Omics and High-Throughput Evidence in GO Annotation

Bottom line: GO annotations derived from omics data are filed under two
evidence-code families. The high-throughput (HTP) codes HDA/HTP/HMP/HGI cover 65,876
rows. RCA, the reviewed-computational-analysis code, covers another 1,557 omics rows.
The same dataset can appear under both codes, and the code says little about what the
annotation claims. What predicts whether an annotation holds up is how the data were
used:

• A location read from a fraction is usually sound. 57% of reviewed rows are
  accepted, rising to 63–89% for systematic localization screens and organelle
  proteomes. The exceptions are vesicle and body-fluid proteomes (6% accepted) and the
  generic term membrane (23%). These rates describe the reviews as inherited on
  2026-10-05; this project has since re-reviewed both clusters by hand (Recommendation 1),
  so the current vesicle and membrane actions are its own output, not independent
  evidence.
• A function read from location or co-purification is unreliable under either code.
  Examples are "cadherin binding" from an E-cadherin proximity screen (2 of 50 accepted)
  and the RCA matrisome "ECM structural constituent" rows (56%).
• A process read from differential expression has been coded HDA. These are 0 of 9
  accepted.
• NOT annotations from exclusion exist only under RCA. There are 627 such rows and
  just 4 HTP-family NOTs, and 43% of the RCA cytosol NOTs are contradicted elsewhere in
  GOA.

The same papers make the point directly. For the twelve papers cited under both codes,
our reviews accepted 74% of the HDA location rows and 56% of the RCA function rows.

This project replaces the earlier RCA_EVIDENCE project. Its RCA analysis is now the
sub-page RCA: omics-derived annotations.

Scope

The framework is the GO consortium's guidance for annotating from high-throughput studies
(PMID:30715275), which introduced the HTP codes so that users could separate
screen-derived annotations from hypothesis-driven ones.

Methods and reproducibility

python3 projects/OMICS_EVIDENCE/rca/rca_source_catalog.py     # RCA download (also used for cross-code joins)
uv run python projects/OMICS_EVIDENCE/htp/htp_inventory.py    # HTP download + reviewed corpus

Count claims as distinct gene × term pairs, not rows. A GAF row is repeated once per
annotation extension, so row counts can mislead badly. The 1,535 HMP rows for
metal-dependent deubiquitinase activity (PMID:26412298) are 5 S. pombe DUBs,
each listed once per substrate. The 4,090 HDA rows for RNA polymerase II cis-regulatory region sequence-specific DNA binding (PMID:40015273) are 67
transcription factors, each listed once per target. In GOA, 65,876 HTP-family rows
reduce to 50,795 distinct gene × term pairs.

How omics data become GO annotations

The classification below applies across evidence codes. Each row is a way of turning a
dataset into a GO claim. Reviewed-row figures are from this repository's gene reviews as
of 2026-10-05, before the batch dispositions of 2026-10-06 (Recommendation 1); the
committed htp_inventory_report.txt is that snapshot.
"Rejected" means MARK_AS_OVER_ANNOTATED, REMOVE or MODIFY.

1. Location → location: sound, unless the compartment is a vesicle or "membrane"

Nearly three-quarters of HTP-family pairs are HDA locations, and they are the use the
codes were designed for. Accepted rates are high when the experiment localizes proteins
one by one or purifies a defined organelle:

Vesicle-type terms (extracellular exosome, extracellular vesicle, blood microparticle, vesicle) account for 712 reviewed rows, more than any other HTP
cluster. Reviewers rarely say they are false. 59% are kept as non-core and only 11
are removed. They do say the location carries no information about function. In the ADIRF
review's words, one such paper "annotates 1046 distinct gene products with GO:0070062
and nothing else … a bulk import, not 1046 independent findings". The generic
membrane term from membrane fractions fails differently. The fraction includes soluble
and peripheral proteins (ACOX3 "has no transmembrane domain and is a matrix-soluble
enzyme"), so half of those rows are rejected.

2–3. Function from omics: the assay must measure the function

When the assay measures the activity itself, as with RNA crosslinking in interactome
capture, DNA binding in a TF atlas, or activity-based probes for serine hydrolases, the
molecular-function claim is direct. Errors there come from carry-over of abundant
proteins. TXN and HSPA9 appear as "RNA binders", and 40% of reviewed
interactome-capture RNA-binding rows were accepted.

