Inferred from Expression Pattern (IEP) Evidence Code Review

Bottom line: IEP is the experimental evidence code that infers a gene's role in a process from a change in its own expression, so every IEP row carries a leap from "its abundance moved" to "it takes part". We surveyed all 25,401 IEP annotations in UniProt-GOA and reviewed the 550 IEP rows on 221 genes in this repo as of the 2026-09-13 survey (by 2026-09-26 the repo held 565 IEP rows across 235 gene reviews; the rates below have not been re-derived for those), including a tiered cohort of five miRNAs from one 130-gene batch. We did this to measure whether that leap holds, and to name the ways it fails. The characteristic IEP row is true but peripheral: reviewers accepted only 22.5% of IEP rows (the lowest of any code except IPI) and kept 55.6% as non-core, while 16.7% were flagged as over-annotated, modified or removed. Seven failure patterns recur, led by inducible bystanders (a detox enzyme collecting one response to X term per induction paper) and developmental time-courses read as tissue-building roles. Globally, 1,110 of the 1,147 cellular-component IEP rows that break GORULE:0000006 trace to a single ECO class (ECO:0000279) and could be cleared by one mapping change.

The dispositions come from this repository's own AI reviews, so the rates measure one primed reviewer population. The worked examples below carry the argument, and the open action items (a developmental cohort, the E. coli DNA-damage batch, independent disposition data) are listed at the foot of the page.

Overview

IEP (Inferred from Expression Pattern, ECO:0000270) is the one experimental
evidence code whose underlying observation is not about the gene product's
behaviour at all. IDA watches the protein do something; IMP removes it and
watches what breaks; IPI catches it holding a partner. IEP watches the gene's
own transcript or protein abundance go up or down, and infers participation in
whatever process the experimenter was manipulating.

That extra inferential step is the whole subject of this project. The review
question for an IEP row is therefore not the question asked of a propagated
annotation:

Is this gene product an agent in the annotated process, or is its
abundance merely modulated by it?

The same question is what GO itself asks. The GO best-practices
paper (PMID:23842463) puts
the bar explicitly:

"The ‘response to’ GO terms are intended to annotate gene products that are
required for the response to occur and are a direct result of the organism’s
reaction to the stimuli... It is acceptable to not annotate from such
expression studies since changes in expression of a gene product does not in
itself indicate its contribution to the function or process. Also, expression
studies can seldom support annotations to a Cellular Component or Molecular
Function term. Thus IEP should be used to annotate to terms in Biological
Process only."

The GO wiki entry for IEP
is blunter still — "Use this code with caution!" — and adds two operational
constraints that this project treats as testable: IEP is "usually used in
conjunction with high level GO terms in the Biological Process ontology",
and only normal expression counts (an overexpression or ectopic-expression
experiment is IDA or IMP territory, not IEP). The BP-only restriction in the
quote above is also a hard validation rule:
GORULE:0000006
enforces it for IEP and its high-throughput twin HEP.

Relationship to the sibling evidence-code projects

This page is the third in a set, and the contrast between them is the point.

Because there is no source annotation and no propagation edge to audit, the
review.propagation_review machinery that IBA and ISO reviews rely on only
partly fits IEP; see Action items.

Corpus Snapshot

Three views. IEP/iep_corpus_survey.py (output:
iep-corpus-survey.md) produces two of them: the
GOA view counts rows in the repo's cached *-goa.tsv downloads, and the
review view counts annotations in *-ai-review.yaml, i.e. what reviewers
concluded. Neither is a probability sample of IEP, because a gene directory
exists only because somebody chose that gene for review.

IEP/iep_global_atlas.py (output:
iep-global-atlas.md) supplies the denominator: the
complete set of IEP annotations in UniProt-GOA, downloaded from QuickGO. IEP
is rare enough that no sampling is needed — there are only 25,401 of them.

Within the repo, IEP is rare in the same way it is globally: 555 of 149,995
cached GOA rows (0.4%), roughly one IEP row per 40 IDA rows, spread over 220
gene directories (218 distinct accessions, two of which have a directory under
two different symbols). The review view finds 550 IEP rows in
221 review files. The two sides do not line up exactly, and neither is a subset
of the other: cached GOA files and reviews are refreshed independently, so a
review can retain a row that a later GOA download dropped, and a GOA row can
arrive after the review that would have carried it. The gap is small — under 1%
of rows, one review file — but it means the two views should be read as two
measurements of the same thing rather than as one filtered from the other.

Review-view denominators include all entries in existing_annotations, including
reviewer-proposed NEW annotations. The disposition table reports NEW
separately; those entries are not necessarily annotations supplied by GOA.

The reviewed sample is 2% of global IEP, so before drawing conclusions from it,
see how representative it is. The short
answer: close to global shares for coarse term types, badly skewed by organism
and annotation group. Matching those marginal shares does not establish that
reviewer flag rates are representative.

What "flagged" means, and does not mean

Every disposition on this page — every ACCEPT rate, every flag rate, the
cross-code comparison in the next section — is read out of this repository's own
*-ai-review.yaml files. Those are AI-generated reviews, written under the
project's reviewing guidance, which tells reviewers that "many GO terms are
over-annotations" and that they should "not take existing annotations as gospel,
whether experimental or bioinformatic". Three consequences follow, and they
apply to the recommendations at the foot of this page as much as to the tables:

• This is not independent adjudication of GO curators. It is one reviewer
  population's opinion, and that population was primed to look for
  over-annotation. A cross-code ACCEPT-rate table is therefore partly a
  measurement of reviewer priors, not only of the codes.
• The priors do not single IEP out, but the vocabulary favours the finding.
  The guidance is evidence-code-agnostic — it names experimental and
  computational codes alike as fallible — which is why the cross-code
  ordering is more trustworthy than any single rate: every code was judged
  under the same instructions. What is not neutral is that the action
  vocabulary offers KEEP_AS_NON_CORE and MARK_AS_OVER_ANNOTATED as
  first-class verdicts, making "true but peripheral" an easy call to reach —
  and "true but peripheral" is exactly this page's headline conclusion about
  IEP. A 16.7% flag rate should not be read as "16.7% of IEP annotations in GOA
  are wrong".
• The qualitative claims are the load-bearing ones. The failure patterns
  below are argued from named genes, named papers and quoted reviewer reasons,
  each of which can be checked against the source. The statistics describe how
  often those patterns turned up in this corpus; they are not independent
  confirmation that the patterns are real.

Where a claim survives only in the statistics and not in the worked examples, it
is marked as suggestive in the text.

