Function Knowledge Gaps

What does biology genuinely not know about how a gene works — and how do we state it
rigorously enough that someone could close it?

This project builds a curated, literature-grounded register of function knowledge gaps:
specific, defensible statements of what is unknown about a gene product's molecular function,
mechanism, partners, or biological role. It is the inverse of the rest of this repository.
Everywhere else we adjudicate what is known; here we adjudicate, with the same evidentiary
discipline, what is not.

The framing is deliberately ambitious. If a foundation set out to systematically eliminate the
conserved "unknome," the first deliverable would not be more experiments — it would be an honest
map of where the real gaps are, sharp enough to fund. That map is the goal.

Why this matters: the unknome

Decades into the genomic era, a large fraction of even well-studied proteomes remains
functionally dark:

• ~20% of proteins in well-studied model organisms still lack any informative description of
  their biological role, and many of these are conserved from yeast to human — implying
  fundamental, not peripheral, functions (Wood et al. 2019, Open Biology, "Hidden in plain
  sight", PMID:30938578, DOI).
• The Unknome database, which ranks proteins by an evidence-weighted "knownness" score
  (manual annotation = 0.9, electronic = 0.0; "unknown" = score ≤ 1.0), found 23% of human
  protein clusters still below that threshold — down from 43% a decade earlier, i.e. slow
  progress. RNAi screening of 260 conserved-unknown Drosophila genes found 62 essential for
  viability and 59 with measurable phenotypes — the dark set is dense with real biology
  (Rocha, Freeman et al. 2023, PLOS Biology, PMID:37552676,
  DOI).

Genes stay dark largely for sociological reasons — research concentrates on what is already
famous (the streetlight effect / Matthew effect; Stoeger et al. 2018, PLOS Biology,
PMID:30226837; Edwards et al. 2011, Nature, "Too many roads not taken", PMID:21307913). The
same neglected set is over-represented among unsolved Mendelian disease, unannotated GWAS
peaks, and the druggable-but-unexplored proteome (the NIH IDG program; Oprea et al. 2018,
Nat Rev Drug Discov, PMID:29472638; Kustatscher et al. 2022, Nat Methods, "Understudied
proteins", PMID:35534633). Closing function gaps is the denominator problem under all of these.

Core principle: a knowledge gap is a curated judgment, not a metric

The central methodological commitment of this project: a knowledge gap is determined by
reading the primary literature and exercising judgment, not by a pattern in the annotations.

We tested the alternative and it failed instructively. A score that flags genes with no
specific molecular function (à la Unknome, but using our adjudicated core_functions) marks
~14% of our reviewed genes as "MF-dark" — reassuringly close to Wood's ~20% and Unknome's 23%.
But when the conserved subset of those was inspected, "no molecular function" decomposed into
four completely different things:

What the score called a "gap"	Share	What it actually is	Owner
Structural/accessory subunits (TOM, TRAPP, ESCRT, V-ATPase…)	~64%	Function is "be part of the machine"; the GO MF aspect can't express it	Ontology
Function-kind known, MF term simply not filled (e.g. SOX9)	~14%	Curation incompleteness	Curation
Stub / incomplete reviews (TODO descriptions)	~3%	Data hygiene	—
Genuinely unknown mechanism (CFAP300, tam10, P3R3URF…)	~18%	The true unknome	Experiment

(Snapshot over ~2,100 reviewed genes; proportions are triage-grade, not curated.) A heavily
annotated gene can hide a gaping mechanistic hole; a sparsely annotated one can be perfectly
understood and merely under-curated. Only reading tells them apart. The metric is therefore
demoted to a back-of-house triage aid that produces a read-list, and never appears in the
product.

A taxonomy of function knowledge gaps

Two axes. First, what kind of ignorance (this determines who can fix it):

1. Biology gap — nobody knows. Resolved by experiments. This is the unknome and the
   project's primary target.
2. Curation gap — the knowledge exists in the literature but is not yet annotated, or is
   annotated too generically. Resolved by curation (much of it in-house).
3. Ontology gap — the knowledge exists but no GO term can express it (e.g. "structural
   subunit of complex X", or a novel activity). Resolved by ontology development;
   tracked via proposed_new_terms. Worked exemplar: POLE4 below.

Most real entries are a blend (CFAP300 below is biology-dominant with an ontology shadow),
and naming the blend is the actionable part.

A useful third framing emerged from curation, cutting across the above: the residual sub-gap —
a gene whose core function is textbook-solid but which still hides one sharp, load-bearing
mechanistic hole (e.g. RAB9A's unidentified GEF/GAP, RASA1's catalysis-independent scaffolding,
atg101's WF-finger recruit). These are easy to miss precisely because the gene looks finished;
flagging them is high-value because the hole is often the rate-limiting unknown for the pathway.

Second, which GO aspect is dark — most "dark" genes are not uniformly dark:

• MF-dark — process/location known, molecular mechanism unknown (the most common and most
  insidious: rich BP/CC make the gene look known). The protein binding smell lives here.
• BP-dark — an activity is known but not what it is for (common in microbial/plant
  metabolism).
• CC-dark — function known, but where/when unknown.
• Wholly dark — only root terms / IEA / protein binding survive review (the deep unknome).

The unit of work: anatomy of a gap entry

Each gap is a small, defensible scholarly object. Required elements:

• Gap statement — the specific unknown, stated precisely. Not "role unclear" but
  "the direct substrate / the catalytic activity / the essential partner is undetermined."
• Boundary of knowledge — what is established, so the gap is sharply delimited.
• Provenance for the gap itself — evidence that the unknown is real and not merely
  uncurated. The gold standard is the field's own admissions: primary-literature hedges
  ("remains to be determined", "the precise role is unknown"), "domain of unknown function"
  names, and a literature trajectory that confirms-but-never-mechanizes. Cited with the same
  [PMID:xxxx "supporting text"] discipline this repo uses for positive claims.
• Type judgment — biology / curation / ontology (or blend), per the taxonomy.
• Significance — why closing it matters.
• What would resolve it — the experiment, ontology term, or curation action.

This is strictly more than the existing suggested_questions field (which asks the right things
but cites nothing): the added value is the adjudicated boundary plus provenance for the
unknown.

Structured curation: the KnowledgeGap schema class

The anatomy above is now a first-class schema object, KnowledgeGap, so a gap can be curated
in the gene/module YAML itself rather than only as prose on this page. This mirrors the way
the Monarch dismech knowledge base promotes
unknowns to a first-class Discussion object (with kind / status / proposed_experiment),
and goes further by demanding provenance for the unknown and a biology/curation/ontology
typing.

A KnowledgeGap (see src/ai_gene_review/schema/gene_review.yaml) carries:

• gap_statement (required) and boundary — the precise unknown and the edge of knowledge;
• gap_kind (BIOLOGY / CURATION / ONTOLOGY, multivalued for blends), dark_aspect
  (MF_DARK / BP_DARK / CC_DARK / WHOLLY_DARK / RESIDUAL_SUBGAP), and a dismech-style
  lifecycle status (OPEN / NARROWING / CLOSING / RESOLVED);
• significance and resolution;
• provenance — a list of SupportingTextInReference, so each "the field's own admission of
  ignorance" quote is a verbatim substring checked by the reference validator, exactly like
  supported_by. (When the only source is a DOI-only paper or a local analysis, anchor to a
  file: reference, as elsewhere in this repo.) Caveat: only PMID/PMC-style references are
  verbatim-checked; file: (and the other skip_prefixes in
  conf/reference_validator_config.yaml) are not quote-checked, so a file:-anchored quote
  such as RAB9A's must be verified by hand. Prefer a checkable PMID quote where one exists.
• proposed_terms — inline ProposedOntologyTerms for ONTOLOGY gaps.

It can be attached at five levels — the whole gene (GeneReview.knowledge_gaps), a single
annotation (ExistingAnnotation.review.knowledge_gaps, ideal for residual sub-gaps), a
core function (CoreFunction.knowledge_gaps), a whole module (ModuleReview.knowledge_gaps),
or a single module step (ModuleNode.knowledge_gaps).

Worked YAML exemplars in the KB: genes/human/CFAP300/ (gene-level biology+ontology gap,
two PMID provenance quotes) and genes/human/RAB9A/ (annotation-level residual sub-gap on the
GTPase-activity review, the missing GEF/GAP).

