (IC)	cleaved signal peptide; protease-protected like calnexin's lumenal domain and retained without a motif (PMID:18052930); removes Ero1-derived H2O2 (PMID:20627953)
PYCARD	root	activation-induced move to perinuclear ER/mitochondria clusters (PMID:21124315); recruited adaptor
RIPK2	root; EXP and IEA ACCEPT to KEEP_AS_NON_CORE	endogenous RIPK2 in a PDI-positive ER fraction where ZNRF4 degrades it (PMID:28656966); a regulatory recruited pool
SCP2	root; isoform `source_defects`	mouse source is overexpressed SCP-2 in L-cells (PMID:11003606); soluble PTS1 protein
SLC22A5	root; isoform `source_defects`	ER location is for the OCTN2VT splice form retained in the ER (PMID:17509700)

What ER batch 3 showed:

Retention without a motif. PRDX4 has no KDEL-type signal, but protease
protection, topology and a lumenal partner (Ero1) agree on the lumen. A
retention motif is one line, not a requirement.
When the sub-compartment enters core_functions. Only when an assay
places the activity itself there, not when it is inferred from the protein's
location (rule 6). PRDX4 qualifies: it is shown consuming H2O2 made by the
lumenal oxidase Ero1. POFUT1 does not yet: O-fucosylation is shown in the ER,
so its core location stays at the root even though its NEW lumen row is
well supported.
A new membrane row refines generic membrane rows. Once PGRMC2 has
GO:0005789, its GO:0016020 membrane rows are correct but too general, so they
become MODIFY to the ER membrane term rather than staying ACCEPT.
Isoform-restricted ER locations are common. NPHS2, SCP2 and SLC22A5 join
ABCA7 and ATP7A: UniProt ties the ER to one splice form, often a retained
quality-control product, while the GO rows carry no isoform. Each now has a
source_defects entry.
Shared fractionation papers need one reading. PMID:31536960 (marker-
controlled mito, ER and MAM fractions) underlies NENF, PARK7 and PINK1. All
three are read as real but non-core ER pools, with no MAM row.
The membrane bar. PGRMC2's membrane row rests on a predicted TM plus
endogenous ER-pattern staining without a marker. It is the weakest membrane
row so far, above CGRRF1 (overexpressed construct) and below C18orf32 and
C19orf12 (marker-controlled fractions).
is_active_in: POFUT1 IBA (O-fucosylation shown in the ER) and PRDX4
IBA (peroxidase acting on Ero1 H2O2) are justified; PGRMC2 IBA is soft.
For later checks: NRXN1's ISS rows on GO:0031965, GO:0031982 and
GO:0043025 repeat a wrong claim that mouse Nrxn1 has no experimental
annotation.

2026-11-05 (ER batch 2: 20 human genes)

The next 20 human genes on the root-only ER list (C18orf32 to KIFAP3). GBA was
skipped: genes/human/GBA and genes/human/GBA1 are two folders for the same
accession (P04062), and only GBA1 was reviewed, so KIFAP3 made up the 20. Two
genes gained an ER membrane row (C18orf32, C19orf12) and three changed action
(CHRNA3, HMGB1, KEAP1). The brief carried the batch 1 rules, and the six
optional review findings on batch 1 were fixed in the same PR (caveats on the
BCL2L1 and BAK1 membrane rows, AHSA1's alternative stated, APEX1's quote
extended, and isoform source_defects on ABCA7 and ATP7A).

gene	outcome	evidence
C18orf32	root; rows stay ACCEPT; NEW GO:0005789 ER membrane (IC)	N-terminal hydrophobic anchor necessary and sufficient for ER and lipid-droplet targeting; endogenous protein only in ER-enriched and buoyant fractions; membrane ERAD by gp78/derlin-1 (PMID:29275994)
C19orf12	root; rows stay ACCEPT; NEW GO:0005789 ER membrane (IC)	one predicted TM and no signal peptide; endogenous protein in a marker-controlled ER fraction (PMID:26136767); MAM not asserted
CAVIN1	root; row stays UNDECIDED	UniProt ER line by similarity to rat, which rests on a curator inference; mouse source paper is abstract-only and about Fam198a export (PMID:30188967)
CERKL	root	all ER data on overexpressed tagged constructs (PMID:19158957); soluble protein
CGRRF1	root; rows stay ACCEPT	partial PDI overlap of overexpressed V5-tagged protein (PMID:27485036); TM predicted only, so no membrane row
CHRNA3	root; IEA and ISS ACCEPT to KEEP_AS_NON_CORE	ERAD extraction of alpha3 by p97 shows a real quality-control pool (PMID:26265139); the channel acts at the plasma membrane
DDX41	root	soluble helicase co-fractionating with STING, microsomes and MAM (PMID:21892174); recruited pool
DLG4	root; rows stay UNDECIDED	sole mouse source is an abstract-only knockout paper that never shows the PSD-95 experiment (PMID:16814779)
DNASE1L3	root	endogenous protein in the ER of HT-29 cells, moving to the nucleus in apoptosis (PMID:23229555); secreted and no retention motif
FGFR3	root	direct immunofluorescence plus curated ER N-glycosylation of the wild type; type I receptor in transit
FGFR4	root	core-glycosylated FGFR4 held by ER-resident beta-Klotho (PMID:20683963); transit pool with direct evidence
GBA1	root	mouse immunogold with PGRN in ER and lysosome (PMID:27789271); GCase/LIMP-2 complex forms in the ER (PMID:40159502); soluble hydrolase in transit
GLUD1	root	overexpressed hGDH1-EGFP partly overlaps an ER marker (PMID:19448744); mitochondrial transit peptide
HADHB	root	UniProt curated from an abstract-only paper whose abstract shows viperin, not HADHB, at the ER (PMID:21527675); deferred
HMGB1	root; IEA and IDA ACCEPT to KEEP_AS_NON_CORE	SARS-CoV-2 ORF3a recruits HMGB1 to the ER (PMID:35239449); leaderless protein, recruited pool; the old reason's "secretory pathway" claim was wrong
HSPA13	root; row stays ACCEPT	ER-luminal proximity labelling and translocon binding show residence (PMID:36244454); signal peptide predicted only and no retention motif, so no lumen row
IFIT2	root	UniProt curated from PMID:21190939, whose localization data could not be retrieved; soluble TPR protein
KCNRG	root; rows stay ACCEPT	"an ER-associated protein" retaining Kv1 channels (PMID:19968958, abstract only); PAINT node seeded by KCNRG itself
KEAP1	root; IEA KEEP_AS_NON_CORE to UNDECIDED	sole mouse source's abstract calls Keap1 cytoplasmic (PMID:14517290); the old Nrf1 rationale is refuted (PMID:16687406)
KIFAP3	root	mouse immuno-EM of endogenous KAP3 along ER and Golgi membranes (PMID:19384852); soluble kinesin-II subunit