IEP is used to say one kind of thing

Branch assignment is first match wins in the order above, so a term parented
under both response to stimulus and developmental process — a defence
response that is also a developmental one, say — is counted as
stimulus-response. Those dual-parented rows are excluded from the developmental
bucket. A different bucket priority could change the flag-rate comparison; its
effect has not been measured. The repo shares here are the review view (550
rows); the representativeness tables further down use the GOA view (555 rows),
which is why the same stratum can differ by a point.

Seven out of ten IEP annotations are a "response to X" term, both globally and in
the repo, which follows directly from the experiment type: expose an organism to
a stimulus, see which transcripts move. Globally the most frequent terms are
response to xenobiotic stimulus (512 rows), response to cold (387),
response to bacterium (384) and response to abscisic acid (379).

The flag rates in that table hint at a second-order finding — the developmental
branch may be the riskier one — but the corpus is too small to establish it.
19 of 84 developmental rows were flagged against 64 of 381 stimulus rows, a
difference not separable from noise (two-sided Fisher exact p = 0.21). Splitting
by species does not establish a general developmental excess: the direction
repeats in the three largest species (rat 33.3% vs 25.0%, human 20.6% vs 10.4%,
Arabidopsis 40.0% vs 8.6%), but none of those contrasts reaches significance.
It reverses in Dictyostelium (10.0% vs 18.2%) and in mouse, where 0/4
developmental rows are flagged against 8/8 stimulus rows (Fisher p = 0.002).
The mouse result is an opposite-direction contrast in only 12 rows, not evidence
for a general developmental risk. The developmental rows span 54 gene
directories across seven species, but that spread does not rule out review-batch
or gene-selection confounding.

The mechanistic reason to expect the gap is independent of the numbers, and it
is argued below from worked examples rather than from the flag rate: a "response
to X" row is usually at least true — the transcript really did move when the
stimulus was applied — whereas an "X development" row inferred from a
developmental time-course is a genuine over-reach whenever rising abundance as a
tissue matures reflects demand for the enzyme's product rather than an
instructive role in building the tissue.

The disposition data: IEP is not wrong so much as peripheral

This corpus's reviewers flagged 16.7% of IEP rows (REMOVE +
MARK_AS_OVER_ANNOTATED + MODIFY). That is worse than the other experimental
codes (IDA 6.8%, IMP 6.8%, IGI 6.4%) but comparable to ISO (16.8%) and better
than IEA (20.4%) — not, on its own, a damning number.

The sharper contrast is at the other end of the distribution:

The % core column credits a code whenever a term it carries also appears in
core_functions, even when the term got there on the strength of a different
code annotating it too. It is therefore generous to every code, and most
generous to codes that co-annotate often — including IEP, whose 10.0% falls to
9.1% if only rows the reviewer also ACCEPTed are counted.

Read with the limitations above in
mind, IEP has the lowest ACCEPT rate and the lowest core-function grounding rate
of any code surveyed except IPI — and IPI's position is a known artifact of
protein binding rather than a property of physical-interaction evidence. The
missing IEP mass went to KEEP_AS_NON_CORE, which absorbs 55.6% of IEP
rows, the highest share of any code.

So the characteristic IEP annotation is not false. It is true and peripheral:
a real observation about how the gene is regulated, phrased as a claim about
what the gene does. That is a harder problem than outright error, because
nothing in the annotation is checkably wrong.

Two more measurements sharpen it:

• 72.4% of IEP rows are the sole carrier of their GO term in the review
  — no other evidence code in the same gene supports that term. IEP is not
  mostly redundant confirmation of what IDA/IMP already say; it is mostly
  adding terms nothing else supports. That figure tests exact term-id
  equality, so it is an upper bound: an IEP response to heat sitting beside
  an IDA cellular response to heat counts as sole. Allowing an ancestor or
  descendant under a non-IEP code to corroborate drops it to 53.5% — still a
  majority, and still the substantive point, but the honest number to quote for
  "IEP is load-bearing" is the closure-aware one.
• IEP clusters heavily. The median IEP-carrying gene has one IEP row, but the
  top 10% of them (23 of 221 genes) carry 42.7% of all IEP rows. rat/Hmgcs2
  alone has 33; rat/Hspa8 25; rat/Casp3 23; rat/Tp53 18. A gene that attracts
  stimulus-response papers accumulates a proportional cloud of IEP terms.
  Globally the same shape holds almost exactly: 21 repo genes have 5 or more IEP
  rows, 9.5% of IEP-carrying genes carrying 41.3% of rows, against 1,117 gene
  products globally — 10.5% of IEP-carrying products, accounting for 42.3% of
  all IEP annotations.

Is the reviewed sample representative?

The reviewed corpus is 2.1% of global IEP and was assembled by gene-review
interest, not by sampling IEP. The
global atlas
compares each stratum's repo share against its global share; a ratio of 1.00x
means proportional, above 1 over-sampled, below 1 under-sampled.

Close to global marginal distributions. Several coarse annotation shares
come out close to proportional:

The global counts independently support two descriptive findings: IEP is
overwhelmingly a stimulus-response code, and its annotations concentrate in a
few genes. They do not measure reviewer dispositions. Similar branch shares
therefore cannot establish that developmental annotations are riskier or rule
out selection and review-batch effects on the flag-rate comparison.

Badly skewed by organism and annotation group. Here the sample is a poor
picture of IEP:

Entirely absent from the repo: AgBase (785 rows), ZFIN (326),
CollecTF (211); Gossypium hirsutum (395), Danio rerio (335),
Gallus gallus (279), M. tuberculosis H37Rv (179).

The human over-sampling is expected — the repo is human-centric. The coarse
term-type shares above remain close to the global shares, but organism
imbalances still limit comparisons of reviewer dispositions.

The MGI row is the one that has moved, and it is worth saying why, because it
shows what targeted sampling buys. MGI matters disproportionately: the largest
single-screen IEP batches in all of GOA are MGI's (see
pattern 3),
so the stratum was exactly where the most extreme instance of a pattern this
page documents actually lives. Before the miRNA cohort,
the repo sampled MGI at 0.05x — 1 row against a 3.4% global share. Five
targeted reviews took it to 0.73x, close to proportional at the row level. At
the product level the gap is barely dented: 5 reviewed of the 437 RNAcentral
gene products, and 3 of 294, 5 of 139 and 2 of 103 rows for the three batch
terms those reviews touched. Row-share parity is not cohort coverage.

A slice the repo under-samples but can represent. 840 global IEP rows
(3.3%), on 437 gene products, are RNAcentral entries rather than proteins —
almost all microRNA precursors from miRNA-profiling studies. This is IEP in its
purest form: the only observation is that a non-coding RNA's abundance changed.
The genes/ tree handles ncRNA entries natively (id: URS…, product_type: MIRNA, fetched with ai-gene-review fetch-ncrna), so the gap was coverage, not
capability. Five of them are now reviewed — see
the miRNA cohort below.