Rendered from data: running just aggregate-knowledge-gaps (or
python scripts/aggregate_knowledge_gaps.py) walks every gene/module YAML and regenerates the
Structured Knowledge-Gap Register plus
reports/knowledge_gaps.tsv. The prose worked-entries below remain the curated narrative; the
register is their queryable, schema-backed counterpart and will grow as gaps are recorded in the
YAML.

Worked example: the CFAP300 molecular-function gap

CFAP300 (formerly C11orf70) is a dynein axonemal assembly factor; loss-of-function causes
primary ciliary dyskinesia (PCD). It is in the KB (genes/human/CFAP300/) and is an ideal
exemplar: clear biomedical importance, a clearly established role, and a sharp, durable
mechanistic silence.

Boundary (what is firmly known):
- LOF causes PCD with combined outer + inner dynein arm (ODA+IDA) loss; CFAP300 is required
for cytoplasmic preassembly of axonemal dyneins and their IFT-dependent delivery into the
axoneme (Höben et al. 2018, AJHG, PMID:29727693, DOI;
Fassad et al. 2018, AJHG, PMID:29727692, DOI).
- Localizes mainly to cytoplasm, moves into cilia by IFT, and interacts with the preassembly
factor DNAAF2 (Höben 2018).
- Re-confirmed in 2025: LOF → no CFAP300 protein → total ODA+IDA loss (Demchenko et al. 2025,
IJMS, PMID:40806783, DOI).

Gap statement: The biochemical activity of CFAP300 is unknown. It is unresolved whether
it acts as a chaperone/co-chaperone, a scaffold, or an adaptor; what its client/substrate is;
and whether it acts at one common step upstream of both arm types or in parallel ODA- and
IDA-specific steps.

Provenance that the gap is real (the judgment):
1. Both founding papers describe the gene as "uncharacterized" and frame the mechanism as a
hypothesis ("supporting our hypothesis that C11orf70 is a preassembly factor") — they
establish requirement and localization, never a biochemical activity (PMID:29727692,
PMID:29727693).
2. Its sole domain is DUF4498 — a "domain of unknown function"
(genes/human/CFAP300/CFAP300-deep-research-falcon.md).
3. The trajectory is the strongest evidence: seven years and many cohorts on (Slavic 2019
PMID:30916986; Cypriot 2021 PMID:33715250; Russian 2024 PMID:39180133; ALI-culture 2025
PMID:40806783), every follow-up is diagnostic — confirming loss — not mechanistic. A
durable mechanistic silence despite clear motivation is the signature of a genuine gap, and
reading the arc of the literature is the only way to see it.

Type judgment: biology gap (dominant) with an ontology shadow — even "a dynein-preassembly
factor" has no adequate GO MF term, which is why the gene reads as MF-dark.

Significance: this is the assembly step whose failure removes all axonemal dynein motors —
mechanistically central to motile ciliopathy.

What would resolve it: proximity/affinity proteomics of the CFAP300–DNAAF2 module;
in vitro reconstitution of dynein assembly intermediates; structural characterization of
DUF4498.

Worked gap entries (from the seed read-list)

The following entries were curated by reading each gene's -deep-research-*.md and notes/GOA
files for the field's own statements of ignorance, in the CFAP300 format but condensed. Every
PMID cited below was independently verified against PubMed; ignorance quotes are reproduced
verbatim and attributed to the source file. Where a gene's deep-research file cited sources only
by DOI, or by PMIDs that failed verification, the gap is provenance-anchored to the file path
rather than to an unverified PMID (see the KCTD14 caution).

swrD (BACSU) — ~71-aa swarming-motility factor

• Boundary: Required for swarming on soft agar; increases per-flagellum motor torque/power at
  the level of MotAB stator activity (not stator abundance); deletion cuts single-flagellum
  torque ~6-fold and is rescued by motAB overexpression (Hall et al. 2018, J Bacteriol,
  PMID:29061663, DOI). GOA carries only one BP term
  (GO:0071978, swarming motility) — no MF, no CC.
• Gap statement: SwrD's direct binding partner(s), biochemical activity, membrane/motor
  localization, and structural domain are undefined.
• Provenance (verbatim): "they proposed that SwrD may reduce stator dynamism by facilitating
  stator retention at the motor, but were unable to directly visualize stator dynamics due to
  nonfunctional fluorescent fusions—thus mechanistic details remain hypothetical"; "Experimental
  localization of SwrD ... remains undetermined"; "no resolved structural domain showing
  stomatin/SPFH" (genes/BACSU/swrD/swrD-deep-research-falcon.md).
• Type: biology gap (primary) with an ontology/curation shadow — genetics support an MF in
  stator/torque modulation, but no MF term is assignable.
• Resolve: functional tagged-SwrD co-IP + super-resolution to test direct MotAB binding/
  retention; in vitro reconstitution or cryo-EM of the SwrD–stator interface.

mxaC (METEA) — methanol-oxidation auxiliary VWA protein

• Boundary: In the mxa methanol-oxidation cluster and genetically required for methanol
  oxidation — the founding paper states mxaC is required while "the function of the other two
  genes is still unknown" (Morris et al. 1995, J Bacteriol, PMID:7592474,
  DOI). Encodes a von Willebrand factor A
  (VWA) domain protein; recent work implicates it in Ca²⁺ loading during methanol-dehydrogenase
  (MDH) maturation alongside a MoxR AAA+ ATPase module (Zhou et al. 2025, Nat Commun — cited by
  DOI in the file; no PMID present).
• Gap statement: Whether MxaC's VWA/MIDAS domain itself binds Ca²⁺, and how it delivers Ca²⁺
  into the MxaF active site (adaptor vs scaffold vs transporter), is undemonstrated.
• Provenance (verbatim): "Direct physical interaction data (e.g., co-purification of MxaC with
  MxaR) are not shown in the retrieved excerpts."; "There is no direct localization assay for
  MxaC in the retrieved materials..." (genes/METEA/mxaC/mxaC-deep-research-falcon.md); "Does
  MxaC directly bind Ca²⁺ through its vWFA domain?" (genes/METEA/mxaC/mxaC-notes.md).
• Type: biology gap (primary) + ontology gap (no GO term for "metal incorporation into
  metalloenzyme" / MDH-complex assembly).
• Resolve: MIDAS-mutant Ca²⁺-binding assays on purified MxaC; cryo-EM / co-purification of the
  MxaR/S/C/L module with MDH maturation intermediates.

TRAPPC12 (human) — moonlighting TRAPP subunit

• Boundary: Bona fide TRAPP subunit acting at an early stage of ER-to-Golgi trafficking
  (Scrivens et al. 2011, Mol Biol Cell, PMID:21525244,
  DOI); member of TRAPPIII/autophagy machinery (Kim et
  al. 2016 review, PMID:27066478, DOI). Separately, a
  mitotic "moonlighting" role: depletion gives noncongressed chromosomes and mitotic arrest, it
  regulates CENP-E recruitment to kinetochores, and is phospho-cycled through mitosis (Milev et al.
  2015, J Cell Biol, PMID:25918224, DOI).
• Gap statement: The molecular mechanism of the mitotic kinetochore function — and how it is
  biochemically partitioned from the TRAPP trafficking role — is undefined; the CENP-E interaction
  is captured only as uninformative protein binding.
• Provenance (verbatim): "TRAMM cycles between its role in TRAPP in interphase cells, and its
  newly identified roles during mitosis"; "Small amounts of TRAMM associated with chromosomes"
  (genes/human/TRAPPC12/TRAPPC12-notes.md, re PMID:25918224).
• Type: blend — biology gap (mechanism of the second, non-TRAPP pool) + curation/ontology gap
  (CENP-E recruitment under-captured as protein binding).
• Resolve: separation-of-function/domain-mapping mutants that uncouple kinetochore binding from
  TRAPP assembly; mitotic-phase proximity proteomics of the non-TRAPP TRAMM pool.