What ER batch 2 showed:

The batch 1 rules held. Transit pools with direct evidence stayed
non-core (FGFR3, FGFR4, GBA1, DNASE1L3), and no transit pool got a
topology-forced membrane row.
The membrane line needs more than a predicted TM. C18orf32 and C19orf12
have an ER-specific assay on the endogenous protein plus an anchor; CGRRF1
has only a predicted TM and an overexpression assay, so it stays at the root.
ER does not mean "secretory". HMGB1 is secreted without passing through
the ER; its pool is recruited by a viral protein.
Unreadable sources: defer to a curated human line, otherwise UNDECIDED.
When the human UniProt entry carries a curated ECO:0000269 ER line from a
paper we cannot read, the row stays KEEP_AS_NON_CORE on curator deference
(HADHB, IFIT2). When the row is only an ortholog projection whose donor
paper also cannot be read, it is UNDECIDED (KEAP1, DLG4, CAVIN1). Either
way the reason says which result would settle it.
is_active_in: one ER row carries it (KCNRG IBA), defensible because
the channel-retention activity is proposed to act at the ER.
Duplicate folder. genes/human/GBA duplicates genes/human/GBA1
(same accession); it should be retired in a separate change. Done in
#6088, which removes the GBA folder and points its history at GBA1.

2026-11-04 (ER batch 1: 20 human genes)

The first endoplasmic reticulum (GO:0005783) batch: the first 20 human genes
alphabetically (A1CF to BRSK2) on the regenerated root-only ER list, which now
has 80 human genes. Four gained a sub-compartment term: AGR3 the ER lumen,
and BAK1, BCL2L1 and BNIP3L the ER membrane. Five rows changed action. The
brief adapted the mitochondrion policy to the ER: biosynthetic transit is not
residence, retention motifs are positive evidence, topology decides membrane
vs lumen, microsomes are not ER without a marker, and overexpressed
constructs can be retained artefactually.

gene	outcome	evidence
A1CF	root	UniProt by similarity from rodent immuno-EM: a minor pool on the cytosolic face of rough ER (PMID:12881431)
ABCA4	root	wild-type in calnexin-positive vesicles with "some" ER staining; ER retention is strongest for misfolded mutants (PMID:29847635)
ABCA7	root	ER location is for isoform 2 only, which lacks TM1; canonical type I is at the plasma membrane (PMID:14592415)
ADIPOQ	root	endogenous protein partly overlaps GRP94 in 3T3-L1 adipocytes and insulin speeds its ER export (PMID:10444069); transit pool, no retrieval motif
AGR3	NEW GO:0005788 ER lumen (IC)	cleaved signal peptide and no TM; C-terminal QSEL motif whose deletion sends it to the Golgi (PMID:18086916)
AHCY	root	overexpressed mCherry fusion against GFP-calreticulin only (PMID:28647132)
AHSA1	root	curated from an abstract that calls the transient ER association "preliminary" (PMID:11554768)
AIMP1	root	HeLa staining with calnexin and GM130 markers (PMID:17525271); route to luminal gp96 unresolved
AMBRA1	root; rows stay ACCEPT	moves from dynein to calnexin-positive ER on starvation, where it nucleates autophagosomes (PMID:20921139); recruited, so no membrane term
APEX1	root	SET complex "associated with the endoplasmic reticulum" (PMID:12524539); minor pool of a nuclear enzyme
ASB11	root; rows stay ACCEPT	ER-associated E3 whose substrates (ribophorin 1, BIK) are ER membrane proteins (PMID:24337577, abstract only)
ATP1A1	root; ISS KEEP_AS_NON_CORE to MARK_AS_OVER_ANNOTATED	rat source paper shows assembly in transit and retention only when export is blocked (PMID:18768923)
ATP1A2	root; IEA KEEP_AS_NON_CORE to MARK_AS_OVER_ANNOTATED	mouse source: overexpressed wild-type "slightly" ER, FHM2 mutant retained and degraded (PMID:21731499)
ATP2B2	root; IEA UNDECIDED to MARK_AS_OVER_ANNOTATED	mouse source is a CHO overexpression paper: retained PMCA "normally inactive", no ER marker (PMID:17234811)
ATP7A	root	exon-10-skipped isoform without TM3-4 co-localises with BiP (PMID:9467005); quality-control retention of a natural splice form
BACE2	root	type I protease seen in its biosynthetic ER pool (PMID:11423558); no retention motif
human/BAK1	NEW GO:0005789 ER membrane (IC)	tail anchor without signal peptide; endogenous Bak in the calnexin-positive fraction free of COX IV and Tom40; ER-targeted Bak depletes ER Ca2+ (PMID:12847083, mouse)
BCL2L1	root; IEA ACCEPT to KEEP_AS_NON_CORE; NEW GO:0005789 ER membrane (IC)	tail anchor, cytoplasmic side; ER fractionation (PMID:17428862) and reticulon partners (PMID:11126360)
BNIP3L	root; NEW GO:0005789 ER membrane (IC)	tail anchor; endogenous cardiac NIX at ER/SR sets SR Ca2+ stores (PMID:19065046, mouse)
BRSK2	root	soluble kinase recruited to the ER under tunicamycin stress (PMID:22713462)