Corrections this forces. Two claims made from the repo sample alone need
restating:

1. Aspect violations are not a curiosity. The repo's 23 non-BP IEP rows looked
   like a rounding error. Globally there are 1,223 (4.5% CC, 0.3% MF) — and
   the repo sample was proportionally accurate (1.01x for BP, 0.80x for CC) all
   along. The mechanism proposed from 20 rows holds at scale: see
   below.
2. No GO term is majority-IEP. Within a gene, 72.4% of IEP rows are the sole
   carrier of their exact term (53.5% with closure). But measured per term across
   the whole ontology, IEP is never the dominant support: the most IEP-dependent
   frequent term is seed trichome elongation at 19.6% of its annotations,
   then response to ethanol (15.5%) and
   cellular response to leukemia inhibitory factor (13.7%); most sit below 2%.
   (These shares alone among the atlas figures are not frozen by the committed
   snapshot — the per-term totals come from a live QuickGO count — so they drift
   by a point or two between runs.) Both statements are true and they answer
   different questions — IEP is load-bearing for the gene it sits on, never
   for the term it points at.

Failure Patterns

1. Inducible bystander — the "response to X" cloud

The largest and most systematic class. A detoxification or intermediary
metabolic enzyme has one stable job; because that job is useful under stress,
its transcript is induced by a long list of chemically unrelated insults; and
because each induction was published separately, each becomes an annotation.

rat/Gstt1 carries response to salicylic acid, response to selenium ion,
response to vitamin E and response to xenobiotic stimulus — four separate
PMIDs, one enzyme, one activity. rat/Qdpr, whose actual job is quinonoid
dihydrobiopterin reduction in tetrahydrobiopterin recycling, carries response to aluminum ion, response to lead ion, response to glucagon, cellular response to xenobiotic stimulus and liver development. rat/Hsd11b2 has six
flagged stimulus terms.

What makes these hard is that they are not false. GSTT1 activity is part of
how a cell handles a xenobiotic. The problem is proportion: the annotation set
implies a stimulus-specialist when the biology is one broad-specificity
transferase. GO's own criterion — "required for the response to occur" — is the
right discriminator, and it is exactly the question the induction experiment
does not answer.

A sharp variant is the co-exposure artifact. rat/Gsta4's response to zinc ion and response to herbicide come from the same zinc/paraquat co-exposure
experiment (PMID:20553223); the reviewer noted there is "no mechanistic evidence
that zinc directly modulates GSTA4-4". A factorial exposure design yields one
term per factor regardless of which factor drove the induction.

2. Developmental time-course → tissue-development term

rat/Ckmt2 is the cleanest instance. Sarcomeric mitochondrial creatine kinase
mRNA is undetectable in prenatal heart and rises sharply after birth
(PMID:8086475), so it carries heart development and skeletal muscle tissue development. As the review puts it, that profile "reflects the maturing heart's
increasing metabolic demand for phosphocreatine buffering, not a direct
instructive role of Ckmt2 in heart morphogenesis."

The same shape recurs: rat/Ephx1 and rat/Qdpr → liver development, rat/Gamt →
embryonic liver development, rat/Hmgcs2 → lung development and adipose tissue development, rat/Pgam2 → spermatogenesis. In every case the protein is
a metabolic enzyme whose product the maturing tissue needs more of. The tissue
builds the enzyme; the enzyme does not build the tissue.

This is the mechanism behind the developmental branch's higher flag rate (22.6%
versus 16.8%, though as noted that difference is not statistically separable
from noise on 84 rows): a metabolic enzyme genuinely participates in a stress
response in a way it does not participate in organogenesis. The six genes
above are the evidence; the flag rate is a description of how often the pattern
came up, not a demonstration that it exists.

3. Differential-expression screen batch: one screen, one term, N genes

PMID:21492153 is a 2-D gel proteomics comparison of proliferating versus
differentiated Caco-2 intestinal cells. It reports "53 proteins that were
differently regulated during the differentiation process", 34 of them identified
by MALDI-TOF, and those identifications were curated involved_in
GO:0030855 epithelial cell differentiation with IEP. In
this corpus that single paper is the source for eight genes:
human/ACADVL, human/ACTL8, human/CDK1, human/CPT1A, human/FN3K, human/LGALS3,
human/PGK1 and human/TOLLIP — a very-long-chain acyl-CoA dehydrogenase, a
cyclin-dependent kinase, a glycolytic kinase, a fructosamine kinase, a galectin
and a TLR adaptor. Six of the eight were flagged (five MARK_AS_OVER_ANNOTATED,
human/FN3K REMOVE); the other two, human/LGALS3 and human/TOLLIP, were kept as
non-core with reviewers noting the evidence is "correlative expression-pattern"
and the process "secondary" to the protein's actual role.

human/CDK1 is the case that exposes the underlying logic error. The review notes
that CDK1 "promotes proliferation, which decreases during differentiation" —
CDK1's abundance is anti-correlated with its causal contribution. The paper
itself says as much: "proteins associated with proliferation, cell growth and
cancer were downregulated, reflecting the loss of the tumorigenic phenotype of
the cells." An abundance change carries no sign information about the direction
of the causal role, so "changed during process X" and "promotes process X" are
simply different claims — and here the annotation was made from a change in the
wrong direction.

This pattern is the IEP analogue of the SPKW mapping layer: a single upstream
decision (here, "annotate every hit in this screen") propagates a term across a
functionally unrelated gene set, and the resulting annotations look independent
because they sit on different genes.

At global scale this is the dominant shape of IEP, and eight genes was a small
example. The distribution of IEP over references is sharply bimodal: 62.4% of
the 11,777 references behind global IEP contribute exactly one annotation,
while the top ten contribute 1,130 between them. Every one of the six largest is
a single screen annotated to a single term:

Three of those are miRNA-profiling studies annotating differentially expressed
miRNA precursors, so "changed abundance" is the entire observation. And
PMID:11967071 is the clearest single case in all of IEP: a genome-wide E. coli
DNA-damage transcriptome, whose own title concedes that "over 1000 genes are
involved", yielding DNA damage response for 152 genes including the maltoporin
lamB, the maltose transport subunit malF, fumarate reductase frdA,
D-serine deaminase dsdA and asparagine synthetase asnA. Those are not DNA
repair proteins; they are genes whose transcripts moved. This is
pattern 5 executed 152
times from one experiment.