AGR3 (human) — noncanonical PDI-family protein

• Boundary: ER-resident thioredoxin/PDI-family protein restricted to ciliated airway
  epithelium; required for Ca²⁺-mediated regulation of ciliary beat frequency and mucociliary
  clearance but not for ciliogenesis (Bonser et al. 2015, Am J Respir Cell Mol Biol,
  PMID:25751668, DOI); binds α-dystroglycan and C4.4a
  by yeast two-hybrid in a cancer context (Fletcher et al. 2003, Br J Cancer, PMID:12592373,
  DOI). Has a noncanonical thioredoxin motif (DCYQS,
  lacking the second catalytic cysteine of CXXC PDIs).
• Gap statement: Whether AGR3's noncanonical motif retains any thiol-disulfide catalytic
  activity, what its physiological substrates/folding clients are, and the receptor that mediates
  extracellular-AGR3 signaling, are all unknown.
• Provenance (verbatim): "no definitive AGR3-specific enzymatic substrate, direct folding
  client, or biochemical turnover measurement is established"; "much of AGR3 biochemistry remains
  inferential rather than directly measured" (genes/human/AGR3/AGR3-deep-research-falcon.md).
• Type: biology gap (primary) + curation gap (GOA carries only generic protein binding plus a
  single experimental dystroglycan binding; no MF for the core ciliary/Ca²⁺ role).
• Resolve: thiol-disulfide / reductase assays on WT vs catalytic-Cys-mutant AGR3 plus
  proximity labeling for ER clients; identify the surface receptor for extracellular AGR3.

tam10 (SCHPO) — meiotic sequence orphan

• Boundary: Bona fide protein-coding gene (RT-PCR / RNA-seq / RACE confirmed), named for
  "transcripts altered in meiosis"; deletion is viable / non-essential under standard conditions;
  the only direct functional GO terms (RNA binding, nucleolus) are ISO/ISS transfers, never
  validated in S. pombe (genes/SCHPO/tam10/tam10-deep-research.md; tam10-goa.tsv).
• Gap statement: No experimentally determined molecular activity, RNA/protein substrate,
  partner, or in vivo role exists; the RNA-binding/nucleolus annotations are unverified
  computational transfers.
• Provenance (verbatim): "no specific molecular function has been ascribed to Tam10";
  "classifying it as a 'sequence orphan' with no obvious homology to characterized proteins"
  (tam10-deep-research.md); "No evidence-supported enzymatic activity, substrate specificity,
  transport substrate, or pathway membership can be asserted." (tam10-deep-research-falcon.md).
  (The two providers even disagree on whether it has any annotatable domain.)
• Type: blend — biology gap (primary) + curation gap (unverified ISO/ISS; provider
  disagreement on domain content).
• Resolve: GFP localization + AP-MS / CLIP to test the predicted nucleolar/RNA-binding role;
  meiosis-specific phenotyping of tam10Δ (sporulation, spore viability, meiotic timing).

P3R3URF (human) — uORF microprotein (non-canonical-ORF dark matter)

• Boundary: Protein-coding locus encoding a ~95-aa microprotein from an ORF upstream of
  PIK3R3; UniProt evidence at protein level (PE1); testis / late-spermatid enriched; the sole
  GOA term is IBA (GO:0019221, cytokine-mediated signaling — phylogenetic inference, not
  experiment) (genes/human/P3R3URF/P3R3URF-deep-research-openai.md; P3R3URF-goa.tsv).
• Gap statement: Whether the microprotein is stably expressed as an independent protein, where
  it localizes, and what its molecular function/partner is — every current functional claim is
  computational inference.
• Provenance (verbatim): "direct, P3R3URF-specific functional evidence is very limited in the
  accessible corpus"; "Primary function ... Unknown in the accessible corpus."
  (P3R3URF-deep-research-falcon.md); "While direct experimental evidence is still lacking..."
  (P3R3URF-deep-research-openai.md). In the one dataset that tested for it, the transcript was
  not detected.
• Type: blend — biology gap (primary) + curation gap (IBA-only term risks conflation with
  canonical PIK3R3 / p55γ).
• Resolve: mass-spec detection of a P3R3URF-unique peptide + Ribo-seq initiation evidence in
  testis; ORF-specific perturbation (sparing the PIK3R3 CDS) with interactome/localization readout.
• This is the deliberately-included "we did not even know there was a player" frontier.

KCTD14 (human) — least-studied KCTD-family BTB protein

• Boundary (PubMed-verified): member of the potassium-channel-tetramerization-domain (KCTD)
  family, which "consists of 26 members with mostly unknown functions"; many KCTDs act as BTB-domain
  CUL3 substrate adaptors, but most members remain functionally uncharacterized (Liu et al. 2013,
  Cell Biosci, PMID:24268103, DOI).
• Gap statement: KCTD14's biochemical activity, bona fide substrate/partner, and biological
  process are unknown — including whether it functions as a CUL3 substrate adaptor like several
  paralogs, or via another mechanism.
• Provenance the gap is real (verifiable facts, not summary prose): the gene's GOA record
  (genes/human/KCTD14/KCTD14-goa.tsv) contains no experimental function annotation — its
  entire molecular-function content is five protein binding (GO:0005515, IPI) rows plus one
  electronic protein-homooligomerization term (GO:0051260, IEA), with no IDA/IMP/IGI evidence for
  any specific activity, substrate, or process. Combined with the family review's explicit
  "mostly unknown functions" (PMID:24268103), this establishes an unfilled molecular-function gap
  rather than mere under-curation.
• Type: biology gap (primary; no experiment defines its activity) + curation gap (the
  high-throughput interactors behind the protein binding rows are uninformative and no specific MF
  is captured).
• Resolve: endogenous AP-MS to define the partner/substrate and a CUL3 co-IP to test the
  adaptor hypothesis directly; knockout/knockdown phenotyping with proteomics to assign a process.
• Note: an earlier draft leaned on the -deep-research-cyberian.md summary, which cited PMIDs
  that fail verification (its "PMID:30929316" is an SGLT paper; "PMID:36362138" is a xylanase
  paper). Those claims (including a reported CUL3-non-binding result attributed to an
  unlocatable 2024 paper) are excluded here; the entry now rests only on the verified review and
  the repo's own GOA data.

C18orf21 (human) — ORF-named gene, a closing gap

• Boundary: 220-aa UPF0711-family protein with a single DUF4674 domain and no enzymatic motifs;
  ubiquitously expressed, nucleolar-enriched; a recurrent CRISPR fitness-screen hit. 2025 preprints
  propose it is a metazoan RNase MRP-specific subunit (alias RMRPP1) required for pre-rRNA
  processing (genes/human/C18orf21/C18orf21-deep-research.md;
  C18orf21-deep-research-falcon.md).
• Gap statement: Whether C18orf21 is genuinely a constitutive RNase MRP RNP subunit (and its
  mechanistic role in RMRP stabilization / pre-rRNA cleavage) versus an uncharacterized DUF4674
  protein is not yet established in peer-reviewed, GO-curated form.
• Provenance (verbatim): "evolutionarily conserved protein of 220 amino acids with no
  well-characterized biochemical function"; "current databases report no defined Gene Ontology
  molecular function for this gene"; "To date, there are no published studies focusing
  exclusively on C18orf21's function." (C18orf21-deep-research.md).
• Type: curation / recency gap (leaning) — function-defining evidence currently exists only as
  2025 bioRxiv preprints (no PMIDs) and has not propagated to GOA (empty) or UniProt (still
  "UPF0711").
• Resolve: peer-reviewed validation of RNase MRP-specific membership (reciprocal AP-MS vs
  RPP21; RIP-seq showing RMRP, not RPPH1, enrichment) plus a pre-rRNA-processing readout.
• A clean illustration of a closing gap — and of why preprint-only evidence must not be
  auto-annotated.

Worked gap entries — second batch (read-list deepening)

Curating the un-vetted leads surfaced a useful third category alongside the wholly-dark genes:
the residual sub-gap — a gene whose core function is textbook-solid, but which still hides a
sharp, specific mechanistic hole (RAB9A's missing GEF/GAP, RASA1's catalysis-independent
scaffolding, atg101's WF-finger recruit). These matter because a heavily annotated gene can look
finished while a load-bearing mechanism is undetermined — exactly the failure mode the project's
core principle warns about. All eight leads adjudicated to real gaps (none were spurious);
every PMID below was PubMed-verified, and one mis-attributed citation was caught and dropped.