When the function is inferred from where the protein was found, the claim goes beyond
the data, whatever the evidence code:

• ECM proteomics → "ECM structural constituent" (RCA). The term follows the protein's
  matrisome category, not its biology. See the RCA sub-page.
• E-cadherin proximity proteomics → "cadherin binding" (HDA, PMID:25468996, 255
  genes). 48 of 50 reviewed rows were not accepted. The GLOD4 review describes the
  artefact: "abundant soluble cytosolic proteins are labelled by a membrane-tethered
  biotin ligase and enter GOA as cadherin binders."

Both are the same inference ("found with X, therefore binds or builds X") filed under
different codes. A direct assay code makes the error harder to spot.

4. Expression change coded as a direct assay

PMID:16210410 (a differential membrane proteome during osteoblast differentiation)
supports 42 HDA osteoblast differentiation pairs. None of the 9 reviewed rows was
accepted: "differential expression during osteoblast differentiation does not imply a
functional role in the differentiation process" (CLTC review). This is an expression
pattern used as evidence of process involvement, the problem analysed in the
IEP project. It also uses the wrong code: HEP is the high-throughput
expression code, and HDA asserts a direct assay of the process.

5. Screens → process (HMP/HGI)

Genome-wide RNAi and mutant screens supply 875 HMP and 52 HGI process pairs. The
reviewed sample is small (47 rows). HGI rows are mostly accepted (5/6), while HMP rows
are mostly kept as non-core (22/41): a screen hit in "defense response to Gram-negative
bacterium" or "wound healing" is usually a pleiotropic phenotype rather than the gene's
role.

6. NOT annotations: an RCA-only practice

The HTP-family codes carry 4 NOT rows in all of GOA: three S. pombe DUBs NOT is_active_in nucleolus and one miRNA. RCA carries 627, all from two Arabidopsis
fraction proteomes. The Golgi paper PMID:22430844 shows the split cleanly. It supports
415 positive HDA Golgi apparatus rows and 191 RCA NOT Golgi apparatus rows: the same
experiment, recorded with one code for each polarity. Analysis of the cytosol NOTs
(43% contradicted, including 37 cytosolic ribosomal proteins) is on the
RCA sub-page.

Same papers, two codes

htp_inventory.py finds 12 papers cited by both HTP-family and RCA rows. Ten are
BHF-UCL's ECM proteomics papers, which yield HDA locations (extracellular matrix,
extracellular region) and RCA molecular functions. The other two are the Golgi
paper and a T. brucei mitochondrial-membrane proteome. Reviewed outcomes on those 12
papers:

The data and the curators are the same. The difference is the inference step.

Reviewer checklist

1. Classify the use (table above) before judging the code. The evidence code tells
   you the provenance, and the use tells you how much the data support.
2. Location rows: accept localization screens and confidence-scored organelle
   proteomes. Keep vesicle and body-fluid rows as non-core unless the protein is a known
   vesicle component. Prefer a specific membrane over membrane. Assess target-specific evidence for
   peripheral or transient association as well as membrane anchors; lack of an anchor
   alone does not establish that a membrane localization claim is over-annotated.
3. Function rows: ask whether the assay measured the activity. If it did not, and the
   function is inferred from co-location or co-purification, treat the row as
   over-annotated unless independent evidence exists. If it did (crosslinking, binding
   assay, activity probe), check for abundant-protein carry-over.
4. Process rows from HDA: check whether the data are really differential expression.
   If they are, the row is an IEP-type inference with the wrong code.
5. NOT rows from omics: look for positive annotations to the same compartment. A
   positive observation outranks an exclusion.
6. Judge by distinct gene × term pairs. A gene can carry hundreds of rows for one
   claim through annotation extensions.

Recommendations

1. No rule decides HTP-family rows; both clusters are reviewed by hand (2026-10-10).
   Two default dispositions were agreed and applied as rules on 2026-10-06: vesicle-type
   locations → KEEP_AS_NON_CORE, and generic membrane for proteins without a membrane
   anchor → MARK_AS_OVER_ANNOTATED. Both were withdrawn and every edit undone. A rule
   lets existing annotations (or a category) make the curation call, while the goal is
   manual review with existing annotations used as leads. The audit files
   (htp/disposition-2026-10-06*.yaml) remain as a record of what the rules did.
   htp/apply_dispositions.py is retired; it
   now only supplies helpers to the review scripts.

