No Data (ND): what an ND annotation is telling us

SCOPING PIPELINE

Species: yeast, SCHPO, DROME, CANAL, worm, DANRE, ARATH, human, mouse

No Data (ND): what an ND annotation is telling us

Bottom line: an ND annotation says that a curator searched the literature
for one GO aspect and found nothing to annotate. That is a statement about
evidence for this gene product, not about knowledge. The two diverge in
predictable ways. When this project started (2026-10-24), 361 ND rows sat in
237 gene reviews, and the same situation received different actions in different
reviews: for an ND
whose aspect also carried an IBA or ISO, reviews chose ACCEPT 36 times,
REMOVE 37 times and KEEP_AS_NON_CORE 13 times. Only 90 of the 237 genes
had any knowledge_gaps, and 62 accepted ND rows (51 genes) had nothing else in
their aspect and still no BIOLOGY gap. Current counts are in
the inventory. This project sorts ND rows into a
small set of cases. One rule fixes the action: an ND that follows the
evidence-code guideline is ACCEPT. It follows the guideline when no
literature on this gene product, in this species, identifies a function in that
aspect. Inferences (IBA, ISO, ISS, IEA) in the same aspect do not change this.
Only one case (a true gap) must also be paired with a KnowledgeGap.

What ND means in GO

Two GO rules define it:

Three consequences follow, and the cases below rest on them:

  1. ND is per aspect. An MF ND next to rich BP annotations is the
    MF_DARK pattern from Function Knowledge Gaps.
    It is not a contradiction.
  2. ND is dated. "As of the annotation date" means an ND can go stale. The
    date is only informative if it records a real search. 108 of the 133 SGD ND
    rows in our corpus carry the same date (2013-08-07). That pattern looks like a
    bulk re-stamp, not 108 separate searches.
  3. ND is about literature for this gene product. It is silent on
    inference. In practice a MOD's ND often coexists with a GO_Central IBA in
    the same aspect, though GORULE:0000054 reports this pairing. An IBA is not
    literature for the gene, so it does not make the ND wrong.

The acceptance rule

We reward annotations that follow the evidence-code guidelines, and ND is no
exception. The test for an ND row is:

Is there literature on this gene product, in this species, that would
support a GO annotation in this aspect under the evidence-code guidelines
(an IDA, IMP, IPI, IGI, IEP or similar that a curator would make)?

"Literature" means papers that identify a function, not any mention of the
gene. A screen hit, an expression change or a non-physiological in-vitro
reaction that a curator would not annotate leaves the ND correct (see N4).
A bare protein-binding IPI does not count either: it identifies no function,
which is why the protein-binding policy removes it. Papers on orthologs do not
count, however strong. Per CLAUDE.md, an accepted ND needs evidence that the literature was searched:
cite the deep-research file or the reviewer's own search, not the UniProt
PE line. The rows converted by the re-action pass below carry no such source
yet, because the pass searched nothing, so they do not earn the
nd_search_support credit; the metric therefore counts the searches still to
be recorded.

ACCEPT here endorses the annotation as made. It does not make a root term a
core function: core_functions never cite an ND row, and the function of an
N2 gene is carried by its informative IBA or ISO rows. KEEP_AS_NON_CORE is
not used for ND, and neither is REMOVE when an IBA appears in the same
aspect.

Case taxonomy

The cases are ordered by how often they occur in our corpus and by how much
they matter. Each case answers one question: if the ND is true, why is there
no data?

Case Short name Is the ND true? Knowledge gap? Proposed action
N1 Stale or missed No: data exist No (curation) REMOVE, and add the informative term
N2 Not assayed here, inferable Yes No ACCEPT
N3 Existence in question Yes Yes, of a special kind (does the product exist?) ACCEPT, defer to MICROPROTEINS
N4 True knowledge gap Yes Yes, BIOLOGY, required ACCEPT and a KnowledgeGap
N5 Ontology-blocked No: literature identifies the function No (ontology) MODIFY to the term, or a proposed term
N6 Tested and negative Yes No (or a narrower gap) ACCEPT, and keep the NOT row
N7 Bookkeeping artifact Unclear Decide after reading Re-check, then reclassify

N1 — Stale or missed: data now exist

The ND was true on its date and is no longer true, or the search missed a paper.
GORULE:0000054 already flags the conflict when the new data have been
annotated. When they have not, the review finds the paper and the gap becomes
a curation gap.

Action: REMOVE. Here REMOVE is used in its normal sense: the claim "no
data exist" is now false. Add the informative term (as NEW if it is missing
from GOA). Do not create a knowledge gap. If GOA still has the ND after the
MOD has annotated the aspect, the ND is a candidate source_defects report.

N2 — Not assayed here, but obvious from orthologs

An enzyme, transporter or complex subunit whose direct orthologs are well
characterised and whose function no one has re-assayed in this species. The ND is
true, because nothing was published on this gene product. The function is
nonetheless known by any reasonable standard, and IBA, ISO or ISS rows usually
carry it already. This is not a knowledge gap: no wet-lab experiment would tell
us anything new, beyond "confirm it in species X".