MTC7 (yeast / S. cerevisiae) — telomere-capping sequence orphan

• Boundary: small ~139-aa basic protein; mtc7Δ clusters genetically with short-telomere /
  telomere-maintenance deletions and is synthetically sick with cdc13-1 in a genome-wide screen
  (Addinall et al. 2008, Genetics — cited in the files only by DOI/PMC, no verified PMID;
  DOI). GOA marks every functional aspect ND
  (no data) plus one IEA membrane keyword.
• Gap statement: Mtc7's biochemical activity, substrate/partner, and the mechanism by which it
  influences telomere capping/length are entirely unknown.
• Provenance (verbatim): "MTC7 (YEL033W) encodes a protein of unknown molecular function. No
  enzymatic activity or specific biochemical function has been demonstrated to date"
  (genes/yeast/MTC7/MTC7-deep-research.md); "No study in the retrieved corpus provides a direct
  molecular function for Mtc7 ... MTC7 remains functionally unannotated mechanistically"
  (genes/yeast/MTC7/MTC7-deep-research-falcon.md).
• Type: biology gap (primary) + curation gap (a telomere-maintenance BP is arguably capturable
  from the genetic evidence as IGI, yet GOA still carries only ND).
• Resolve: AP-MS / Y2H for physical partners; GFP localization + telomere-length / TPE assays
  in mtc7Δ.

RAB9A (human) — known Rab, unknown switch (residual sub-gap)

• Boundary: endosome-to-TGN retrograde Rab GTPase that recycles mannose-6-phosphate receptors
  (Lombardi et al. 1993, EMBO J, PMID:8440258,
  DOI); has well-defined effectors including
  the p40 effector (Díaz et al. 1997, J Cell Biol, PMID:9230071,
  DOI) and TIP47/GCC185 (deep-research files).
• Gap statement: The specific GEF that activates RAB9A on late endosomes and the GAP that
  inactivates it have not been definitively identified.
• Provenance (verbatim): "the specific guanine nucleotide exchange factor (GEF) that activates
  RAB9A and the GTPase-activating protein (GAP) that inactivates it have not been definitively
  identified" (genes/human/RAB9A/RAB9A-deep-research-cyberian.md); "Regulators specific to
  RAB9A (cognate GEFs/GAPs) ... remain less well defined"
  (genes/human/RAB9A/RAB9A-deep-research-falcon.md). (A claim in the openai file that DENND2 is
  the GEF rests on a general DENN-domain paper covering other Rabs, not RAB9A — an
  over-extrapolation that reinforces that no validated RAB9A GEF exists.)
• Type: biology gap — the regulators are genuinely undiscovered.
• Resolve: in vitro GEF assays across candidate DENN-domain GEFs; a TBC-domain GAP screen with
  a CI-MPR mis-sorting readout on knockdown.

RASA1 (human) — catalysis solved, scaffolding unsolved (residual sub-gap)

• Boundary: p120 RasGAP; accelerates Ras GTP hydrolysis ~10⁵-fold via the arginine finger
  Arg789, structurally defined with the transition-state (AlF) mimic (Scheffzek et al. 1997,
  Science, PMID:9219684, DOI); multidomain
  (SH2-SH3-SH2, PH, C2, GAP) with tandem-SH2 phosphotyrosine engagement
  (genes/human/RASA1/RASA1-deep-research-falcon.md).
• Gap statement: The molecular mechanism of RASA1's GAP-activity-independent (scaffolding)
  functions — e.g. how p190RhoGAP recruitment drives directed cell movement and contributes to
  blood-vessel formation independently of its own Ras-GAP activity — is unresolved.
• Provenance (verbatim): "Experimental evidence indicates RASA1 is necessary for directed cell
  movement in vitro, and this role depends on its ability to recruit p190^RhoGAP (independent of
  RASA1's own Ras-GAP activity)"; "the embryonic blood vessel defects in RASA1-null embryos are
  partly due to Ras-independent actions of RASA1" (RASA1-deep-research-openai.md); "How these
  two functions are coordinated, and whether they can be separated therapeutically, warrants
  further study." (RASA1-deep-research-cyberian.md).
• Type: blend — biology gap (the scaffolding mechanism is unresolved) + curation gap (GO
  captures the catalytic GAP branch; the p190RhoGAP-recruitment role in migration is
  under-annotated).
• Resolve: GAP-dead (Arg789) vs scaffold-dead (SH2/SH3) separation-of-function knock-ins
  scoring migration / vascular tube formation; BioID of GAP-dead RASA1 to map the
  catalysis-independent interactome.

BAIAP2L2 (human) — Pinkbar, dark in its native tissue

• Boundary: epithelial I-BAR/IMD protein ("Pinkbar") that binds phosphoinositide membranes and
  generates planar membrane sheets, localizing to Rab13 vesicles and intercellular junctions in
  intestine/kidney (Pykäläinen et al. 2011, Nat Struct Mol Biol, PMID:21743456,
  DOI). In cochlear hair cells it is a row-2 stereocilia-tip
  component (deep-research files).
• Gap statement: The molecular function of BAIAP2L2/Pinkbar in its name-defining native
  intestinal/renal epithelium — what membrane/junctional structure it builds, and through which
  partner — is unknown, because knockout mice show no overt epithelial phenotype.
• Provenance (verbatim): "mice lacking BAIAP2L2 display normal kidney and colon tissue
  morphology and maintain normal electrolyte homeostasis and tissue architecture under
  physiological conditions"; "BAIAP2L2's relationship to microvillar formation and maintenance in
  intestinal brush borders remains incompletely understood."
  (genes/human/BAIAP2L2/BAIAP2L2-deep-research-perplexity.md).
• Type: biology gap (primary; the epithelial MF was never measured) + curation gap (the
  epithelial GO terms are IEA/ISS/IBA inferences).
• Resolve: challenge-condition / conditional-KO phenotyping of intestinal & renal epithelium
  (barrier integrity, brush-border architecture under stress); Pinkbar proximity-labeling
  interactome in polarized enterocytes.

SCGB1C1 (human) — orphan secretoglobin

• Boundary: small secreted secretoglobin-fold protein localized to Bowman's glands of the
  olfactory mucosa, with a hydrophobic cavity capable of binding small ligands; in a mouse
  OVA-asthma model, recombinant SCGB1C1 suppressed Th2 inflammation and expanded Tregs (Kim et al.
  2024, Int J Mol Sci, PMID:38892470, DOI). GOA carries
  only a single IEA extracellular-region term.
• Gap statement: The endogenous physiological ligand(s) SCGB1C1 binds in vivo, and the
  cell-surface receptor / signaling mechanism behind its immunomodulatory (Treg-expanding) effect,
  are unidentified.
• Provenance (verbatim): "The specific hydrophobic ligands that SCGB1C1 binds in vivo remain
  incompletely characterized."; "The exact receptors and signaling cascades through which SCGB1C1
  mediates these immunomodulatory effects remain to be definitively identified, representing an
  important area for future investigation." (SCGB1C1-deep-research-perplexity.md).
• Type: biology gap (ligand + receptor genuinely undiscovered) + curation gap (only one IEA CC
  term; the established secretoglobin fold and the mouse phenotype are uncaptured).
• Resolve: a biochemical ligand-binding screen (lipidomic / odorant affinity) on recombinant
  human SCGB1C1; receptor identification by pulldown/proximity-labeling on Tregs + LOF validation.

FGFRL1 (human) — a receptor that signals without a kinase

• Boundary: atypical FGFR with three Ig-like ectodomains but no tyrosine-kinase domain;
  binds FGF ligands and heparin and acts as a decoy receptor (Trueb et al. 2003, J Biol Chem,
  PMID:12813049, DOI); forms constitutive dimers and
  mediates HSPG-dependent cell adhesion (Rieckmann et al. 2007, Exp Cell Res, PMID:18061161,
  DOI); essential for kidney, diaphragm and skull
  development, with the cytoplasmic tail dispensable (deep-research files).
• Gap statement: The precise molecular mechanism by which kinase-dead FGFRL1 modulates
  FGF/FGFR signaling in vivo (pure ligand sink vs inhibitory FGFR complexes vs intracellular
  Sprouty/Spred recruitment), and the identity of its Ig3 cell-fusion partner, are unknown.
• Provenance (verbatim): "manipulating FGFRL1 levels in vitro did not measurably change cell
  proliferation or ERK phosphorylation"; "How exactly does FGFRL1 regulate FGF signaling in vivo?
  Is it purely by sequestering FGFs (acting as a sink), or does it form inhibitory complexes with
  the signaling receptors?"; "The 'target protein' involved in FGFRL1-mediated cell fusion is
  currently unknown." (genes/human/FGFRL1/FGFRL1-deep-research.md).
• Type: biology gap (primary; transduction mechanism undetermined) + curation gap (the GOA term
  GO:0005007 "fibroblast growth factor receptor activity" overstates canonical signaling for a
  kinase-dead receptor — a candidate MODIFY).
• Resolve: FGFRL1–FGFR1/2/3/4 co-IP + live-cell co-imaging during ligand stimulation to test
  sink-vs-complex models; a fusion-defective Ig3 point-mutant knock-in plus a screen for the Ig3
  partner.