Vesicle-row manual review. Candidates are the 283 HDA/HTP vesicle-type rows (164
genes) whose action was not already a settled non-core or removal call. That is the
241 rows the withdrawn rule had changed plus the 42 ACCEPTs it flagged.
vesicle_review/build_dossiers.py
gathers material for each row and decides nothing:
- UniProt location, function and signal-peptide/TM features;
- the source sample (urine, prostatic secretion, B cells, saliva, plasma, CSF...);
- how many independent EV studies list the protein;
- the gene's other location annotations, offered as leads to check.

Each protein was judged with MISEV2018 (PMID:30637094, cached) as the guide:
- Expected EV content, true but not core → KEEP_AS_NON_CORE. This covers
transmembrane proteins of the plasma membrane or endosomes (category 1) and
cytosolic proteins, including promiscuous enzyme and cytoskeleton cargo (category 2).
- Over-annotated → MARK_AS_OVER_ANNOTATED. This covers co-isolated non-EV
constituents (category 3). MISEV names ALB, APOA1/2 and APOB as negative markers;
APOE and APOL1 ride on the same plasma lipoproteins, so their plasma-microvesicle
rows were treated the same way; residents of the ER, Golgi, mitochondria or nucleus (category 4); and
secreted or lumenal proteins with no EV receptor shown (category 5).
- ACCEPT only for ALIX/PDCD6IP, on protein-specific grounds: UniProt records that
ALIX sorts multivesicular-body cargo into intralumenal vesicles and regulates exosome
biogenesis (PubMed:22660413). MISEV itself puts ALIX in category 2a with TSG101 and the
CHMPs, so this is not a MISEV category distinction.
- Category 2 is split as MISEV splits it: 2a, cytosolic proteins with lipid or
membrane-protein binding ability (ESCRT, annexins, HSPA8), and 2b, promiscuous cargo
(actins, abundant enzymes). Both are KEEP_AS_NON_CORE. Ribosomal proteins are
category 3b (co-isolated aggregates), so they are MARK_AS_OVER_ANNOTATED.

Where an earlier reason made a factual claim, it was checked. The SCAMP3 ACCEPT said
SCAMP3 is "required for extracellular vesicle biogenesis", but the cited papers
(PMID:19158374, PMID:23418353) show ESCRT-dependent MVB sorting of EGFR, so the row is
KEEP_AS_NON_CORE. GRID1, a CNS-restricted receptor detected in urine, is
over-annotated as a likely peptide misassignment.

decisions_draft.py holds one
decision and basis per protein, and
decisions.yaml lists every row.
249 rows changed (excluding the two CD2AP rows; see the session notes):

Final vesicle-row actions: 233 KEEP_AS_NON_CORE, 47 MARK_AS_OVER_ANNOTATED, 3 ACCEPT
(PDCD6IP). For most cytosolic cargo the manual outcome matches the withdrawn rule,
because under MISEV such proteins really are in EVs. As first applied, it differed from
the rule on 91 rows: 52 that the rule would have made non-core are over-annotated, and 39 ACCEPTs
that the rule left in place are changed. The edited reviews validate. The quote errors
in ATP6V1C1 and NRAS, which predated this work, are fixed (ATP6V1C1 in this PR, NRAS
on main).

Second pass: rows already set to non-core. The 418 HDA/HTP vesicle-type rows (220
genes) that reviewers had already set to KEEP_AS_NON_CORE were not candidates above.
build_dossiers.py --settled built their dossiers
(dossiers_settled.yaml),
and each protein was judged the same way, with one decision per protein in
decisions_settled.py.
Proteins reviewed in both passes (PDCD6IP, ACTB, AHCY, ALPL, ASS1, ATP6AP2) got the
same call in both. APOE did not: its pass-1 row is a plasma-microvesicle sample, but its
rows here come from urine, prostatic secretion, B-cell culture and CSF. APOE is not a
MISEV marker, and van Niel et al. (PMID:26387950) show it on intraluminal vesicles
secreted as exosomes, so those four rows stay KEEP_AS_NON_CORE. GAS6 and APOH also stay
KEEP_AS_NON_CORE: both bind phosphatidylserine or anionic phospholipids, the lipid-receptor
route by which MISEV category 5 allows secreted proteins onto EVs. The same applies to
every category-5 protein with experimental proteoglycan, heparin or lipid binding: FN1,
NID1, COMP, PSAP, GM2A, ITIH1 and ITIH2 stay KEEP_AS_NON_CORE here, and THBS1, THBS4 and C1QB become
KEEP_AS_NON_CORE in pass 1. Mitochondrial and ER proteins whose own reviews keep an
experimental cell-surface or plasma-membrane pool (HSPD1, ATP5F1A/B, CALR, GOT2, P4HB; LMAN2 and
SYNE2 in pass 1) also stay KEEP_AS_NON_CORE, since small EVs budding from the plasma membrane can
carry that pool, and the same holds for secreted proteins with such a pool (SMPD1, LPO here;
C3, CLU and NEU1 in pass 1). The surface-pool rule does not override category 3b:
RACK1 and RPS3 have experimental plasma-membrane rows, but MISEV2018 lists ribosomal
proteins as co-isolated aggregates, and a transient membrane-recruited pool does not make
the aggregate an EV component, so their rows stay MARK_AS_OVER_ANNOTATED. 71 rows in 43 genes
changed
(decisions_settled.yaml):