Two review consequences follow. A batch-sourced IEP row should be judged against
its cohort, not on its own, because the cohort reveals the annotation rule that
produced it. And because these batches are concentrated in MGI, which the repo
sampled at 0.05x before this project touched it, the reviewed corpus was blind
to the most extreme form of the pattern; the cohort below is the start of fixing
that, and it moved the MGI row share to 0.73x on five reviews.

Cohort review: 130 miRNAs, one term, one paper

PMID:25858512 is a natural experiment, because the paper itself sorts its miRNAs
into tiers. It detected 372 miRNAs in hippocampal slices, found that only
12 changed during LTP, and functionally validated three — miR-26a,
miR-384-5p and let-7a — by electrophysiology, time-lapse spine imaging and 3'
UTR reporter assays. MGI annotated 130 miRNA precursors to
long-term synaptic potentiation. The paper's own abstract describes what the
rest amount to: "presents a catalogue of candidate 'LTP miRNAs'".

Five members were reviewed, one from each tier, so that the batch is tested with
an internal positive control rather than assumed to be wrong:

The tier predicts the verdict exactly. This matters because it separates two
claims that are easy to conflate: the batch is not wrong because it is a batch,
it is wrong for the members whose only qualification is having been detected.
mouse/Mir30e is the instructive middle case — its expression genuinely changed,
so IEP is the correct evidence code and the observation is sound; what fails is
the leap from "changed" to "acts upstream of or within".

The batch also has a defect visible only from the cohort: it omits let-7a,
one of the three miRNAs the paper establishes as required, while including 127
that it does not. The annotation set is not merely over-inclusive, it is
misaligned with the paper's conclusions at both ends.

Three further batch papers annotate these same five miRNAs, and two produced
sharper findings than the LTP batch itself:

• A source that refutes its own annotation. mouse/Mir26a-1 carries
  response to bacterium from PMID:24205035, a circulating-miRNA survey in a
  cecal-ligation-and-puncture sepsis model. The authors asked whether bacterial
  sensing drives the increase and concluded it does not — it persists in Tlr2,
  Tlr4 and NF-kB knockouts, indicating the change "was not directly mediated
  by the TLR2/NF-κB or TLR4/NF-κB pathway, and pathways induced by exposure to
  the gram-positive or gram-negative bacteria". The annotation asserts precisely
  what the cited paper tested and rejected, which is what makes it a REMOVE
  rather than an over-annotation. The companion response to wounding row
  describes the surgery used to perturb the animals.
• The pattern inside a single paper. PMID:20439489 supports both a validated
  IMP row on mouse/Mir100 (positive regulation of stem cell differentiation,
  via direct repression of Smarca5 — miR-100 is one of the paper's three title
  miRNAs) and an unvalidated IEP batch row (cellular response to leukemia inhibitory factor, one of 291). The same study, the same gene, one row
  recording what was demonstrated and one recording what was merely observed.
  The IEP term is also a poor description of the experiment, which induced
  differentiation by withdrawing LIF.
• Wrong direction as well as wrong evidence. mouse/Mir100's
  sensory perception of sound comes from an ageing survey (PMID:23646144)
  whose full text never mentions miR-100 and whose subject is degeneration of
  the organ of Corti during hearing loss, not normal hearing.

Across the 13 IEP rows in this cohort: 2 ACCEPT, 7 MARK_AS_OVER_ANNOTATED,
4 REMOVE — an 85% flag rate against 16.7% corpus-wide, and 4 REMOVEs added to a
corpus that previously held 7 in total. Targeting batch cohorts rather than
individual rows is therefore a high-yield review strategy, which is the practical
lesson for the remaining candidates.

4. Promiscuous hub inversion: the regulator annotated as a responder

ARATH/PIF3 carries seven IEP rows — response to heat, response to cold,
response to ethylene, response to auxin, response to abscisic acid,
response to salt, response to water-immersion restraint stress — six of them
from a single gene-family expression survey (PMID:23708772). All seven were
flagged.

PIF3 is a phytochrome-interacting transcription factor: it runs the light and
hormone response programmes. Its transcript answering every stimulus is what a
signalling hub's transcript does. Annotating the hub as a responder to each
stimulus inverts the regulator/target relationship and buries the actual
function, which the gene's IMP annotations already carry. The seventh row is a
granularity failure rather than an inversion, and is taken up in
pattern 7.

5. Regulon membership is a property of the promoter

ECOLI/arnF is annotated response to iron(III) ion because the arnBCADTEF
operon is induced by iron through the BasS-BasR two-component system. The
review's verdict
is precise: "the IEP evidence code is technically appropriate... However,
annotating a gene to 'response to iron(III) ion' based solely on transcriptional
induction conflates regulation with function. ArnF is a flippase that
translocates undecaprenyl phosphate-alpha-L-Ara4N; it does not participate in
iron sensing, binding, or detoxification." The iron-responsiveness
belongs to the operon's promoter; the flippase does not sense or handle iron.
(This case also appears in the IBA project,
where the same gene's IBA rows are analysed.)

yeast/THI22 is the same shape with an extra twist: thiamine-dependent regulation
is real, but the paper that establishes the regulation (PMID:10383756) also
establishes that THI22 is not required for thiamine biosynthesis, so the
regulon membership points at a process the gene demonstrably does not carry out.

6. Marker-gene circularity in Dictyostelium

Dictyostelium development is staged, and stage-specific genes are used as
stage markers precisely because their expression tracks the stage. A
developmental transcriptome (PMID:25887420) supplies IEP rows for 17 genes here.

The circularity has to be judged case by case, and this corpus contains both
verdicts. DICDI/cotB is a prespore marker: the review flags
slug development involved in sorocarp development because "there is no
evidence that the SP70 protein participates causally in slug development... the
gene's core function lies in spore coat structure, not in the morphogenesis of
the slug."
DICDI/mhcA gets the same treatment for aggregation involved in sorocarp development. But DICDI/ecmB was accepted as core for culmination involved in sorocarp development — ecmB is a prestalk extracellular-matrix protein, so
the late-development induction the transcriptome records is the production of
the material culmination consumes. The marker is the product.

7. Wrong-granularity terms, in both directions

PIF3's seventh row is the over-specific case: response to water-immersion restraint stress (GO:1990785) is a rodent stress-model term applied to a plant
submergence experiment, and DICDI/acaA's response to imidacloprid names one
insecticide from one exposure.