atg101 (SCHPO) — known subunit, unknown recruit (residual sub-gap)

• Boundary: core subunit of the S. pombe Atg1/ULK autophagy-initiation complex (Yu et al.
  2021, J Cell Sci, PMID:34499173, DOI); forms an obligate
  HORMA heterodimer that stabilizes the Atg13 HORMA domain, and its conserved "WF finger" recruits
  downstream factors (Suzuki et al. 2015, Nat Struct Mol Biol, PMID:26030876,
  DOI).
• Gap statement: The molecular identity of the downstream factor(s) recruited by the Atg101
  WF-finger surface in S. pombe is unknown, despite that surface being genetically required for
  autophagy independent of Atg13 binding.
• Provenance (verbatim): "A WF-finger triple mutant (W110A, P111A, F112A) retained Atg13
  binding but impaired autophagy, indicating Atg101 has functional roles beyond stabilizing/binding
  Atg13." (atg101-deep-research-falcon.md; the underlying WF-finger result is from a thesis, no
  PMID); "The precise protein targets of the WF finger motif remain incompletely characterized,
  though WIPI family proteins represent likely candidates." (atg101-deep-research-perplexity.md).
• Type: biology gap (the binding partner is unidentified) + minor curation gap (no MF term for
  the WF-finger recruitment; the key evidence sits in an uncited thesis, not a curatable PMID).
• Resolve: IP-MS comparing WT vs WF-finger AAA mutant in starved S. pombe; targeted binding
  tests against candidate WIPI / Atg18 orthologs.

irg-1 (worm) — famous reporter, mystery protein

• Boundary: intestinal infection-response gene transcriptionally induced by Pseudomonas
  aeruginosa via the bZIP factor ZIP-2 (Estes et al. 2010, PNAS, PMID:20133860,
  DOI); annotated to antibacterial innate immune
  response by IEP (expression) evidence; the protein carries predicted NADAR/YbiA-like domains (a
  domain prediction only).
• Gap statement: The actual biochemical/enzymatic activity of the IRG-1 protein — whether its
  predicted NADAR/YbiA-like domain confers a real catalytic function, and on what substrate — is
  undefined; everything known concerns its transcriptional induction, not what the protein
  does. GOA records molecular_function as ND.
• Provenance (verbatim): "irg-1 enables GO:0003674 molecular_function ... ND"
  (genes/worm/irg-1/irg-1-goa.tsv); "Despite domain predictions (NADAR/YbiA-like), no direct
  enzymatic reaction, substrate specificity, or transport function has been experimentally defined
  for the IRG-1 protein in C. elegans." (genes/worm/irg-1/irg-1-deep-research-falcon.md).
• Type: biology gap (MF genuinely unknown), honestly reflected as an MF ND annotation — not a
  curation/ontology defect. A clean BP-known / MF-dark exemplar.
• Resolve: assay recombinant IRG-1 for NADAR-family activity (NAD / ADP-ribose-related
  hydrolase) against candidate substrates; structure-guided catalytic-residue mutagenesis + rescue.

Worked gap entries — third batch (cluster + subunit cases)

This batch closes out the taxonomy: it adds the project's first worked ontology-dominant entry
(POLE4 — the structural-subunit pattern that the core-principle table estimates at ~64% of
apparent MF-darkness), a consolidated cluster gap (the M. extorquens lanthanophore accessory
genes), and three more residual sub-gaps. With these, all three gap kinds (biology, curation,
ontology) and the residual-sub-gap framing have curated exemplars.

M. extorquens methanol/lanthanide cluster — uncharacterized accessory genes (consolidated)

• Scope & boundary: the mxa (Ca²⁺-dependent MDH), xox (lanthanide-dependent MDH), and
  mll (lanthanophore biosynthesis) clusters drive methanol oxidation in METEA. Several accessory
  components are now firmly defined and are not gaps: XoxG (cytochrome c_L) and XoxJ are
  biochemically/structurally characterized (Featherston et al. 2019, ChemBioChem, PMID:31017712,
  DOI); the lanthanide-switch regulatory cascade
  MxcQE→MxbDM→mxa/xox is mapped (Skovran et al. 2011, J Bacteriol, PMID:21873495,
  DOI); and the lanthanophore methylolanthanin plus its
  biosynthetic machinery were identified in 2024 (Zytnick et al. 2024, PNAS, PMID:39078674,
  DOI).
• Shared gap statement: for the mll accessory genes (mllG, mllH, mllJ and the NIS-synthetase
  components mllA/mllF), no enzyme of the methylolanthanin pathway has been biochemically
  reconstituted — the specific reaction, substrate, and direct molecular role of each protein are
  inferred from homology and cluster context, not demonstrated.
• Per-gene provenance (verbatim): mllG — "No enzymatic activities have been directly
  demonstrated for mllG through standard biochemical assays in the current literature."
  (genes/METEA/mllG/mllG-deep-research-perplexity.md); mllA — "No direct biochemical studies
  have been conducted on the mllA enzyme, and its precise substrate specificity, catalytic rate
  constants, and reaction mechanism await experimental determination..."
  (genes/METEA/mllA/mllA-deep-research-perplexity.md); mllH — "Direct enzymatic characterization
  of recombinant META1p4137 protein has not yet been reported in the peer-reviewed literature..."
  (genes/METEA/mllH/mllH-deep-research-perplexity.md); mllJ — "The primary molecular function of
  MexAM1_META1p4138 (mllJ) remains partially characterized due to its recent discovery and the
  presence of a domain of unknown function classification."
  (genes/METEA/mllJ/mllJ-deep-research-perplexity.md).
• Type: biology gap (dominant — no reconstituted enzymology for the mll accessory proteins) +
  ontology shadow (GO lacks precise terms for lanthanide / metallophore handling).
• Resolve: in vitro reconstitution of the methylolanthanin pathway with purified MllA/F/H plus
  carriers (as done for aerobactin/petrobactin); single-gene (not whole-cluster) deletions of
  mllG/H/J with lanthanide-bioaccumulation phenotyping and localization to assign individual roles.

MAP7D1 (human) — paralog-specific mechanism unknown (residual sub-gap)

• Boundary: MAP7-family microtubule-associated protein; binds MTs via its N-terminal half and
  recruits/activates kinesin-1 via the MAP7 domain; all four paralogs bind kinesin-1 and act
  redundantly in HeLa (Hooikaas et al. 2019, J Cell Biol, PMID:30770434,
  DOI); MAP7D1 is specifically required to maintain
  acetylated/stable microtubules, mechanistically distinct from MAP7D2 (Kikuchi et al. 2022, Life
  Sci Alliance, PMID:35470240, DOI).
• Gap statement: The precise molecular mechanism by which MAP7D1 specifically maintains
  acetylated/stable microtubules (vs its paralogs), and how it scaffolds nuclear DNA-damage-response
  factors from the cytoplasm, are unknown.
• Provenance (verbatim): "How exactly does MAP7D1 maintain acetylated tubulin levels? Does it
  regulate acetyltransferases, inhibit deacetylases, or protect acetylated microtubules from
  depolymerization?"; "What is the precise molecular mechanism by which MAP7D1 participates in DNA
  damage response? How do cytoplasmic MAP proteins interact with nuclear DDR machinery...?"
  (genes/human/MAP7D1/MAP7D1-deep-research-cyberian.md).
• Type: blend — biology gap (the acetylation-maintenance and DDR-scaffolding mechanisms are
  unresolved) + curation gap (GOA carries only family-level/IBA terms plus uninformative protein binding; the experimentally supported kinesin-1-activation and acetylated-MT-maintenance roles
  are not captured as specific MF/BP terms).
• Resolve: test whether MAP7D1 depletion alters ATAT1/HDAC6 activity or protects the
  K40-acetylated lattice (in-cell + in-vitro); map the MAP7D1 region binding DDR factors and test
  whether it is required for damage-focus formation.