The other 347 rows stay KEEP_AS_NON_CORE. They cover cytosolic cargo and PM/endosomal
membrane proteins, including the ESCRT subunits (CHMP*, TSG101, VPS28, VPS37B–D,
MVB12A/B). MISEV2018 lists TSG101 and the CHMPs in the same category 2a as ALIX, so
accepting ALIX alone rests on its protein-specific role in exosome biogenesis, not on
MISEV. The edited reviews validate; A2M's quote error, which predated this work, is
fixed.

Membrane-row manual review. Candidates are all 163 HDA/HTP membrane rows (159
genes) for proteins without a structured UniProt anchor feature. A protein with a
TM segment, intramembrane region or lipid anchor is in a membrane by construction.
membrane_review/build_dossiers.py
gathers the material for each row and decides nothing. A dossier holds:
- UniProt location and function text, verbatim;
- the gene's other location annotations with their codes and references, offered as
leads to check;
- the source paper's title;
- the current action and reason.

151 of the 163 rows come from one NK-cell membrane proteome (PMID:19946888). Its
authors stripped peripheral proteins before analysis and report that about 60% of
their identifications are not predicted membrane proteins. So the question for each
row was whether there is protein-specific evidence of stable membrane association:
- Yes (an experimental UniProt membrane location, or a membrane annotation whose
source supports it): KEEP_AS_NON_CORE. The term is true but generic; the specific
membrane term carries the location.
- No: MARK_AS_OVER_ANNOTATED. Earlier REMOVE calls on such proteins were confirmed
rather than churned.
- ACCEPT was kept only where membrane association is the core of the activity and
no specific term carries it (ANXA11, PIK3C3).

Leads were followed where they mattered. Examples:
- DNAJB11: UniProt's ER-membrane note applies only to a C-terminally tagged construct,
so the row is over-annotated.
- EIF5A: its ER-membrane EXP lead (PMID:8660923) could not be checked, because no
abstract is cached. The cached localization study shows nuclear and cytoplasmic
distribution only, so the row stays over-annotated and the reason says so.
- APOE: its plasma-membrane leads are Reactome rows for APOE bound to receptors, so it
is over-annotated.

decisions_draft.py records
a basis for every row, including the 97 confirmed unchanged.
decisions.yaml is the
reviewable output. 66 rows changed:

Final membrane-row actions: 93 MARK_AS_OVER_ANNOTATED, 56 KEEP_AS_NON_CORE, 11
REMOVE, 2 ACCEPT, 1 MODIFY. All 64 edited reviews validate.
2. Report the mis-coded clusters to their groups:
- HDA→HEP for expression-derived process rows (PMID:16210410).
- HDA for RCA Use C rows that are really direct localizations (see the RCA sub-page).
3. Report the location-to-function inferences: BHF-UCL for the matrisome mapping
(RCA sub-page), and the E-cadherin proximity screen (PMID:25468996, 255 genes,
cadherin binding).