Examples: CANAL/LPL1 (phospholipase B, inferred from the S. cerevisiae
branch), CANAL/ERG2, CANAL/MON1 and DANRE/gmppb. The inventory signal
inferable_by_homology marks 106 rows, and most of them belong to this case.

Action: ACCEPT. Nothing was published on this gene product in this
species, so the ND follows the guideline. The informative IBA or ISO carries the
function, and is reviewed on its own row. The ND records that the function is
unverified in this species. Do not add a KnowledgeGap. REMOVE is wrong: 37
reviews removed an ND for this reason alone ("superseded by the IBA"), which
penalises a correct annotation.

Subcase N2b — the family is known but the specificity is not. If the
ortholog assignment is ambiguous (paralog-rich family, possible pseudoenzyme,
substrate range that differs between subfamilies), the function is not
obvious. yeast/SIA1 (possibly a pseudophosphatase) is the type example. This
subcase belongs in N4 with dark_aspect: RESIDUAL_SUBGAP or MF_DARK. The
test is the IBA one in projects/IBA_REVIEW.md: is the target inside the clade
that inherited the function?

N3 — Existence in question: microproteins and dubious ORFs

The basic unknown is whether a functional gene product exists at all:
sORF peptides, uORF and alternative-frame products, yeast "dubious" and
"putative" ORFs, and lineage-restricted ORFs without homologs. See
Microproteins and the just fetch-gene --alias rule for
alternative-ORF folders. Example: human/P3R3URF.

A length filter does not identify this case. The inventory's ≤100-aa flag
also catches yeast/SMX2 (an Sm protein) and yeast/COA4, which are canonical,
well-conserved short proteins and fall in N2 or N4. The MICROPROTEINS census
found the same split.

Action: ACCEPT. Where translation or conservation evidence
exists but function does not, the gene gets a BIOLOGY gap whose
gap_statement puts the existence question first.

N4 — True knowledge gap

Nobody knows. The review has read the literature and the deep research and
confirms that no experiment, and no defensible inference, assigns a term in this
aspect. Action: ACCEPT. Only this case should be coupled to
knowledge_gaps. The coupling should be mandatory:

ND aspect Required KnowledgeGap.dark_aspect
MF MF_DARK (or WHOLLY_DARK)
BP BP_DARK (or WHOLLY_DARK)
CC CC_DARK (or WHOLLY_DARK)
all three WHOLLY_DARK

The gap also has gap_kind: [BIOLOGY] and a provenance quote showing that the
unknown is real. Reading the existing reviews turns up recurring sub-flavours.
They are worth recording in gap_statement because each points to a different
experiment:

N5 — Ontology-blocked: the function is known but GO cannot say it

Structural or accessory subunits whose function is "be part of machine X" often
get an MF ND because no MF term fits, even though literature on the gene
identifies its role. Under the acceptance rule that literature makes the ND
incorrect. That accounted for about 64% of the conserved
"MF-dark" genes in the Function Knowledge Gaps triage (worked example:
human/POLE4). No ND row in the corpus is classified N5 yet. DROME/Vps53 looks
like one, but the tethering literature behind it is on human EARP, so under the
rule its ND is N2 and the tethering activity sits on its own inferred row.

Action: MODIFY to the expressible term (a contributes_to complex
activity or a structural constituent term). If none exists, MODIFY to an
NTR and add a matching proposed_new_terms entry. Per CLAUDE.md, an
ONTOLOGY-only knowledge_gaps entry is not allowed.

N6 — Tested and negative

The function people expected was tested and ruled out. Action: ACCEPT:
a negative result identifies no function. The ND then sits next to a NOT
row, as for SCHPO/alo1: the positive ISO to D-arabinono-lactone
biosynthesis was withdrawn and PomBase now carries NOT IGC for that process. The
ND is honest, but the aspect is not empty: the NOT row is information. If an
alternative function is plausible but unknown, record an N4 gap with the
negative result as its boundary.

N7 — Bookkeeping artifact

The ND reflects how the record was handled, not a search. Signs:

These are not a separate biological category. Re-check the ND and reclassify it
into N1–N6, then record the provenance issue in the notes.

Considered and folded in

Current state of the corpus

From the inventory
(uv run python projects/NO_DATA/scripts/nd_inventory.py), regenerated with
each change to this project; on 2026-10-31 it counted 427 ND rows in 268 reviews. A bare protein-binding
IPI is not counted as experimental (it identifies no function).