POLE4 (human) — a structural subunit GO can't describe (ontology-dominant)

• Verdict & framing: this is the project's worked exemplar of the ontology gap — the
  apparent MF-darkness is GO's inability to express "be part of / stabilize the machine," not true
  biological ignorance. POLE4's biology is largely solved.
• Boundary: accessory subunit of DNA polymerase ε with no catalytic activity of its own; a
  structural scaffold required for holoenzyme stability (in Pole4-KO mice the whole Polε complex
  is destabilized) (Bellelli et al. 2018, Mol Cell, PMID:29754823,
  DOI); with POLE3 it is a bona fide histone
  H3–H4 chaperone coupling replication to nucleosome assembly (Bellelli et al. 2018, Mol Cell,
  PMID:30217558, DOI).
• Gap statement: Whether POLE4 has any molecular function beyond a histone-fold
  structural/scaffolding subunit and (with POLE3) an H3–H4 chaperone — and how that scaffolding
  role should be expressed in GO — is the open question, not its biology.
• Provenance (verbatim): "Whether POLE4 plays a functional role in ATAC-mediated transcription
  or whether this association reflects promiscuous histone fold interactions remains to be
  determined." (genes/human/POLE4/POLE4-deep-research-cyberian.md); "The 2024 human Pol ε–PCNA
  structures do not resolve the non-catalytic lobe (POLE2–POLE3–POLE4)..."
  (genes/human/POLE4/POLE4-deep-research-falcon.md).
• Type: ontology-dominant + curation cleanup. GOA captures the biology adequately at the
  complex level (part_of epsilon DNA polymerase complex GO:0008622; involved in DNA replication),
  but the MF rows fall back to uninformative protein binding (×7) and an inferred
  DNA-directed DNA polymerase activity (GO:0003887, TAS) that POLE4 does not itself perform.
• Resolve (ontology + curation, not experiment): annotate the scaffolding role with a
  structural-constituent MF term (structural molecule activity, GO:0005198, or a replisome child)
  and a histone-chaperone-activity MF term for the H3–H4 function; MODIFY/demote the misleading
  polymerase-activity row and the generic protein binding. (Biology residue: a dimerization-vs-
  chaperone separation-of-function allele to test whether processivity and histone-recycling
  phenotypes dissociate.)

AP3B2 (human) — neuronal adaptor, unknown cargo (residual sub-gap)

• Boundary (PubMed-verified): AP3B2 encodes the neuron-specific β subunit of the AP-3
  adaptor complex; the ubiquitous paralog AP3B1 instead causes Hermansky-Pudlak syndrome type 2, and
  biallelic AP3B2 variants cause an early-onset epileptic encephalopathy with optic atrophy (Assoum
  et al. 2016, Am J Hum Genet, PMID:27889060, DOI;
  further cases in Dilber et al. 2022, Clin Neurol Neurosurg, PMID:36356440,
  DOI). In mouse, the neuronal isoform knockout
  (Ap3b2⁻/⁻) reduces selected synaptic-vesicle proteins — the opposite of the ubiquitous
  Ap3b1⁻/⁻ — establishing isoform-divergent SV biogenesis (Newell-Litwa et al. 2009, Mol Biol
  Cell, PMID:19144828, DOI); AP-3 also regulates SV size
  in a brain-region-specific manner (Newell-Litwa et al. 2010, J Neurosci, PMID:20089890,
  DOI).
• Gap statement: The complete repertoire of neuron-specific cargoes selected by the β3B (AP3B2)
  AP-3 isoform — and the molecular basis by which that selection differs from the ubiquitous β3A
  (AP3B1) isoform — is not defined.
• Provenance the gap is real (verifiable facts): the Ap3b2 vs Ap3b1 knockouts have opposite
  effects on SV protein content (PMID:19144828) yet no study enumerates the β3B-specific cargo set;
  correspondingly, the AP3B2 GOA record (genes/human/AP3B2/AP3B2-goa.tsv) carries no
  experimental (IDA/IMP/IGI) cargo or function annotation — every term is IBA/IEA/NAS/ISS/TAS — so
  the isoform-specific cargo specificity is genuinely undetermined, not merely uncurated.
• Type: blend — biology gap (the β3B-specific cargo set / isoform divergence is
  uncharacterized) + ontology dimension (GO annotates the subunit gene rather than the
  β3B-containing complex).
• Resolve: comparative quantitative proteomics of AP-3 vesicles immunoisolated from neuronal
  (β3B) vs non-neuronal (β3A) cells; β3B-specific-KO neuron SV proteomics vs β3A rescue.
• Note: an earlier draft leaned on the -deep-research-cyberian.md summary, whose PMIDs fail
  verification ("PMID:7545544" → nNOS/dystrophin; "PMID:19116307" → ADAMTS13); those are excluded
  and the entry now rests on PubMed-verified primary papers plus the repo's GOA data.

atg2 (SCHPO) — known lipid bridge, unresolved directionality (residual sub-gap)

• Boundary: chorein-N-family lipid-transfer protein that bridges membranes through an extended
  hydrophobic channel, carrying tens of lipids at once (Lees & Reinisch 2020, Curr Opin Cell
  Biol, PMID:32213462, DOI); acts at the phagophore
  rim / ER–phagophore contact and is required for autophagosome formation in S. pombe (Wang et
  al. 2023, Nat Commun, PMID:37553386, DOI).
• Gap statement: How directionality, timing, and regulation of lipid flow are enforced in the
  coupled Atg2–Atg9(scramblase)–Atg18/WIPI system at ER–phagophore contact sites is
  mechanistically unresolved.
• Provenance (verbatim): "precise molecular coordination among ATG2, ATG9, and Atg18/WIPI
  (timing, directionality, and regulation of lipid flow) remains incompletely resolved."
  (genes/SCHPO/atg2/atg2-deep-research-falcon.md); "the calculated transfer rates appeared
  insufficient to account for the tens of millions of lipids required to build an entire
  autophagosome" (genes/SCHPO/atg2/atg2-deep-research-perplexity.md).
• Type: biology gap (residual mechanistic). The core MF/BP/CC are correctly captured by
  experimental GO annotations, so this is not a curation or ontology gap — only the mechanism of
  directional lipid coupling is open.
• Resolve: reconstitute Atg2–Atg9–Atg18 on asymmetric proteoliposomes with leaflet-specific
  reporters to test whether scramblase coupling sets net directionality; cryo-EM of the Atg2–Atg9
  junction at the phagophore rim.

EryCIV / EryCV (S. erythraea) — desosamine-pathway enzymes with unresolved catalytic identity (curation+experiment gap)

From the BGC erythromycin curation (terms/erythromycin_biosynthesis/). These are not unknown
genes — both are clearly TDP-D-desosamine biosynthesis enzymes of the ery cluster — but the
precise catalytic reaction of each is undetermined, and the existing annotations actively
conflict.

• Boundary (what is established): both contribute to TDP-D-desosamine biosynthesis (a
  monosaccharide donor for erythromycin); EryCIV (A4F7N3) is a PLP-dependent DegT/DnrJ/EryC1-family
  enzyme; EryCV (A4F7N2) is a radical-SAM-type enzyme (binds a [4Fe-4S] cluster and SAM). The
  cluster's dedicated PLP sugar aminotransferase is the separate gene EryCI.
• Gap statement: EryCIV's reaction — PLP-dependent 3,4-dehydratase/isomerase (UniProt/EMBL
  name) vs transaminase (GOA term) — is unresolved; EryCV's reaction — 3,4-enoyl reductase
  (UniProt/EMBL name) vs ammonia-lyase (GOA term) — is unresolved.
• Provenance for the gap (checkable repo facts): the GOA files carry only electronic (IEA)
  assignments with no experimental evidence code (genes/SACEN/eryCIV/eryCIV-goa.tsv,
  genes/SACEN/eryCV/eryCV-goa.tsv), and those IEA terms contradict the UniProt/EMBL
  ECO:0000313 names recorded in each gene's *-uniprot.txt. No primary biochemical
  characterization was located (the erythromycin module deep-research run, modules/…-falcon.md,
  explicitly states it "deliberately avoids asserting … the full enzymatic sequence of deoxysugar
  biosynthesis … because direct primary textual support … was not successfully retrieved"). Both
  reviews carry action: UNDECIDED for the catalytic MF.
• Type: curation+experiment gap — the function-kind is known; the exact MF needs a primary
  source (likely Salah-Bey et al. 1998 deoxysugar gene-inactivation work, not yet read) or new
  reconstitution.
• Significance: correct MF terms for these would complete the desosamine arm and let the
  pathway module assert the donor-biosynthesis steps rather than leaving them UNDECIDED.
• Resolve: read the primary desosamine-pathway literature (gene-inactivation / heterologous
  reconstitution) to adjudicate; or reconstitute EryCIV (PLP) and EryCV ([4Fe-4S]/SAM) in vitro on
  the TDP-keto-deoxysugar intermediates. Tracked per-gene in suggested_questions/suggested_experiments.