Action items

• [x] Decide and document the vesicle-type and generic-membrane dispositions
  (Recommendation 1), then apply them as a batch to existing reviews (2026-10-06).
• [x] Replace the vesicle rule with row-by-row review, including the 42 flagged ACCEPTs
  (2026-10-10).
• [x] Extend the vesicle review to the 418 vesicle-type rows that reviewers had already
  set to KEEP_AS_NON_CORE (2026-10-10; 104 changed).
• [ ] Decide whether ESCRT subunits (TSG101, CHMP*) should be treated like ALIX, which
  MISEV2018 places in the same category 2a.
• [ ] Review a sample of the unreviewed MF clusters with the use-2/use-3 test: serine
  hydrolase activity-based probes (PMID:33827210), copper/cobalt/zinc ion binding from
  plant mitochondrial metal-affinity proteomics (PMID:20018591), kinase
  autophosphorylation (PMID:21477822).
• [ ] Review the 33 unreviewed osteoblast differentiation pairs from PMID:16210410, and
  look for other differential-expression papers coded HDA for process terms.
• [ ] Sample the largest HTP-family location sources that have no reviews yet: the
  T. brucei protein map (PMID:36804636, 5,439 genes) and the Arabidopsis
  membrane-oligomerization profiling (PMID:28887381, 3,263 rows), which also
  supplies most of the HDA rows that contradict the RCA cytosol NOTs.
• [ ] Extend the HTP-family inventory with a per-paper confidence signal (screen vs
  fraction vs vesicle) so that reviewers can triage rows automatically.

Sub-pages

• RCA: omics-derived annotations: source catalogue of all RCA,
  the matrisome category mapping, NOT contradictions, and out-of-scope RCA notes
  (YeastPathways, proteome-scale predictions, NEW rows).

Session notes

2026-10-26: FKBP5 membrane reassessment in the androgen audit

The FKBP5 GO:0016020 membrane review in PR #5554 now explicitly supersedes the
2026-10-10 manual KEEP_AS_NON_CORE → MARK_AS_OVER_ANNOTATED decision with
UNDECIDED. The original provenance is retained in the gene's rationale.
PMID:19946888 reports an NK-cell membrane fraction that includes proteins predicted
to associate transiently with membranes. Its accessible abstract does not establish
FKBP5's target-specific identification, enrichment or contamination controls, and
the publisher full text was unavailable. The unresolved question is biological
membrane association, which does not require an integral membrane anchor. This is
a source-specific reassessment, not a rule applied to other soluble proteins.
The location checklist is clarified accordingly; historical audit counts above
remain snapshots of their original sessions.

2026-10-17: optional follow-ups from the final automated review

Three suggestions from the review that approved the merged PR, none of which changes a
decision:

• The page now says that the surface-pool rule does not override category 3b (RACK1,
  RPS3).
• The positive CAT5_BOUND check now requires a directly assayed GOA row (EXP, IDA, IMP
  or IPI), matching the "(IDA)" its bases cite; the negative CAT5 check keeps the wider
  evidence set, which is the conservative direction there. All ten genes still pass.
• CAT3B's "ribosomal protein" is now checked against GO (ribosome or structural
  constituent of ribosome). RPS3 and RACK1 pass. Category 2a's "lipid or membrane-protein
  binding" half is the one template clause still taken on trust.

Re-running the first pass on current main reports three files (AHCY, CTNNB1, CUL3) whose
vesicle-row text has been rewritten by later reviews, CUL3's with row-level evidence from
the urinary-exosome database. Their actions all still match this project's decisions, so
they are left as the later curators wrote them, not reverted to the template text.

2026-10-12: fourth round of automated-review fixes

The reviewer found one category-5 protein the GOA check had missed: GM2A, whose lipid
chaperone activity (IDA) extracts GM2 from intralysosomal luminal vesicle membranes. Its
two exosome rows return to KEEP_AS_NON_CORE (CAT5_BOUND) and the file is back to its
original text. The search had looked for named binding terms, not for lipid chaperone or
carrier activity. apply_decisions.py now checks the template's last clause itself: a row
whose reason says the protein records no experimental proteoglycan, heparin or lipid binding
fails if that gene's GOA has an experimental binding, chaperone, carrier or transfer function
for a lipid, heparin, heparan, glycosaminoglycan or proteoglycan. No other category-5
protein trips it.

The re-review then found SCGB1A1, whose evidence is in UniProt rather than GOA: it binds
phosphatidylcholine and phosphatidylinositol. Those are ligands held in the central
hydrophobic cavity of its homodimer (PubMed:7664082), not an EV-surface association, so
its exosome row stays over-annotated, but its reason now says this instead of the
template clause. The check now also reads each gene's UniProt FUNCTION text for an
unhedged "binds" sentence naming a lipid or glycosaminoglycan, and matches a list of GO
ids (lipoprotein particle, phosphatidic acid and retinol binding among them) as well as
term names. Across both passes SCGB1A1 is the only hit, and its reason no longer
carries the clause.