What ER batch 1 showed:

Where transit stops being a location. Direct evidence on the gene
itself (ABCA4, BACE2) or the endogenous protein in its native cell
(ADIPOQ) keeps a biosynthetic ER pool as KEEP_AS_NON_CORE. A synthesis row
whose source shows only transit, or retention of an overexpressed or
mutant protein (ATP1A1, ATP1A2, ATP2B2), is MARK_AS_OVER_ANNOTATED.
Topology does not force a sub-compartment on a transit pool. ABCA4,
ABCA7 and BACE2 would be in the ER membrane if they resided there, but they
only pass through, so no GO:0005789 row: a NEW row would claim residence.
Tail anchors are the ER counterpart of the TM line. A tail-anchored
protein with no signal peptide cannot be lumenal, so with one ER-specific
assay (a marker-controlled fraction, an ER Ca2+ function, ER-membrane
partners) it earns GO:0005789 (BAK1, BCL2L1, BNIP3L). No MAM was asserted;
none of the sources had a MAM marker.
Retention plus signal peptide means lumen. AGR3's KDEL-like motif,
shown necessary by deletion, plus a soluble signal-peptide chain is the
lumenal case.
Recruited cytosolic pools stay at the root, as on mitochondria (A1CF,
AHSA1, AMBRA1, APEX1, BRSK2). AMBRA1 stays core because its autophagy
function is shown at the ER.
is_active_in: three ER rows carry it. AGR3 IBA and ATP1A2 IEA have no
activity shown in the ER and join the open-question worklist; BNIP3L IBA is
defensible for its Ca2+-store role.
For the Golgi batches: ATP1A1's Golgi ISS row has the same rat source and
the same transit logic.

2026-11-03 (batch 4: the last 24 human genes)

The remaining 24 human genes on the root-only mitochondrion list (MAP2K1 to
TICAM1). None gained a compartment term; all 24 stay at the root. Eight
changed action: MAP2K1, MAP2K2, MAPK1, MAPK3, PNPO and SCP2 went up to
KEEP_AS_NON_CORE, and MOTS-c, humanin and NENF went down from ACCEPT to
KEEP_AS_NON_CORE. The brief carried the batch 3 lessons (source defects,
proteome rescue, contact sites) plus cautions for mitochondrially encoded
peptides and for signalling genes recruited to the surface.