Methodology

For each candidate gene:

1. Deep literature search — use cached publications, the gene's -deep-research-*.md,
   PubMed, and full text where available. Read primarily for the boundary and for the field's
   own statements of ignorance.
2. Harvest ignorance signals — author hedges ("remains unknown / unclear / to be
   determined"), DUF / "uncharacterized" / hypothetical-protein naming, orphan-activity vs
   orphan-gene framing, and the diagnostic-not-mechanistic trajectory.
3. Judge the type — separate biology gaps from curation gaps from ontology gaps. This is
   the irreducible human/curatorial step.
4. Write the gap entry with full provenance.
5. Route it — experiment (suggested_experiments), ontology (proposed_new_terms), or
   curation (a normal review action).

Selection of which genes to read is curatorial, in the spirit of PomBase's "priority
unstudied genes" determination (Wood 2019) — informed, but not decided, by the triage
read-list.

Reviewing deep-research output

Deep-research -deep-research-{provider}.md files are LLM-generated and are leads, not sources.
Curating this project surfaced a consistent set of failure modes; the table below is grounded in
actual errors caught during this work, because the cost of not catching them is a curated artifact
that cites a real-looking paper that says something else.

Failure mode	Why it's dangerous	Examples caught here
Mismatched-but-real PMID (well-formed, resolves, wrong paper)	Passes format validation; only a content check catches it	RASA1 "36323259" (→ILC paper); KCTD14 "30929316/36362138" (→SGLT/xylanase); AP3B2 "7545544/19116307" (→nNOS/ADAMTS13); Edwards/Oprea/Kustatscher in the intro
Citation by DOI / PMC / citekey only (no PMID)	Not verifiable by PMID tooling; easy to copy a wrong one	falcon files broadly; mxaC (Zhou 2025), tam10, C18orf21, MTC7 (Addinall)
Over-extrapolation (real PMID, but it doesn't support the claim)	The citation "checks out" yet the claim is unsupported	RAB9A: DENND2-as-GEF cited to a general DENN-domain paper on other Rabs
Unlocatable citation (claimed paper not findable at all)	Can't be confirmed or refuted	KCTD14's reported CUL3-non-binding "2024 paper" / PMC10856315
Cross-provider disagreement	One file's confident claim contradicts another's	tam10 domain content; C18orf21 "characterized" vs "uncharacterized"

Mechanisms, layered from automated to manual:

1. Automated PMID resolution + title-match (the missing layer). Today
   src/ai_gene_review/tools/validate_pmid_references.py checks only PMID format and that cited
   PMIDs appear in the review YAML — it cannot catch a real PMID pointing at the wrong paper. The
   high-leverage addition: for each PMID in a deep-research file, fetch metadata (reusing the
   refresh_pmid_titles / publication-cache machinery) and compare the fetched title/first-author/
   year against the text adjacent to the citation; flag low-similarity matches and emit a per-file
   citation-reliability score. This directly catches the dominant failure mode.
2. Ground claims against authoritative repo data, not prose. Each gene folder already ships
   non-LLM ground truth — -goa.tsv (GO terms + evidence codes) and -uniprot.txt. Any claim about
   annotation/function state should be checked against these, and a gap's provenance should prefer a
   checkable repo fact (e.g. "GOA molecular_function is ND / IEA-only") over an LLM hedge. (This is
   exactly how KCTD14 and AP3B2 were re-sourced above.)
3. Cross-provider corroboration. Treat a claim asserted independently by ≥2 providers as
   stronger; treat a single-provider mechanistic claim as a lead; route disagreements to a human.
4. Provider-reliability tracking. Log citation-accuracy per provider over time (with a
   small-sample caveat) so curators can weight accordingly — e.g. in this small sample the
   cyberian files produced mismatched PMIDs twice (KCTD14, AP3B2), and falcon tends to cite by
   DOI rather than PMID.
5. Promotion rule (human-in-the-loop). Nothing enters a curated artifact (review YAML, project
   page) unless its citation is independently PubMed-verified or anchored to a checkable repo
   fact. Verbatim LLM hedges are starting points, never the provenance of record.
6. Quote-and-source discipline. Require a verbatim supporting quote plus a locatable source for
   each promoted claim; a claim with no findable quote+source is, by construction, unverified.

Recommendation: implement mechanism 1 as a validate-deep-research check (or an opt-in,
network-gated mode of the existing tool), reporting a reliability score and a list of
mismatched/unresolvable PMIDs per file. Mechanisms 2, 5 and 6 are process rules already adopted in
this project; 3 and 4 fall out for free once 1 produces structured per-file output.

Relation to existing KB machinery

The building blocks already exist and are reused, not replaced:

Signal	Existing field	Gap role
Open questions	suggested_questions	Seed for biology-gap statements (to be elevated with provenance)
Proposed experiments	suggested_experiments	The "what would resolve it"
Missing GO terms	proposed_new_terms	Ontology-gap entries
Unresolvable annotation	action: UNDECIDED	Often a curation gap (esp. literature inaccessible)
protein binding avoidance rule	curation guideline	Flags MF-mechanism gaps

Schema direction (implemented): this has now been elevated into a provenance-bearing
knowledge_gaps structure (gap statement + boundary + verbatim provenance quotes + kind/aspect/
status typing) so gaps are first-class and queryable — see the Structured curation section above
for the KnowledgeGap class. The project first proved the unit and method on the curated examples
below; the schema element now makes them machine-tractable.

Prioritization (curatorial, not computed)

When choosing what to read next, weight toward gaps that are both high-value and tractable:

• Conservation depth — LECA-deep conservation implies fundamental function (Wood filter).
• Disease / GWAS / IDG overlap — for human genes; the funder hook. (Disease-mechanism
  prioritization can be handed to / shared with Monarch's dismech.)
• A coevolution handle — phylogenetic-profile or co-expression clustering with a known
  module gives instant guilt-by-association and a ready hypothesis.
• Microproteins / smORFs / non-canonical ORFs — a deliberately included frontier where the
  gap is "we did not even know there was a player" (e.g. P3R3URF).

Seed read-list (triage candidates — to be read, not yet curated)

From the conserved-MF-dark triage, after stripping structural-subunit and curation-completeness
false positives. These are candidates for deep reading; the eight marked worked now have
gap entries above.

Gene	Org	Status	Why interesting
CFAP300	human	worked	Dynein preassembly; mechanism unknown; DUF4498 (full exemplar)
P3R3URF	human	worked	95-aa microprotein from a uORF; non-canonical-ORF dark matter
tam10	SCHPO	worked	Meiotic sequence orphan; only ISO/ISS transfers
AGR3	human	worked	Noncanonical PDI-family; catalytic activity / clients unknown
swrD	BACSU	worked	~71-aa swarming-motility enhancer; mechanism unknown
mxaC	METEA	worked	VWA auxiliary protein for methanol oxidation; mechanism unknown
TRAPPC12	human	worked	Moonlighting TRAPP factor; mitotic function ill-defined
KCTD14	human	worked	Least-studied KCTD BTB protein; no experimental function in GOA
C18orf21	human	worked	ORF-named; a closing gap (2025 RNase MRP preprints)
MTC7	yeast	worked	Telomere-capping sequence orphan; all GO aspects ND
RAB9A	human	worked	Residual sub-gap: known Rab, cognate GEF/GAP unidentified
RASA1	human	worked	Residual sub-gap: catalysis solved, scaffolding mechanism unsolved
BAIAP2L2	human	worked	Pinkbar; native epithelial function unmeasured (KO normal)
SCGB1C1	human	worked	Orphan secretoglobin; ligand + receptor unidentified
FGFRL1	human	worked	Kinase-dead FGFR; signaling mechanism unresolved
atg101	SCHPO	worked	Residual sub-gap: WF-finger recruitment partner unknown
irg-1	worm	worked	BP-known / MF-dark; IRG-1 protein activity undefined (MF ND)
METEA mll cluster	METEA	worked	Consolidated: lanthanophore accessory enzymes unreconstituted
MAP7D1	human	worked	Residual sub-gap: paralog-specific MT-acetylation mechanism
POLE4	human	worked	Ontology gap exemplar: structural Polε subunit GO can't express
AP3B2	human	worked	Residual sub-gap: β3B vs β3A neuron-specific cargo undefined
atg2	SCHPO	worked	Residual sub-gap: directionality of Atg2–Atg9 lipid flow