The next re-review found ITIH1 and ITIH2, with hyaluronan binding by IDA, which the
name and id lists missed: GO files hyaluronic acid binding under carboxylic acid binding,
not glycosaminoglycan binding. Both return to KEEP_AS_NON_CORE (CAT5_BOUND) and their
files are back to their original text; ITIH3, which UniProt says only may carry
hyaluronan and which has no hyaluronan-binding row, stays over-annotated with that stated.
The check now takes its GO terms from the ontology itself (every is_a descendant of lipid,
glycosaminoglycan, proteoglycan and lipoprotein particle binding, lipid transfer and lipid
chaperone activity, plus hyaluronic acid binding), and its UniProt scan reads "binding"
as well as "binds" but skips hedged sentences. No row carrying the clause trips it.

The same re-review found the category-4 template had the same flaw: it called HSPD1,
ATP5F1A, ATP5F1B and CALR intracellular residents, while their own reviews keep
experimental cell-surface rows, and GOT2, with weaker evidence of the same kind, was kept
non-core. All five, and LMAN2 in pass 1, now share one basis (CAT4_SURFACE,
KEEP_AS_NON_CORE): small EVs budding from the plasma membrane can carry a documented
surface pool. HSPD1, ATP5F1A, ATP5F1B and CALR return to their pre-review text
(ATP5F1A/B keep only their membrane-pass edits); LMAN2's two rows move from
MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE. The category-4 template now says "mainly
resident" and that the protein's own review keeps no experimental surface pool, and
apply_decisions.py checks that claim against each review. The GO closure check now also
fails loudly if a root term is missing from the loaded GO release.

Checking the category-1 template the same way found six proteins that UniProt gives no
membrane span or lipid anchor: ANKFY1, DNAJC13, ATP6V0D1, ATP6V0D2 and ATP6V1G1 are
peripheral proteins on the cytosolic face (now category 2a, like the other V-ATPase V1
subunits), and CUBN is held at the cell surface by its transmembrane partner AMN (now its
own basis). All six rows were already KEEP_AS_NON_CORE and stay so; only the recorded basis
changed, so no review file is edited. apply_decisions.py now requires a UniProt
transmembrane, intramembrane, lipidation or GPI-anchor feature for any category-1 basis.

The category-2 template said "Cytosolic protein (UniProt)", but for 39 proteins UniProt
records no cytoplasmic location and the cytosol or cytoplasm annotation is in GO; those
now cite GO (CAT2_GO), and the script checks each attribution against its source. Three
bases were rewritten because the template did not fit at all: APOE, which is secreted,
and ARL15 and ARMC3, which have no recorded location other than EV proteomes. DNAJA2, a
farnesylated co-chaperone, moves to category 2a. No action changed.

The category-4 surface check had matched only the cell surface and plasma membrane terms
themselves. It now uses their GO descendants (is_a and part_of), which found P4HB
(external side of plasma membrane, IDA) and SYNE2 (lamellipodium and filopodium membrane,
IDA). Both now use CAT4_SURFACE: P4HB's row returns to its original text, and SYNE2's
row moves from MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE.

The last template claims to be checked were the compartments. UniProt gives CR1, SCAMP3
and EGF only a generic "Membrane" location, so their category-1 basis now says the
compartment comes from GO (CAT1_GO). The script checks every category-1, -2 and -5 location
claim against the source it cites, UniProt or GO.

The re-review found that the category-1 location check accepted bare words such as
"vesicle", so SYPL1, whose UniProt locations are cytoplasmic vesicle and melanosome
membranes, passed it. The pattern now names only plasma-membrane, endosomal and lysosomal
locations, and SYPL1, ATP6V0A1, ATP6V0C and NCSTN now cite GO for the compartment. Two more
claims are now checked: the category-4 residency clause (a nuclear, mitochondrial, ER,
Golgi, peroxisomal or lipid-droplet location in UniProt or GO) and the category-2a
cytosolic clause. SYNE2's
reason now answers the earlier reviewer's point about peptide-level detection of a giant
protein.