gene	outcome	evidence
MAP2K1	root; IEA and TAS UNDECIDED to KEEP_AS_NON_CORE	mouse subfractionation puts MEK1/2 in outer membrane and IMS fractions (PMID:18545666); phospho-MEK1/2 in rat brain mitochondria (PMID:15030409); pan antibodies, no ER marker, recruited pool
human/MAP2K2	root; IEA and TAS UNDECIDED to KEEP_AS_NON_CORE	same evidence as MAP2K1; the antibodies cannot separate MEK2 from MEK1
MAPK1	root; IEA and TAS MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE	projection from mouse cardiac IDA (PMID:11884367); ERK1/2 in outer membrane and IMS fractions (PMID:18545666); transient pool
MAPK3	root; IEA and TAS MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE	~15% of ERK1 in purified HeLa mitochondria checked for ER markers, leaving on serum stimulation (PMID:19847302)
MBTPS1	root	fraction separated from ER and Golgi markers (PMID:35362222); proteinase K reading of matrix conflicts with ER/Golgi type I topology, one study
MLDHR	root; rows stay ACCEPT	evidence is on the uORF peptide itself (PMID:33406399), not PTEN; no sub-mitochondrial assay
MT-RNR1 (MOTS-c)	root; IDA and IEA ACCEPT to KEEP_AS_NON_CORE	organelle-level imaging of the endogenous peptide; its activities are nuclear and extracellular
MT-RNR2 (humanin)	root; 3 IDA and IEA ACCEPT to KEEP_AS_NON_CORE	organelle-level co-staining and uptake; activities are as an extracellular ligand and BAX sequestration in the cytosol
MYL10	root	IBA rests on a LIFEdb GFP fusion; mouse MitoCarta HDA is independent
NENF	root; IEA and IDA ACCEPT to KEEP_AS_NON_CORE	fractionation with mito, ER and MAM markers (PMID:31536960); soluble secreted protein, so no MAM membrane term
NLRP3	root	recruited via MAVS on activation; the authors call it a "mitochondrial/MAM fraction" (PMID:23582325)
PFDN2	root	UniProt curation from an abstract-only paper that names only URI; deferred; HPA is a second line
PFDN4	root	same URI paper (PMID:17936702) as PFDN2; deferred to the curator
PNPO	root; IBA and HTP MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE	MitoCoP (human, PMID:34800366) and MitoCarta3.0; PAINT seeds are plant and yeast mitochondrial orthologues; matrix vs IMS disagree
PYCARD	root	activation-induced move to perinuclear ER/mitochondria clusters, no MAM marker (PMID:21124315)
RRM2B	root; IEA stays MARK_AS_OVER_ANNOTATED	inferred from mtDNA replication; the enzyme supplies dNTPs from the cytosol (PMID:18997010)
SCP2	root; IEA, ISS and EXP MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE; source defect	EXP paper PMID:17157249 never mentions mitochondria (full text); pool stands on mouse PMID:26901662 and MitoCarta3.0
SHMT1	root; IBA stays MARK_AS_OVER_ANNOTATED; source defect	PAINT node seeded only by SHMT2-type enzymes; recorded as MISSING_NEGATION for GO_CENTRAL
SQSTM1	root	projection from mouse CCCP colocalisation (PMID:32047033); recruited mitophagy pool
STUB1	root	stress-induced pool in rat neurons (PMID:29934347); no targeting sequence
SYNE2	root	immunogold EM paper is abstract-only (PMID:15671068); KASH anchor is in the nuclear envelope and SR
TARDBP	root	protease-protected pool (PMID:28794432, PMID:33031745); matrix (mt-tRNA binding) and inner membrane (TIM22 import) lines disagree
TAX1BP1	root	virus-induced recruitment to MAVS (PMID:27736772)
TICAM1	root	poly I:C-induced recruitment of tagged TRIF (PMID:21703541)

What batch 4 showed:

Signalling genes are the bulk of the tail. Ten of the 24 (the four
MAPK-pathway kinases, NLRP3, PYCARD, SQSTM1, STUB1, TAX1BP1 and TICAM1)
reach mitochondria on a stimulus. All stay at the
root as non-core; none earns an outer-membrane term (TRAF3/BRAF precedent).
Location is not activity. Rule 2 cuts both ways: MOTS-c, humanin and
NENF are shown in mitochondria by direct assays, but their activities are
elsewhere, so the organelle is non-core rather than ACCEPT.
Conflicting sub-compartment lines are common. TARDBP, MAPK3, PNPO and
MBTPS1 each have a sub-mitochondrial assay, but the lines point to
different compartments, so rule 4 keeps them at the root.
Two more source defects. SCP2's EXP row cites a peroxisome paper, and
SHMT1 inherits an SHMT2-only PAINT node with no NOT on the SHMT1 clade.
is_active_in: four GO:0005739 rows carry it. MYL10, PNPO and
SHMT1 (all IBA) have no activity shown in the organelle; TAX1BP1's IDA is
defensible for its virus-induced pool. The first three join the
open-question worklist.
To follow up: the URI paper (PMID:17936702) behind PFDN2 and PFDN4 is
abstract-only; its full text would settle both. SYNE2's immunogold paper
likewise.

2026-10-25 (batch 3: 20 human genes)

The next 20 alphabetically (CFAP91 to LDHA). One gained a compartment term,
DGAT2 at the mitochondria-associated ER membrane; 19 stay at the root. Six
changed action (CKB, DTYMK, FKBP4, FOXO1, GLUL, KYNU). The brief now carried the batch 2 lessons (check crude
fractions for ER, note is_active_in, YAML pitfalls).