The commoner error runs the other way. ARATH/CRY1 and ARATH/CRY2 each carry a
generic response to light stimulus IEP row, both MODIFYed — and the reviewers'
stated reason is granularity, not agency: the cryptochromes really are
blue-light photoreceptors, so "responds to light" is true and merely
uninformative next to the response to blue light and blue light signaling pathway terms proposed in its place. CRY2's separate IEP rows on response to blue light and response to low fluence blue light stimulus by blue low-fluence system were both ACCEPTed, which is the same judgment made from the other side:
IEP on a photoreceptor's own stimulus is fine once the term names the stimulus
the protein actually senses. ARATH/SOC1's IEP row on positive regulation of DNA-templated transcription was likewise MODIFYed to the Pol II-specific child.

Neither direction warrants REMOVE. The observation is sound and the term is in
the right lineage; only its level is wrong, which is precisely what MODIFY says.

GORULE:0000006 violations are an ECO-mapping artifact

Globally 1,223 IEP annotations (4.8%) sit on non-BP terms, violating the hard
validation rule: 1,147 cellular-component and 76 molecular-function. They are not
sloppy curation, and the diagnosis is worth recording because it is mechanical
and it accounts for nearly all of them.

"IEP" in a GAF is not one ECO class. Of the 25,401 rows, 23,747 (93.5%) are
literally ECO:0000270; the other 1,654 use a more specific descendant class
that collapses to IEP in the GAF projection. Splitting the aspect violations by
ECO class localises the problem almost perfectly:

The generic parent class is 99.8% BP — essentially rule-compliant. One descendant
class, ECO:0000279, contributes 1,110 of the 1,147 CC violations in GOA
(96.8%), and 1,132 of all 1,223 aspect violations (92.6%), while being 5.6% of
IEP.

The repo's 23 examples are that global picture in miniature (its aspect mix is
proportional to global at 1.01x for BP and 0.80x for CC). Twenty are SynGO
cellular-component annotations (postsynaptic density,
glutamatergic synapse, presynapse, presynaptic active zone) on rat/Hspa8,
rat/Mapk1, mouse/Hspa8, mouse/Casp3, mouse/App, mouse/Notch1, human/APOE and
human/HTT. In the cached GOA rows these carry ECO:0000279, "qualitative
western immunoblotting evidence used in manual assertion" — the evidence class
for detecting a protein in a biochemically fractionated preparation such as a
synaptosome or PSD prep.

ECO:0000279
has two relevant ancestors, and they map to different GAF codes:
ECO:0000314 (direct assay evidence used in manual assertion → IDA) is a direct
parent, and ECO:0000270 (expression pattern evidence used in manual assertion
→ IEP) is reached one step further up, via
ECO:0000284
"protein expression evidence used in manual assertion". Collapsing the specific ECO class down
to a three-letter GAF code forces a choice between them, and the pipeline picks
IEP — dragging a localization assay into a BP-only code. GORULE:0000006 itself
names the correct resolution: "For CC annotations that assess the localization of
a gene product, IDA should be used." The remaining three violations are DisProt
molecular-function rows (human/NFE2L2 ubiquitin protein ligase binding,
human/BAG6 molecular function activator activity, human/ADAM10 protein homodimerization activity); human/BAG6's was reviewed and REMOVEd.

The lesson generalises beyond IEP: where a specific ECO class is multiply
parented across GAF code boundaries, the GAF round-trip is lossy, and a
downstream hard rule then fires on an annotation whose evidence was never
actually expression-pattern evidence.

Where IEP Is Legitimate

IEP is not a code to review adversarially. 22.5% of IEP rows were accepted, and
55 IEP rows ground a term in a gene's core_functions — 50 of them rows the
reviewer both accepted and let reach core, the other 5 riding on a co-annotated
non-IEP row for the same term. The accepted cases share
a single property, and it is the discriminator this project turns on: the
gene's job is the response itself.

• Inducible stress-response effectors. For a heat-shock protein, induction
  by heat is not a correlate of the function — being made when the cell is hot
  is the function. worm/hsp-16.2, ECOLI/DnaK, ECOLI/DnaJ and ARATH/HSP17.6A
  all keep response to heat; worm/hsp-4 keeps endoplasmic reticulum unfolded protein response (four separate accepted rows); worm/hsp-6 and worm/hsp-60
  keep mitochondrial unfolded protein response. These are regulon outputs
  whose entire deployment is the stress programme.
• Infection-inducible immune effectors. worm/irg-1 and worm/irg-2 keep both
  antibacterial innate immune response and defense response to Gram-negative bacterium: strong, specific, replicated induction by P. aeruginosa is the
  defining property of an infection-response gene. worm/lys-1 is the same call
  with independent backing: the same paper that shows infection-inducible
  expression also shows that overexpression confers resistance.
• Clock genes. ARATH/CCA1 and ARATH/TOC1 keep circadian rhythm. Oscillating
  expression is not evidence about a clock component; it is the mechanism by
  which the oscillator works.
• Corroboration inside a converging set. ARATH/UVR8 (response to UV-B),
  ARATH/SOS1 (response to salt stress), ARATH/CBF1 (cold acclimation),
  MEDTR/NFP (nodulation), worm/skn-1 (response to oxidative stress) and
  worm/xbp-1 (IRE1-mediated unfolded protein response) all keep IEP rows that
  restate what IMP and IDA independently establish. Here IEP costs nothing and
  adds a line of evidence — the 27.6% of IEP rows that are not the sole carrier
  of their term on an exact term match, or 46.5% once a non-IEP annotation on a
  parent or child term counts as corroboration.
• Markers that are the differentiated product. DICDI/ecmB, as above.

The contrast with the failure cases is not about evidence quality; the
induction experiments behind rat/Gstt1 are as sound as those behind worm/hsp-4.
It is about whether the gene product exists for the annotated process. A
heat-shock chaperone is deployed only during heat shock. A glutathione
transferase does the same chemistry whether or not the animal was dosed with
selenium.

Reviewer Checklist

Before accepting or flagging an IEP row.

Read the experiment, not just the term

• Identify what was measured (transcript or protein), by what method, and under
  what perturbation. Northern/microarray induction, a developmental time-course,
  and a differential-abundance proteomics screen fail in different ways.
• Check the paper is a normal-expression study. Overexpression and ectopic
  expression are IDA/IMP evidence, not IEP, per the GO wiki.
• Check whether one paper supplies IEP rows for many genes (a screen) or many
  terms for one gene (a stimulus cloud). Both are batch artifacts and neither is
  visible from a single annotation row.

Ask the agency question

• Is the gene product required for the annotated process to occur, or does
  the process merely change its abundance? This is GO's own wording and it
  decides most cases.
• If the gene is a regulator, check for inversion: does it drive the response
  programme it is annotated as responding to (ARATH/PIF3)?
• If the gene is an enzyme, ask whether the annotated process needs more of its
  normal product (rat/Ckmt2 in maturing muscle) rather than a distinct
  activity.
• Check the sign. An abundance change carries no direction — human/CDK1 falls
  during the differentiation it is annotated to.
• Distinguish regulon membership from function: an iron-responsive promoter does
  not make an iron-responsive protein (ECOLI/arnF).