The next re-review applied the surface-pool test to category 5, which its template's own
wording invites ("specific or promiscuous receptors"): five secreted proteins have an
experimental cell-surface or plasma-membrane row their reviews keep. C3 (covalent C3b
deposition, EXP ×5), CLU (receptor-bound, IDA), SMPD1, LPO and NEU1 now share a
CAT5_SURFACE basis (KEEP_AS_NON_CORE). SMPD1 and LPO return to their original text; C3's
three rows move from UNDECIDED, CLU's four from ACCEPT and NEU1's two from
MARK_AS_OVER_ANNOTATED to KEEP_AS_NON_CORE. The category-5 template now also states that the
protein's review keeps no experimental surface pool, which the script checks. Two template
clauses were then still taken on trust: the "lipid or membrane-protein binding" half of
category 2a, and category 3b's "ribosomal protein" (RPS3, RACK1). The second is now
checked (see the 2026-10-17 note); the first remains on trust.

2026-10-10: third round of automated-review fixes

The reviewer found that the category-5 template still claimed "none is shown" for an
EV-surface receptor on proteins whose own annotations record exactly the routes MISEV
names. The GAS6/APOH argument is now applied to the whole class rather than only to the
proteins named in the review. Every category-5 protein's GOA was checked for experimental
heparin, proteoglycan or lipid binding:

• Back to non-core (CAT5_BOUND): FN1, NID1, COMP and PSAP, restored to their
  original text; THBS1, THBS4 and C1QB, previously ACCEPT or UNDECIDED. C1QB was not
  named in the review but has PS binding by IDA.
• Still over-annotated, with a softened template: the rest. The template now says
  only what holds: MISEV asks for evidence of EV association, a bulk proteome cannot
  separate surface binding from co-isolation, and the protein's annotations record no
  experimental proteoglycan, heparin or lipid binding. AGRN (ISS only), HSPG2 and MGP
  are in this group.
• Wording: category 2a now also covers cytosolic proteins recruited to membranes
  (ARF1, ARL6, ARL8A/B, APPL2, AP2M1, PARD6B, PLCG2, the V-ATPase V1 subunits).
• Script: the revert path now refuses to run on a row with an S quote drop, since
  it cannot restore dropped supported_by entries.

2026-10-10: second round of automated-review fixes

The automated reviewer re-checked the earlier fixes against MISEV2018 Table 3 and found
four places where a reason misstated the authority it cites. Fixed:

• Ribosomal proteins (RPS3, RACK1). Table 3 puts ribosomal proteins in category 3b,
  a co-isolated aggregate, not category 2. Their four vesicle rows return to
  MARK_AS_OVER_ANNOTATED, their pre-review action, and the files are back to their
  original text.
• GAS6 and APOH. Category 5 allows secreted proteins onto EVs through promiscuous
  lipid receptors, and both bind phosphatidylserine or anionic phospholipids. Their rows
  return to the original KEEP_AS_NON_CORE.
• Category 2a vs 2b. ESCRT subunits, annexins and HSPA8 now cite category 2a
  (lipid or membrane-protein binding) rather than 2b's "promiscuous" wording. ALIX's
  ACCEPT is restated as protein-specific (UniProt; PubMed:22660413), since MISEV puts
  ALIX in 2a with TSG101.
• APOB. Its one row is a urine sample, where MISEV's plasma/serum marker scope does
  not apply directly. It now has its own basis: APOB lipoproteins are not normally
  filtered into urine.
• Citation. MISEV2018 is cited with its PMID exactly once per reason.

2026-10-10: fixes from the automated PR review

The automated reviewer (ai4c-reviewer) found problems in how the manual passes were
written back. Fixed:

• APOE. Four non-plasma vesicle rows had been over-annotated as MISEV category 3,
  but MISEV does not name APOE, and each row cites van Niel et al. (PMID:26387950)
  showing ApoE on exosome-bound intraluminal vesicles. Those rows are restored to their
  original KEEP_AS_NON_CORE. Only the plasma-microvesicle row stays over-annotated, and
  its basis now says why. This page had also wrongly listed APOE and APOL1 among
  MISEV's named markers.
• Provenance. Every reason written by the passes now cites MISEV2018 as
  MISEV2018 (PMID:30637094).
• Stale summaries. The apply scripts rewrote action and reason but left
  summary, which then contradicted the action ("remains unresolved", "Retain"). Every
  changed row now gets a summary that matches its decision. The scripts are idempotent:
  a re-run edits nothing.
• Per-row wording. The vesicle decisions can now override a protein's call row by
  row (R). ALB's five rows each name their own sample.
• Contradicting quotes. Six supported_by entries were a previous reviewer's own
  argument for the action this review overturned (ARF1 ×3, GLA, PI15, PRDX4). They were
  dropped (S in the decisions files). The other inherited quotes were checked: they
  are source-paper text, UniProt text, or evidence the new basis already addresses
  (e.g. DNAJB11's epitope-tagged construct).
• CD2AP. Its rows are recorded as superseded by ai4curation/ai-gene-review#4287
  rather than edited.