gene	outcome	evidence
CFAP91	root; rows stay MARK_AS_OVER_ANNOTATED	only short isoforms were mitochondrial; canonical AAT-1L is diffusely cytoplasmic, endogenous protein axonemal
CKB	root; IEA and 2 ISS MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE	mouse BAT: ~10% of CKB protease-protected while ER marker CANX is degraded; import needs an internal MTS (PMID:33597756)
CRYAB	root	crude stress-induced fraction without ER marker behind the projection; kept as a stress pool
CYP1B1	root	fractionation controlled with an ER-only CYP1A1 mouse; inner membrane only by CYP1A1 analogy
DGAT2	NEW GO:0044233 MAM (IC)	enriched in Percoll-purified MAM with calnexin; DGAT activity in mito+MAM fraction (PMID:19049983); multi-pass ER membrane
DTYMK	root; IBA and HTP MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE	MitoCoP and MitoCarta3.0 list it; 1979 rat activity assay cannot tell it from CMPK2
ERN1	root; IEA stays MARK_AS_OVER_ANNOTATED	type I ER membrane protein; mouse source paper does not mention IRE1 in its abstract
FANCD2	root	mouse co-staining and fractionation without ER marker; nucleoid by co-IP alone
FBXO7	root	CCCP-induced recruitment; surface (PINK1/Parkin) and "into mitochondria" readings disagree
FKBP4	root; EXP and IEA KEEP_AS_NON_CORE to MARK_AS_OVER_ANNOTATED; source defect	PMID:21730050 is about FKBP51; FKBP52 appears only in its introduction and methods (full text checked)
FOXO1	root; ISS UNDECIDED to KEEP_AS_NON_CORE; source defect	ISS WITH/FROM is FCoR, a Foxo1-binding corepressor, not a homologue; claim stands on PMID:30974111
GLUL	root; IEA REMOVE to MARK_AS_OVER_ANNOTATED	rests on a curator inference for rat GS from a 1978 purification; consistent with the ISS row
GLYCTK	root	GFP study shows only the truncated isoform 2 mitochondrial; MitoCoP row
HSP90AA1	root	recruited via Tom70 with IRF3/Bax on infection
IFIH1	root	crude 10,000 g pellet with MAVS only; kept for the established MAVS platform
IFIT3	root	curated UniProt mitochondrion (PMID:21813773); soluble TPR adaptor bridging TBK1 to MAVS
IRF3	root	Tom70-dependent recruitment; fractions without ER marker
KIF5B	root	cargo association via TRAK/Miro; projection from one mouse IDA
KYNU	root; IBA and IDA to UNDECIDED	PAINT node seeded only by KYNU's own 1984 fractionation; UniProt and HPA say cytosol (PTS precedent)
LDHA	root; IBA stays MARK_AS_OVER_ANNOTATED	seed paper: gold-labelled LDHA not seen in mitochondria, unlike LDHB

What batch 3 showed:

Source defects are a distinct class. FKBP4's experimental row comes
from a paper about its paralog, and FOXO1's ISS row takes its WITH/FROM
from a binding partner. Both are recorded as source_defects for UniProt
rather than argued only in the reason.
Proteome datasets rescue minor pools. DTYMK and CKB went back to
non-core on MitoCoP/MitoCarta or a controlled protease-protection assay,
as ADO did in batch 2.
A contact site is a compartment too. DGAT2's evidence names the MAM
directly (Percoll MAM fraction with calnexin), so the bold call is
GO:0044233 rather than a mitochondrial sub-compartment. STOM did not
qualify because it had no MAM marker.
is_active_in: five GO:0005739 rows in batch 3 carry it (CKB ISS;
CYP1B1, DTYMK, KYNU and LDHA IBA). CKB's activity is shown in mouse
mitochondria, and KYNU's seed measured activity in a mitochondrial
fraction whose reality is unresolved. CYP1B1, DTYMK and LDHA have no
activity shown in the organelle and join the open-question worklist.
To follow up: GLYCTK's full-length N-terminus looks like a
presequence (arginine-rich, acid-free) that UniProt does not annotate; a
targeting predictor would say whether it is a matrix candidate. (FKBP4's
nucleus rows, which quoted the same FKBP51 paper, were re-supported in
review of PR #5539.)

2026-10-24 (batch 2: 20 human genes)

The next 20 human genes, alphabetically, with root-only mitochondrion rows
from synthesis evidence (ABCC9 to CFAP410). Only one gained a
sub-compartment. This slice is dominated by signalling and cytosolic
proteins that reach mitochondria transiently, so the main work was checking
whether the organelle claim holds at all.

gene	outcome	evidence
ABCC9	root	Ensembl projection rests on mouse IDA/IMP; mitochondrial forms are short splice products; inner-membrane KATP location only assumed
ABL1	root; NAS row ACCEPT to KEEP_AS_NON_CORE	kinase recruited under ROS or ER stress; no presequence or TM
ACACA	root; IBA stays MARK_AS_OVER_ANNOTATED	PAINT node seeded only by ACC2-lineage genes; ACC1 is cytosolic
ACBD3	root	By similarity, from rodent PAP7 binding the mitochondrial PBR; one indirect line
ADO	root; IBA, IEA and HTP rows to KEEP_AS_NON_CORE (were MARK_AS_OVER_ANNOTATED / REMOVE)	mouse fractionation and MitoCarta HDA behind the projection; the earlier "UniProt says cytosol" was wrong
AGTPBP1	NEW outer membrane (IC)	mouse sub-mitochondrial fractionation (PMID:20620870); soluble, no presequence or TM
ANK2	root	contact points on the mitochondrial surface by imaging; no anchor
ARAF	root	sub-compartment lines disagree (rat IMS/matrix fractionation vs cytosolic-face MST2/BAD roles)
ARL16	root	organelle-level co-staining only, of an N-extended variant
ASAH2	root; rows to MARK_AS_OVER_ANNOTATED	the original GFP construct was N-terminally truncated; full-length ASAH2 goes to the plasma membrane (PMID:15845354)
ATP7B	root; HTP row UNDECIDED to KEEP_AS_NON_CORE	only the processed 140 kDa form is mitochondrial; MitoCoP call deferred to the curator, as for AMACR
AZIN2	root; row to MARK_AS_OVER_ANNOTATED	projection from a crude mouse fraction; the mouse protein also has a NOT row that was not projected
BRAF	root	Bag1-dependent pool recruited to the mitochondrial surface; no anchor
C0HM83 (SHMOOSE)	root	no presequence, no TM, no sub-mitochondrial assay
CALR	root; row stays MARK_AS_OVER_ANNOTATED	one rat study; signal peptide, KDEL retention, curated ER lumen
CAMK2A	root; ISS row UNDECIDED to MARK_AS_OVER_ANNOTATED	the mouse donor's antibody is pan-"CaMKII"; alpha is never named
CASP4	root	immuno-EM on ER and mitochondria; fractions ER-contaminated
CAVIN1	root; rows stay UNDECIDED	combined mitochondrial/nuclear pellet; full text not retrievable
CCN6	root	two direct sources; ER signal peptide; sub-mitochondrial data not citable
CFAP410	root	one hedged 1997 antisera study; MTS only predicted