Check the term and the aspect

• Verify the aspect is Biological Process (GORULE:0000006). If it is CC or MF,
  check the underlying ECO class before blaming the curator — the GAF collapse
  may have chosen the wrong parent.
• Check granularity in both directions. GO advises high-level terms for IEP;
  a hyper-specific stimulus term from one exposure experiment (response to imidacloprid) is over-reach, and the correct fix is usually MODIFY to the
  parent rather than REMOVE.
• Check whether the term already has non-IEP support in the same gene, and check
  the parents and children too, not just the exact term. If something in that
  lineage is covered by IDA/IMP, the IEP row is cheap corroboration; if nothing
  is (53.5% of rows), the IEP row is load-bearing and deserves the full agency
  question.

Prefer the right action

• KEEP_AS_NON_CORE is the correct default for a sound induction observation
  that is not the gene's job. It is already the modal IEP outcome and should
  stay that way.
• MARK_AS_OVER_ANNOTATED fits when the observation is real but the term
  overstates the role — the inducible-bystander and developmental cases.
• REMOVE should be reserved for a categorical mismatch, where the process is
  not merely peripheral but belongs to a different functional class entirely
  (human/RB1 Ras protein signal transduction, human/FN3K epithelial cell differentiation). Only 11 of 550 rows met that bar, four of them added by the
  miRNA cohort.
• Do not REMOVE an IEP row merely because the evidence code is weak. The
  induction happened; what is at issue is the term, not the experiment.

Recommendations

For reviewers. Treat an IEP row as a statement about regulation and ask
whether it has been mis-phrased as a statement about function. The most common
correct answer is "true, keep it, but it is not what this gene is for."

For curators and GO.

1. Batch-annotating a screen deserves a second look. The signature is
   mechanical and cheap to detect: one reference, one GO term, many gene products.
   Globally the top six such references produce 72–291 annotations each, and
   reference × term × gene-product-count would flag every one of them at
   submission time rather than at review time. PMID:11967071 is the case for
   doing so — a paper titled "Over 1000 genes are involved in the DNA damage
   response of E. coli" should not yield DNA damage response for a maltoporin.
2. Cap stimulus clouds. When a gene accumulates many response to X terms
   from independent exposure papers and has a single well-characterised
   activity, the informative annotation is the parent term once, not the
   catalogue. rat/Gstt1's four sibling stimulus terms say less together than
   response to xenobiotic stimulus alone.
3. Developmental time-courses may need a higher bar than stimulus responses.
   The mechanism is clear enough — metabolic demand tracks tissue maturation
   without any instructive role, as rat/Ckmt2, rat/Ephx1, rat/Qdpr, rat/Hmgcs2,
   rat/Gamt and rat/Pgam2 each show individually. The corpus-level flag-rate gap
   (22.6% versus 16.8%) points the same way but is not statistical support:
   at 19/84 versus 64/381 it is within noise (Fisher p = 0.21). This
   recommendation rests on the worked cases, and testing it properly needs a
   developmental-branch cohort sampled for the purpose.
4. Fix the ECO→GAF collapse rather than the annotations. The GORULE:0000006
   violations come from a multiply-parented ECO class (ECO:0000279) whose GAF
   projection picks IEP over the equally valid IDA. Choosing the parent by the
   annotation's aspect would clear 1,110 of the 1,147 CC violations in GOA
   in one change, with no curation effort at all.
5. response to terms could carry the requirement explicitly. GO's
   best-practice text says these terms are for products "required for the
   response to occur", but nothing in the term definitions or the evidence code
   enforces it. An annotation-extension or qualifier distinguishing "acts in the
   response" from "is induced during the response" would let the two claims
   coexist instead of competing.

Action Items

• [x] Build a reproducible two-view IEP survey (GOA rows + reviewed rows) with
  an evidence-code baseline rather than IEP in isolation.
• [x] Quantify the ACCEPT / core-function-grounding gap that distinguishes IEP
  from the other experimental codes.
• [x] Catalogue the failure patterns against reviewed examples.
• [x] Diagnose the GORULE:0000006 aspect violations.
• [x] Establish the global denominator (all 25,401 IEP annotations) and measure
  the reviewed sample's bias against it, so page claims can be marked as
  sample-robust or sample-dependent.
• [x] Make every figure on this page regenerable from the committed scripts, and
  attach a significance test to the one comparison that carries a
  recommendation.
• [ ] Sample a developmental-branch cohort for the purpose, so the branch
  flag-rate difference is either established or dropped. 4,110 global rows
  are available and the current 84 are far too few (Fisher p = 0.21).
• [ ] Get a disposition signal that is not this project's own reviewers — e.g.
  compare against annotations GO itself has since removed or re-coded —
  so the limitations on the
  cross-code table can be lifted rather than merely stated.
• [ ] Review the remaining 22 PENDING IEP rows, all on rat/Casp3 — the largest
  single unreviewed block already in the repo.
• [x] Review a batch cohort rather than another single gene, sampling tiers
  within the cohort so the batch is tested rather than assumed. Done for
  PMID:25858512 — see
  the miRNA cohort.
• [ ] Do the same for PMID:11967071 (152 E. coli genes → DNA damage response), the cleanest test of the regulon-membership pattern, in an
  organism this repo already covers well. The tiering is available there too:
  the paper distinguishes SOS-regulon members from the wider induced set.
• [ ] Continue closing the MGI gap at the product level. The five miRNA
  reviews took MGI's row share from 0.05x to 0.73x of its 3.4% global share,
  which flatters the coverage: cellular response to leukemia inhibitory factor has 3 rows reviewed of 294, long-term synaptic potentiation 5 of
  139, sensory perception of sound 2 of 103.
• [ ] Review the IEP-heaviest gene products, none of which are in the repo: rat
  Ppargc1a (50 IEP rows), Il6 (48), Nos3 (45), Serpine1 (44), Tnf (43), Star
  (43), Hif1a (41), Ccl2 (40), Il1b (39). These are the extreme form of the
  inducible-bystander pattern — cytokines and stress hubs whose transcripts
  move under every perturbation — and rat/Hmgcs2's 33 rows, the repo's
  current maximum, is smaller than any of them.
• [ ] Work through data/iep_review_candidates.tsv:
  4,807 not-yet-reviewed gene products sampled up to 3 per global term
  stratum, so the long tail of terms is reachable and not just the head.
• [ ] Decide whether IEP warrants structured review fields. PropagationReview
  is documented as covering "a propagated or inferred annotation" and its
  TERM_SCOPING_PROBLEM, NO_FAILURE_NON_CORE, ROLE_CONFLATION,
  GRANULARITY_MISMATCH and CONTEXT_OR_TISSUE_MISMATCH values already fit
  IEP, but there is no value for the central IEP failure — correlation
  asserted as participation. Candidate additions: an
  EXPRESSION_CORRELATION_ONLY failure mode and a REGULON_MEMBERSHIP
  subtype.
• [ ] Extend the survey to HEP (the high-throughput IEP twin), which shares
  GORULE:0000006 and every pattern on this page but is generated at screen
  scale.
• [ ] Check the reviewed IEP corpus against the "normal expression only" rule —
  any row whose source paper is an overexpression study is mis-coded, not
  merely over-annotated.
• [x] Establish that the RNAcentral slice (840 rows, 437 miRNA precursors) is in
  scope. It is, via ai-gene-review fetch-ncrna; five miRNA reviews now
  exist. 432 gene products remain.