Provisional signal ND rows Likely case(s)
aspect_empty (nothing else in the aspect) 216 N4, N3, some N1 (missed)
inferable_by_homology (IBA/ISO/ISS present, no experimental) 118 N2, N2b→N4
electronic_or_author_only (incl. bare protein-binding IPIs) 52 N2 or N4, needs reading
short_orf/* (≤100 aa) 34 N3, or canonical short proteins
coexists_with_experimental 7 N1

The signals come from GOA alone and are only for triage. As in Function
Knowledge Gaps, the case is decided by reading the literature, not by a pattern
in the annotations. Actions are currently inconsistent within each signal (see
the signal × action table in the summary). One example is the 4 NEW ND rows
(SCHPO/tam10 ×2, human/P3R3URF, MYCTU/Rv0898c), where reviewers used NEW
ND to assert a gap. ND is the contributing group's statement about its own
search, so we cannot meaningfully propose one. These rows should become
knowledge_gaps entries.

Coupling to knowledge gaps (implemented)

  1. Schema. review.nd_case (NoDataCaseEnum: STALE, INFERABLE,
    EXISTENCE, TRUE_GAP, ONTOLOGY_BLOCKED, NEGATIVE) records why an ND
    row has no data. Bookkeeping artifacts (N7) are re-checked and given one of
    these values. The ACCEPT description now says what ACCEPT means on an ND
    row.
  2. Validation (nd_policy, best-practice warnings):
  3. KEEP_AS_NON_CORE, MARK_AS_OVER_ANNOTATED or TOO_GENERAL on an ND;
  4. REMOVE or MODIFY on an ND without nd_case STALE or
    ONTOLOGY_BLOCKED, which catches removal "because an IBA exists";
  5. ACCEPT with nd_case STALE or ONTOLOGY_BLOCKED, or with
    proposed_replacement_terms;
  6. NEW on an ND row (record a knowledge_gaps entry instead);
  7. nd_case TRUE_GAP or EXISTENCE without a BIOLOGY gap whose
    dark_aspect matches the ND aspect (MF_DARK, BP_DARK or CC_DARK,
    or WHOLLY_DARK), on the row, the gene or a core function;
  8. nd_case on a non-ND row;
  9. ACCEPT without evidence that the literature was searched: the row must
    cite a publication (PMID, DOI, PMC), or quote in supported_by the gene's
    own deep-research or notes file where it records the queries and their
    result (a search that finds nothing has no paper to cite), as DROME/CG42763
    does. The quote goes in supporting_text, the field the verbatim check
    confirms (supporting_text_fulltext is not checked and does not count), so
    a bare file citation, a file in additional_reference_ids, another gene's
    file, or a MOD gene description (-deep-research-sgd/-flybase/-wormbase,
    curator text rather than a search) does not count. A publication id needs
    no quote: a paper checked and found silent on the aspect has no line to
    quote, and the id is checkable on its own. The UniProt record, or its
    protein-existence (PE) line, is not a search. Any verbatim line of the
    gene's own file passes, so the check narrows the loophole rather than
    closing it: the quote should be the line that records the search. Quotes listed in
    conf/local_quote_baseline.yaml only warn when not verbatim, so for those
    few the check does not confirm the file says it. On 2026-11-06 this flags
    210 of 321 accepted ND rows in 150 reviews, mostly rows the re-action
    pass converted.
  10. Compliance (qc_config.yaml, the evidence-code × action rule table):
    nd_search_support credits an ACCEPTed ND that gives a reason and cites
    the literature search the same way the warning requires: a publication, or
    a supporting_text quote from the gene's own deep-research or notes file
    (not a MOD gene description, nor the UniProt record). The pass adds no
    such citation itself: the credit is for a search a reviewer records, so
    the rows the pass converted score 0 until one is.
    nd_stale_support credits a REMOVEd or MODIFYd ND only for a PMID/DOI
    quote from the literature that makes it stale.
  11. Upstream (to do). N1 rows where the MOD has already annotated the aspect
    are GORULE:0000054 conflicts. They can go to source_defects and on to the
    per-resource reports. N7 bulk-date clusters are a question for the MOD,
    not a defect.

Re-action pass

The decisions are in data/nd-decisions.yaml,
one entry per ND row the pass changed, and
scripts/apply_acceptance_rule.py applies them. They were made by reading each
row and the review's other rows in the same aspect (2026-10-24, extended on
2026-10-26 to reviews merged from main, and re-checked on 2026-10-31 after
review), not by a new literature search, and the generated text says so.
The script restores each row as it was before the pass, then edits only that
row's review block, so it can be re-run. The pre-pass rows are read from git,
so re-running needs those commits to be reachable; after a squash merge it can
no longer restore them.

Not yet swept (as of 04e2ffa3e). Reviews merged from main after the last
pass carry 42 ND rows that nd_policy flags: 37 MODIFY rows in 13 mouse
keratin-associated protein reviews (Krtap3-2 to Krtap29-1), whose reasons rest
on family-level literature; human/CDKAL1 BP; ARATH/PGR3; and rat Actl7a, Aph1a
and Pik3c2a. They are waiting for a decision.

Plan

Open questions