What batch 2 showed:

Transient recruitment is the common case. ABL1, BRAF, ARAF, ANK2
and CASP4 reach the mitochondrial surface under signalling or stress. The
outer membrane is tempting, but none has an anchor or a fractionation
that separates surface from contact sites. As with TRAF3 they stay at the
root.
Projections carry their source's weakness. AZIN2's projection dropped
a NOT row on the mouse source; CAMK2A's donor IDA used a pan-CaMKII
antibody. Checking what the source annotation rests on decided both.
The MAM lesson from STOM recurred. CASP4's fractions were
ER-contaminated by the authors' own account, and CALR is an ER-lumen
protein supported by one rat study. Neither paper separates mitochondria
from ER contact membranes.
Two earlier calls were reversed. ADO, where an earlier review
misread UniProt, and ASAH2, where a 2005 follow-up showed the 2000
construct was truncated.
Tooling: just backfill-record-sections writes unquoted flow lists
([CC:SUBCELLULAR LOCATION]) that the reference validator cannot parse.
The agents quoted them by hand in the files they touched.

2026-10-24 (STOM and PTS re-checked)

Batch 1 left both genes at the root with every row KEEP_AS_NON_CORE, flagged as
possible organelle-level over-annotations. Two new sources settled what could be
settled.

STOM: full text of PMID:25262680 does not support mitochondrion. The
cached copy is abstract-only, so the curator's AgBase IDA was not second-guessed
in batch 1. The user supplied the full text (Kim et al. 2014, J Microbiol
Biotechnol 24:1744-1754, doi:10.4014/jmb.1409.09063). The "mitochondrial"
signal is co-fractionation in a crude heavy-membrane pellet that also carries ER:

"This fraction was enriched with mitochondria, as evidenced by the presence of
Hsp60, a mitochondrial matrix protein [28]. In this fraction, we also detected
the ER marker Grp78, probably because of the physical linkages between some
fractions of the ER and mitochondria."

Sub-fractionation keeps the ER marker with stomatin:

"Analysis of the resulting four fractions by immunoblotting revealed that
HCV NS proteins had co-fractionated with stomatin, Hsp60, and calreticulin, an
ER lumen marker"

and detergent pulls the mitochondrial marker away from stomatin:

"When the P15 fraction was treated with 1% NP40, NS5B and stomatin remained in
the same fraction, while Hsp60, which was released from mitochondria-derived
compartments, was detected in the bottom fractions"

The authors themselves read the compartment as MAM, an ER region:

"our results suggest that the DRM in the P15 fraction might represent the
mitochondria-associated membranes, which are known to be reversibly tethered
to mitochondria and consist of a region of ER [44]."

and imaging shows no mitochondrial pattern:

"Stomatin exhibited a distinct cytoplasmic and perinuclear localization (Fig. 2C)."

Both GO:0005739 rows (the IDA and the circular Ensembl-projection IEA) are now
MARK_AS_OVER_ANNOTATED, not REMOVE: the co-fractionation is real, it is the
curator's source, and the paper does place stomatin on membranes tethered to
mitochondria. GO:0044233 (mitochondria-associated endoplasmic reticulum membrane
contact site) fits the authors' reading but rests on one detergent experiment
with no MAM marker, so no NEW row. The reference now carries a
reference_review saying the paper does not support mitochondrion.

PTS: conflicting weak lines, so UNDECIDED. The only mitochondrial line is the
LIFEdb GFP-fusion screen, which also alone seeds the PAINT node behind the IBA.
The Human Protein Atlas detail (ENSG00000150787, downloaded 2026-10-24) is one
antibody, HPA001481, passing protein-array validation, with immunofluorescence
in A-431, U-251MG and U2OS showing cytosol only, at reliability "Approved". Its
counterstains are nucleus, microtubules and ER, with no mitochondrial marker.
One low-resolution direct row for and one approved single-antibody row against
cannot be resolved without new data, so both rows are UNDECIDED (the enum's use
for unresolvable evidence) rather than kept as if the claim stood.

What this adds to the policy: a crude-fraction IDA is a red flag of its own when
the fraction is not checked for ER. MAM co-purifies with mitochondria in every
P15/heavy-membrane prep, so "found in the mitochondrial fraction" without an ER
marker cannot distinguish mitochondrion from MAM.