2026-10-10: merge with main; CD2AP left to its newer review

Merging main brought in a separate CD2AP re-review (ai4curation/ai-gene-review#4287) that set its two exosome rows
(PMID:23533145, PMID:19056867) to UNDECIDED pending the source tables. The vesicle review
had made them KEEP_AS_NON_CORE. The newer, row-specific review was kept, so
vesicle_review/decisions.yaml overstates those two changes.

2026-10-10: settled vesicle rows reviewed by hand

At the user's request ("Yes"), extended the vesicle review to the 418 vesicle-type rows
already set to KEEP_AS_NON_CORE (220 genes). Dossiers were built with
build_dossiers.py --settled, which skips rows from the first pass, and every protein
was read. 104 rows changed (103 to MARK_AS_OVER_ANNOTATED, 1 to ACCEPT) and 314 were
confirmed. 59 of the 60 edited reviews validate (A2M's error predates this work). One
history record per touched gene (60). APOE's four rows were later restored (see the
next note), leaving 100 changed in 59 genes.

2026-10-10: vesicle rows reviewed by hand; vesicle rule withdrawn

At the user's request ("Review"), undid all 241 rows of the vesicle rule (byte-exact for
142 files), retired apply_dispositions.py, cached MISEV2018 (PMID:30637094) and
reviewed the 283 vesicle candidate rows protein by protein from dossiers
(vesicle_review/): 248 changed, 35 confirmed. 138 of 140 edited reviews validate (2
unrelated errors that predate this work). One history record per touched gene (163).

2026-10-10: membrane rows reviewed by hand; rule withdrawn

The user pointed out that the ID-based membrane rule still let a rule make the curation
call, with existing annotations treated as authoritative instead of as leads. Undid all
29 rows the rule had changed, restoring each to its original reviewer's action and
reason (byte-exact where the file had no other change). Withdrew rule M from
apply_dispositions.py; a dry run now changes nothing. Built per-row dossiers
(membrane_review/) and reviewed all 163 candidate rows individually: 66 changed, 97
confirmed. All 64 edited reviews validate. One history record per touched gene.

2026-10-06: membrane rule corrected to use identifiers

The user pointed out that the membrane-association check relied on labels: it
searched UniProt's location free text for the word "membrane". Replaced it with an
identifier-based test (UniProt anchor features, plus GOA rows to GO:0016020 or its
is_a/part_of closure in GO, excluding NOT and high-throughput rows). Rerunning the script
restored 11 rows to their original action and reason, re-noted 28 and added 1 (DNAJB11).
All 40 touched reviews validate. The script is idempotent: a further dry run changes
nothing.

2026-10-06: dispositions applied

The user agreed to the two default dispositions. Added htp/apply_dispositions.py
(minimal text edits, re-parse check per file, dry-run by default) and applied it: 280
rows in 183 gene reviews, one history record per gene. A first dry run would also have
changed proteins whose UniProt location names a membrane (e.g. AGPS peroxisome membrane,
PICALM cell membrane), as well as explicit ACCEPTs. The membrane rule was narrowed to
proteins with no documented membrane anchor or association, and ACCEPT rows were left in
place and listed for follow-up.

2026-10-05: created from RCA_EVIDENCE; HTP-family inventory

Renamed and broadened RCA_EVIDENCE into this project at the user's request, because
the omics analyses for RCA and HDA overlap. The RCA page and its scripts moved to
OMICS_EVIDENCE/rca.md and OMICS_EVIDENCE/rca/. Added htp/htp_inventory.py, which
covers 65,876 HTP-family GOA rows (50,795 distinct pairs) and 2,799 reviewed rows in
1,435 gene folders, and found the six uses above. HEP is left to the IEP project. No
gene reviews were edited.