Session Notes

2026-09-13 (review follow-up — reconcile dispositions and branch claims)

• Clarified that review-view totals include reviewer-proposed NEW rows, which
  the disposition table now displays separately. The totals are unchanged.
• Corrected the claim that no species stratum reaches significance: mouse has
  an opposite-direction contrast (0/4 developmental versus 8/8 stimulus rows
  flagged; Fisher p = 0.002). The pooled developmental excess remains
  inconclusive (p = 0.21), and gene and species coverage does not eliminate
  review-batch confounding.
• Removed the inference from matching global branch proportions to reliable
  branch-specific flag rates. The global atlas measures annotation composition,
  not independent error rates. Earlier session claims are corrected below.

2026-09-05 (fourth pass — statistics tightened, review response)

Nothing in the qualitative argument changed; what changed is how confidently the
numbers behind it are stated and whether the committed script regenerates all of
them.

• Retracted the developmental-branch finding as a statistical result. The
  first pass reported 23.7% versus 15.1% and the second promoted it to a
  recommendation. On the current corpus it is 22.6% versus 16.8%, i.e. 19/84
  against 64/381, which a two-sided Fisher exact test puts at p = 0.21 —
  within noise. The survey script now computes that test, and splits both
  branches by species so batch confounding is visible: the direction repeats in
  rat, human and Arabidopsis and reverses in Dictyostelium. Correction
  (2026-09-13): this entry originally said no stratum was significant, omitting
  the opposite-direction mouse result (p = 0.002). The developmental rows span
  54 gene directories across seven species, not six; that coverage does not
  exclude batch confounding. Recommendation 3 rests on the six worked cases.
• Added a limitations section,
  What "flagged" means. Every
  disposition on the page comes from this project's own AI-generated reviews,
  produced under guidance that primes reviewers to look for over-annotation, so
  the cross-code ACCEPT-rate table is partly a measurement of reviewer priors.
  The guidance does not single IEP out — it names experimental and computational
  codes alike — which is why the cross-code ordering survives better than any
  single rate; what is not neutral is that KEEP_AS_NON_CORE and
  MARK_AS_OVER_ANNOTATED are first-class verdicts, and "true but peripheral"
  is precisely this page's conclusion about IEP.
• Partly closed the reproducibility gaps. The survey script added per-gene
  concentration statistics, the ACCEPT∩core count, and an UNREVIEWED column.
  Correction (2026-09-13): the columns still omitted NEW rows, so the claim
  that they summed to the total was premature; the missing column is now added.
  Two page figures that had no counterpart in the script output are
  corrected: the concentration stat is now "top 10% of genes (23 of 221) carry
  42.7%", and core grounding is 55 rows of which 50 were also ACCEPTed (the page
  previously quoted the ACCEPT∩core number as if it were the total).
• Bounded the sole-carrier claim. 72.4% tests exact term-id equality; with
  is_a/part_of closure — a non-IEP annotation on a parent or child counts as
  corroboration — it is 53.5%. Both are now given, and the checklist tells
  reviewers to look at the lineage rather than the exact term.
• Cached PMID:23842463,
  the GO best-practices paper the page's central quote comes from and previously
  the one citation with no publications/ file. The quote is extended with the
  sentence that states the BP-only rule at source, which is a better primary
  citation for GORULE:0000006 than the rule file alone.
• Verified the ECO:0000279 double parentage against OLS rather than
  asserting it. ECO:0000314 (→IDA) is a direct parent; ECO:0000270 (→IEP) is
  one step further up via ECO:0000284 "protein expression evidence used in
  manual assertion". Both are now linked.
• Moved ARATH/CRY1 and ARATH/CRY2 out of the hub-inversion pattern. A
  photoreceptor annotated response to light stimulus is not inverted, and the
  reviewers' stated reason for MODIFY was term breadth. They now sit in a
  reworked pattern 7 covering
  granularity errors in both directions, which is also where PIF3's seventh row
  and ARATH/SOC1 belong.
• Fixed a miscount in the atlas. The "gene products with IEP" repo cell was
  counting gene directory names, so rat/Hspa8 and mouse/Hspa8 collapsed
  into one. The correct figures are 218 distinct accessions across 220
  directories, not 212, and both are now reported.
• Smaller fixes. Branch classification is now documented as first-match-wins.
  Correction (2026-09-13): its effect on the branch contrast has not been
  measured, so the earlier claim that it biases against the finding is withdrawn.
  % core is documented as
  crediting a code for terms other codes also carry; both scripts normalise the
  one GOA cache that spells the aspect P; the atlas snapshot date is recorded
  in data/global_iep_snapshot.json instead of being read from file mtime,
  which a fresh clone resets (it was showing 2026-08-08 for a 2026-07-27
  download); both outputs record the GO release the closure was computed
  against (2026-07-26), since the coarse branch tallies move by a row or two
  between releases; the atlas now says which of its figures the committed
  snapshot does not freeze (the per-term IEP-dependence shares, whose
  denominator is a live QuickGO count — response to ethanol had drifted from
  13.4% to 15.5%); and both scripts take an IEP_GO_ADAPTER override for
  environments where the default sqlite:obo:go build is unreachable.