2026-10-20 (batch 1: 20 human genes)

Applied the policy to 20 human genes with root-only mitochondrion rows from
synthesis evidence. Nine gained a sub-compartment, eleven stayed at the root.
No synthesis row was MODIFYed; all kept their action, with reasons recording
the PAINT-seed or mapping check.

gene	outcome	evidence
NSUN2	NEW matrix (IDA)	matrix-only APEX2 labelling and protease protection (PMID:31276587)
RAD51C	NEW nucleoid (IDA)	salt-resistant nucleoid fractions with TFAM; D-loop mtDNA IP (PMID:29158291)
AMACR	NEW matrix (IC)	soluble, no TM, import signal, mitochondrial beta-oxidation; signal not cleaved
DUT	NEW matrix (IC, isoform 3)	processed presequence, no TM, mtDNA replication context
FEN1	NEW matrix (IC)	protease protection, PMF-dependent import, soluble; internal pool is mostly FENMIT
APEX1	NEW IMS and matrix (IC)	Mia40-dependent import and PLA with Mia40; matrix pool repairs mtDNA (one group)
ADCK2	NEW inner membrane (IC); GO:0016020 row MODIFY to it (in review)	predicted TM at the COQ8A anchor position, protease protection, yeast Cqd1
GPX4	NEW inner membrane (IC, mitochondrial isoform)	rodent sub-mitochondrial fractionation, import, cardiolipin
ABHD18	NEW inner membrane (IC)	predicted presequence, cardiolipin substrate, yeast Cld1 analogy; least certain call
AKAP10	root	predicted MTS only; organelle itself contested
DHFR	root	sub-mitochondrial data were for DHFR2 and later reattributed
FADS1	root; rows to KEEP_AS_NON_CORE	ER enzyme; one imaging study; isoform error in old review fixed
STOM	root; rows to MARK_AS_OVER_ANNOTATED (2026-10-24)	Ensembl projection is circular; full text shows ER/MAM co-fractionation
ASB9	root	overexpression and crude pellet only
SECISBP2	root	MTS vs negative fractionation disagree
RFK	root	yeast FMN1 is inner membrane, human RFK has no TM
PTS	root; rows to UNDECIDED (2026-10-24)	PAINT node seeded only by PTS's own GFP-screen row; HPA says cytosol
TRAF3	root	recruited via MAVS; no outer-membrane evidence
HLCS	root	single functional line (matrix carboxylases), like PYCR2
GLOD4	root	no presequence, no TM, crude fractionation only

What batch 1 showed:

IBA seeds are often thin. Several PAINT nodes are seeded by one gene,
sometimes the target itself (PTS) or a distant organism (SECISBP2, from a
trypanosome screen). DHFR's node includes the paralog DHFR2 but is carried
by rodent Dhfr. None of these was wrong about the organelle, but two (PTS,
STOM) may be over-annotated at the organelle level; they were kept
consistent with their direct rows pending a decision.
Sub-compartment calls cluster by evidence type. Two came from a direct
sub-compartment assay (IDA); the rest are IC from at least two agreeing
lines. Every root-staying gene failed rule 4 for a stated reason.
Core-function locations were left alone in all 20 except NSUN2, where
the mt-tRNA methylation activity is shown in the matrix.

2026-10-18 (MMADHC core-function location)

Asked OpenScientist whether MMADHC's mitochondrial pool does part of the core
AdoCbl-supply job in the matrix (report in
genes/human/MMADHC/MMADHC-hypotheses/mmadhc-mitochondrial-pool-adocbl/). The
verdict was "partially supported": the matrix location is well supported, but
no matrix-side partner or step has been shown, and a cytosolic branch-point
model fits the data equally well. A cached mouse paper (PMID:41411131), which
the report did not use, ties mitochondrial Mmadhc accumulation to more AdoCbl.
So the pool matters, but where its step happens is unknown. Outcome: the matrix
stays as the NEW location row but is removed from the core function's
locations; the mitochondrion rows stay non-core; the open question is a
knowledge gap with a compartment-restricted rescue experiment. This is a useful
general point for the worklist: a sub-compartment can be a well-supported
location without being where the gene's core activity is shown to happen.

2026-10-18 (revision: add, don't substitute)

Agreed with cjm after the PR #5208 review. The pilot's IBA/IEA mitochondrion
rows on AARS2, CLYBL and MMADHC no longer MODIFY to the matrix: their sources
correctly assert the organelle and could not see further, so the row's claim
is true and they are kept (ACCEPT for AARS2 and CLYBL; KEEP_AS_NON_CORE for
MMADHC, matching its other mitochondrion rows). The matrix term stays as the
NEW IC row on each gene. The AARS2 and MMADHC IBA propagation reviews go back
to NO_FAILURE (the GRANULARITY_MISMATCH label meant "true but uninformative",
which contradicts rule 1). With every root row on one action, the validator
exemption for the MODIFY split is no longer needed and was removed. The
comparator that decided it: DARS2 and MMUT carry the matrix term alongside
their mitochondrion rows in GOA. The pilot table below records the first
version.

2026-10-18 (pilot: four human matrix candidates)

Applied the policy to AARS2, CLYBL, PYCR2 and MMADHC. Three were moved to the
matrix and one was held at the root.

gene	outcome	synthesis rows	direct rows	lines of evidence
AARS2	NEW matrix (IC)	IBA, IEA: ACCEPT -> MODIFY	3 kept ACCEPT	predicted MTS, no TM; charges mt-tRNA(Ala); matrix amino-acid pool for editing; matrix lactylation substrate PDHA1
CLYBL	NEW matrix (IC)	IEA: ACCEPT -> MODIFY	2 kept ACCEPT	predicted MTS, no TM; soluble trimer (structure); matrix CoA-ester substrates; protects matrix MMUT
MMADHC	NEW matrix (IC)	IBA, IEA: KEEP_AS_NON_CORE -> MODIFY	2 kept KEEP_AS_NON_CORE	MTS (deletion relocates it to cytosol); import needs membrane potential; degraded by matrix LONP1; feeds matrix MMAB/MMUT
PYCR2	stays at root	IEA kept ACCEPT, with reason	3 kept ACCEPT, reasons added	no presequence (N-terminus sequenced), cytoplasmic pool, only two-way fractionation: rule 4 not met

What the pilot showed:

None of the matrix calls rests on a direct sub-mitochondrial assay.
All three NEW rows are IC built from several agreeing lines. That is what
rule 4 asks for, and it is why PYCR2, with only a functional hint, stayed at
the root.
The IBA red flag is often structural. For AARS2 and MMADHC, the PAINT
node that gives the mitochondrion IBA is seeded only by the gene's own
root-level annotation. The IBA cannot be finer than its seed. Fixing the
seed (a matrix annotation on the human gene) is the way to improve the
propagation.
(Superseded by the revision entry above: the MODIFY split was dropped, so no
exception is needed.) The action-consistency check warns on the policy's own pattern. ACCEPT
or KEEP_AS_NON_CORE on direct rows alongside MODIFY on synthesis rows for
the same root term gives "Inconsistent review actions for term
GO:0005739". The warning is expected under this policy. It needs an
exception (or a different rule) before the policy is rolled out.
PYCR1's matrix term is Reactome TAS only; there is no sub-mitochondrial
assay for either PYCR paralog.

2026-10-16 (human mitochondrion cases)

Human reviews carry 1,699 existing rows to GO:0005739 mitochondrion across
534 genes. In 244 rows (135 genes) the review gives no mitochondrial
sub-compartment. Worked examples, grouped by what the policy says to do:

Be bold (direct evidence is fine, biology says matrix): AARS2 (5 rows,
all ACCEPT; transit peptide; aminoacylates mitochondrial tRNA), CLYBL (the
reviewer's own reason says "all of CLYBL's substrates ... are matrix
species", yet there is no matrix term), PYCR2, MMADHC.
IBA red flags from paralog propagation, already caught: SHMT1 (the
mitochondrial paralog is SHMT2), ACACA, LDHA; all MARK_AS_OVER_ANNOTATED.
IEA red flags from ortholog projection (GO_REF:0000107): CRY1 (from
mouse, REMOVE), CALR (from rat), MAPK1, BRAF.
Recruited to the outer membrane or ER-mitochondria contacts (case 3):
IRF3, NLRP3; KEEP_AS_NON_CORE.
Assay-limited, ACCEPT is right: SMIM8 (MitoCoP proteomics only). It has
a predicted TM segment, so GO:0031966 mitochondrial membrane (which
membrane unknown) would be a bolder call that the evidence supports.
Dual location: AMACR (peroxisome and mitochondrion), GPX4
(isoform-specific).

2026-10-16 (policy)

Decided with cjm. TOO_GENERAL is wrong here, because the organelle is still
informative. The call is evidence-based: direct evidence that cannot resolve
the sub-compartment is ACCEPT; IBA, IC and propagation IEA are red flags to
examine; and reviews should propose a sub-compartment when the gene has none
and several lines of evidence agree. The script now classes evidence codes as
direct, synthesis or author (TAS/NAS) and writes the red-flag worklist for the
first-pass organelles.

2026-10-16

Project opened with the question about root organelle terms. The survey script
uses OAK sqlite:obo:go ancestors over is_a and part_of, and treats a
descendant as membrane if it is under GO:0016020 membrane and as lumen if it
is under GO:0031974 membrane-enclosed lumen. "Other" is mostly complexes,
chromatin and sub-organellar structures. Runtime is about 3 minutes over the
full corpus. The ten roots are hard-coded in the script; extend ROOTS to add
more.

Warnings (9)

Organelles

The question

First survey of the corpus

Policy

Survey under this policy

Open questions

STATUS

NOTES

2026-11-11 (tethering route added to rule 7)

2026-11-06 (ER batch 4: the last 19 human genes)

2026-11-05 (ER batch 3: 20 human genes)

2026-11-05 (ER batch 2: 20 human genes)

2026-11-04 (ER batch 1: 20 human genes)

2026-11-03 (batch 4: the last 24 human genes)

2026-10-25 (batch 3: 20 human genes)

2026-10-24 (batch 2: 20 human genes)

2026-10-24 (STOM and PTS re-checked)

2026-10-20 (batch 1: 20 human genes)

2026-10-18 (MMADHC core-function location)

2026-10-18 (revision: add, don't substitute)

2026-10-18 (pilot: four human matrix candidates)

2026-10-16 (human mitochondrion cases)

2026-10-16 (policy)

2026-10-16