2026-08-02 (third pass — the first batch cohort reviewed)

• Retracted the scope claim. The second pass said the RNAcentral slice was
  "out of scope by construction". That was wrong: the repo already reviews ncRNA
  entries (id: URS…, product_type: MIRNA; see human/MIR155, human/XIST) and
  has a dedicated ai-gene-review fetch-ncrna command. The gap was coverage, not
  capability.
• Reviewed five mouse miRNA precursors from the PMID:25858512 LTP batch, chosen
  to span the tiers the paper itself defines: 372 detected, 12 changed, 3
  validated, 130 annotated. Full details in
  the cohort section.
• The tier predicted the verdict in every case. Validated members
  (mouse/Mir26a-1, mouse/Mir384) ACCEPT; changed-but-untested (mouse/Mir30e)
  MARK_AS_OVER_ANNOTATED; detected-only (mouse/Mir100, mouse/Mir127) REMOVE. The
  middle tier is the useful one: mouse/Mir30e's expression change is real and IEP
  is the right code, so the failure is isolated to the inference, not the
  evidence.
• Two findings stronger than the LTP batch itself. mouse/Mir26a-1's
  response to bacterium is refuted by its own source, which shows the
  circulating-miRNA increase survives Tlr2, Tlr4 and NF-kB knockout and
  concludes bacterial sensing is not the driver. And PMID:20439489 supplies
  mouse/Mir100 with both a validated IMP row and an unvalidated IEP batch row —
  the pattern visible within one paper and one gene.
• The LTP batch omits let-7a, one of the three miRNAs the paper establishes
  as required, while including 127 it never tested. Batch annotation is not just
  over-inclusive here; it is misaligned at both ends.
• Cohort flag rate 11/13 (85%) versus 16.7% corpus-wide, and 4 REMOVEs added to a
  corpus that previously held 7. Reviewing batch cohorts is high-yield, and
  reviewing them tier-by-tier is what makes the verdicts defensible rather than
  reflexive.
• Corpus figures updated throughout: reviewed IEP rows 525 → 550 (including 12
  added by other work merged from main), flagged 15.0% → 16.7%, REMOVE 7 → 11.

2026-07-27 (second pass — global denominator)

• Added iep_global_atlas.py and
  iep-global-atlas.md. The first pass measured only the
  repo, whose GOA files exist because genes were picked for review; this pass
  downloads all 25,401 IEP annotations in UniProt-GOA from QuickGO as the
  denominator. IEP is small enough that this is the complete population, not a
  sample.
• The reviewed corpus is 2.1% of global IEP. Crucially, it is
  representative for the stratifications the page's findings rest on —
  stimulus-response branch 1.03x, developmental branch 0.89x, involved_in 0.97x,
  BP aspect 1.00x, gene-concentration 1.02x — and skewed by organism: human
  4.58x, Dictyostelium 6.34x, mouse 0.33x, fly 0.30x, MGI 0.05x, with AgBase,
  ZFIN, CollecTF, cotton, zebrafish, chicken and M. tuberculosis absent entirely.
  Correction (2026-09-13): the representativeness claim above was too broad.
  These are marginal annotation shares, not evidence that the reviewed sample's
  branch-specific flag rates generalize or are free of batch confounding.
• Correction (aspect violations). Querying QuickGO with
  evidenceCodeUsage=exact initially suggested non-BP IEP was ~0.2% of the code
  and that the repo had over-sampled it. Using descendants — which is what a
  GAF's IEP column actually means — gives 1,223 non-BP rows (4.8%), and the
  repo's 4.3% was proportionally accurate all along. The mechanism proposed from
  20 rows scales: ECO:0000279 alone supplies 1,110 of the 1,147 CC violations
  (91%) while being 6% of IEP, whereas the generic ECO:0000270 parent is 99.8%
  BP and essentially rule-compliant. Fixing the ECO→GAF parent choice by aspect
  would clear 91% of the violations with zero curation.
• The screen-batch pattern is the dominant global shape, and the Caco-2 case was
  a small instance. References are sharply bimodal: 62.4% contribute exactly one
  IEP annotation, while the top six each annotate 72–291 gene products to exactly
  one term. The largest is PMID:20439489 (291 mouse gene products →
  cellular response to leukemia inhibitory factor). Three of the six are
  miRNA-profiling studies. PMID:11967071 is the flagship: a paper titled "Over
  1000 genes are involved in the DNA damage response of E. coli" yields
  DNA damage response for 152 genes including maltoporin lamB and maltose
  transporter malF.
• The MGI gap is the consequential one. The biggest batches are MGI's, and MGI
  is the group the repo samples at 0.05x — so the reviewed corpus systematically
  misses the most extreme form of a pattern the page documents.
• New framing: IEP is load-bearing per gene, never per term. Within a gene,
  72% of IEP rows are the sole carrier of their term; measured per term across the
  ontology, no frequent term exceeds 19.6% IEP support
  (seed trichome elongation), and most are under 2%.
• Wrote data/iep_review_candidates.tsv:
  4,807 unreviewed gene products stratified across the global term distribution
  (≤3 per term, seeded), so future review can sample the tail rather than the head.

2026-07-27 (first pass — reviewed corpus)

• Created the project. Built iep_corpus_survey.py,
  which surveys 147,856 cached GOA rows and 210 gene reviews, and writes
  iep-corpus-survey.md. Everything below is measured
  on the repo only; see the second pass for how representative that is.
• Headline finding: IEP's problem is not error rate (15.0% flagged at the
  time, between ISO and IEA) but centrality. At 22.9% ACCEPT and 9.9%
  core-function grounding it was the weakest experimental code by both measures, with 56.8% of
  rows landing in KEEP_AS_NON_CORE — the highest of any code. The
  characteristic IEP annotation is true and peripheral.
• Initial second finding (subsequently qualified): the first pass described
  the developmental branch (23.7% flagged) as riskier than the stimulus-response
  branch (15.1%), despite the latter carrying 70.5% of all IEP rows. That general
  claim was not established; see the corrected comparison and Recommendation 3.
• Third finding: the 23 GORULE:0000006 aspect violations are an ECO→GAF
  mapping artifact, not curator error. ECO:0000279 (qualitative western
  immunoblotting) descends from both ECO:0000270 (→IEP) and ECO:0000314 (→IDA);
  the GAF projection picks IEP, so SynGO's fractionation-based CC annotations
  trip a BP-only rule.
• Identified PMID:21492153 (Caco-2 differentiation proteomics) as a
  single-screen batch source: 8 unrelated human genes, all annotated
  epithelial cell differentiation, 6 of them flagged and the other 2 kept only
  as non-core.
• Recorded the legitimate-use criterion: IEP holds when the gene's job is the
  response (heat-shock and UPR chaperones, infection-inducible effectors, clock
  genes, stage markers that are the differentiated product), and fails when a
  constitutively-functioning protein is merely swept up by the condition.

Slides

• Slides (Marp source: IEP-slides.md) — AI generated