Reproducible read-list (ignorance-signal scan)

A grep of every genes/**/*-deep-research-*.md for author hedges (remains unknown / unclear / undetermined / elusive / uncharacterized / to be determined) and for precise (function|role| mechanism) ... (unknown|unclear|not) surfaced ~60 files carrying explicit ignorance
statements — a reproducible candidate pool beyond the triage table. The strongest leads from that
scan have now been curated (the second- and third-batch entries above). Remaining un-vetted leads
worth reading next, spotted in the same scan: human/PUS3, human/FGFRL1-adjacent CFAP418,
human/SOCS4/SOCS5, human/RFT1, worm/pef-1, worm/fshr-1, SCHPO/alo1, and DESVH/Q72DT1.

Curation caution (learned here): deep-research files frequently cite sources only by DOI,
and some cite PMIDs that do not resolve to the right paper (e.g. the KCTD14 cyberian file). Every
PMID promoted into a gap entry must be verified against PubMed before use; otherwise anchor the
gap to the deep-research file path and verbatim quote.

Prior art and references

• Wood et al. 2019, Open Biology — "Hidden in plain sight" (PMID:30938578)
• Rocha, Freeman et al. 2023, PLOS Biology — Functional unknomics / Unknome database (PMID:37552676)
• Stoeger et al. 2018, PLOS Biology — why important genes are ignored (PMID:30226837)
• Edwards et al. 2011, Nature — "Too many roads not taken" (PMID:21307913)
• Oprea et al. 2018, Nat Rev Drug Discov — unexplored therapeutic genome / IDG (PMID:29472638)
• Kustatscher et al. 2022, Nat Methods — understudied proteins initiative (PMID:35534633)
• de Crécy-Lagard et al. 2025 — enzymes/proteins of unknown function, prediction error types (PMID:40703034)
• Related internal projects: STRUCTURE_FUNCTION.md (dark proteome via structure),
  IBA_REVIEW.md (how orthology propagation manufactures false knowledge),
  OVER_ANNOTATION_PATTERNS.md (annotation that masks gaps).
• External model: monarch-initiative/dismech — Claude-Code-curated disease-mechanism KB with
  the same per-record-YAML + verbatim-quote evidentiary discipline; derives gap/priority
  dashboards over its corpus.

Status

• [x] Literature scan (unknome / Wood; Unknome database; dismech approach)
• [x] Survey of how gaps are captured in the KB today
• [x] Core principle established: gaps are curated judgments, not metrics
• [x] Gap taxonomy (biology / curation / ontology; MF/BP/CC darkness)
• [x] Unit defined (anatomy of a gap entry)
• [x] One worked exemplar (CFAP300)
• [x] Curate the seed read-list into worked gap entries (8 added: swrD, mxaC, TRAPPC12, AGR3, tam10, P3R3URF, KCTD14, C18orf21; all cited PMIDs PubMed-verified)
• [x] Reproducible ignorance-signal read-list established (~60 deep-research files)
• [x] Read-list deepening, batch 2 (8 added: MTC7, RAB9A, RASA1, BAIAP2L2, SCGB1C1, FGFRL1, atg101, irg-1; all adjudicated to real gaps; PMIDs PubMed-verified)
• [x] Third gap framing identified: the residual sub-gap of otherwise well-characterized genes
• [x] Read-list deepening, batch 3 (5 added: METEA mll cluster, MAP7D1, POLE4, AP3B2, atg2; PMIDs PubMed-verified)
• [x] All three gap kinds now have worked exemplars (biology = most; ontology = POLE4; curation = woven through MAP7D1/AP3B2)
• [x] Re-sourced summary-only entries (KCTD14, AP3B2) onto verified primary literature + repo GOA data
• [x] Documented deep-research review mechanisms (failure-mode table + layered checks)
• [ ] Implement validate-deep-research: PMID resolution + title-match + per-file reliability score (extends validate_pmid_references.py)
• [ ] Read-list deepening, batch 4: PUS3, CFAP418, SOCS4/SOCS5, RFT1, pef-1, fshr-1, alo1
• [ ] Decide unit granularity (per-gap vs per-gene narrative)
• [x] Decide home: standalone register vs knowledge_gaps schema element — done: added a first-class KnowledgeGap schema class (gene/annotation/core-function/module/module-node), with the structured register rendered from it
• [ ] Conservation / disease prioritization pass over candidates

Notes

• 2026-06: Project initiated. Brainstorm grounded in (a) the unknome literature, (b) the
  dismech curation model, and (c) how gaps are recorded in this KB today. Key finding from a
  throwaway triage script: apparent "missing molecular function" is dominated (~64% of the
  conserved subset) by GO's inability to describe complex subunits, not by true ignorance —
  reinforcing that the gap call must be made by reading, not counting. CFAP300 curated as the
  first worked exemplar.
• 2026-06: Eight seed-list genes curated into worked gap entries by reading their
  -deep-research-*.md / notes / GOA for the field's own ignorance statements (verbatim quotes).
  Established a reproducible "ignorance-signal grep" over deep-research files (~60 hits) as a
  standing read-list. Verified every promoted PMID against PubMed and caught a deep-research file
  (KCTD14 cyberian) citing PMIDs that resolve to unrelated papers — codified the verify-or-anchor-
  to-file-path rule. Span of gap types now covered: enzyme-adjacent bacterial (swrD, mxaC),
  moonlighting eukaryotic (TRAPPC12), noncanonical-fold (AGR3), sequence orphan (tam10),
  microprotein (P3R3URF), unstudied family member (KCTD14), and a closing gap (C18orf21).
• 2026-06: Second curation batch — eight read-list leads (MTC7, RAB9A, RASA1, BAIAP2L2, SCGB1C1,
  FGFRL1, atg101, irg-1) each adjudicated by reading their deep-research/GOA files; all eight were
  genuine gaps (no spurious calls). Surfaced the residual sub-gap category (RAB9A GEF/GAP, RASA1
  scaffolding, atg101 WF-finger) — a sharp mechanistic hole inside an otherwise well-characterized
  gene. PMID verification again paid off: PMID:36323259, cited by a deep-research summary as the
  RASA1 tandem-SH2 reference, resolves to an unrelated stem-cell/ILC paper and was dropped;
  RASA1's domain architecture is anchored to the file path instead. Sixteen worked entries now
  span bacterial, fungal (budding + fission yeast), nematode, and human genes.
• 2026-06: Third curation batch (5 entries) completes the taxonomy. Added the first worked
  ontology-dominant gap (POLE4 — a structural Polε subunit whose apparent MF-darkness is GO's
  inability to say "structural constituent / scaffold," not real ignorance; the fix is curation +
  a structural-molecule/histone-chaperone MF term, not an experiment), a consolidated cluster
  entry for the M. extorquens lanthanophore (mll) accessory genes (unreconstituted enzymology,
  carefully excluding the now-solved XoxG/XoxJ/MxcQE/MxbDM and the 2024 methylolanthanin discovery,
  PMID:39078674), and three residual sub-gaps (MAP7D1, AP3B2, atg2). PMID verification caught two
  more mis-citations in a deep-research file (AP3B2 cyberian: "PMID:7545544" → nNOS/dystrophin;
  "PMID:19116307" → ADAMTS13), so AP3B2 was initially flagged. 22 worked entries now
  cover all three gap kinds plus the residual-sub-gap framing.
• 2026-06: Re-sourced the two summary-only entries (KCTD14, AP3B2) from primary literature after a
  review flagged that anchoring a gap to an LLM-generated -deep-research-*.md — especially one
  whose PMIDs were just shown to be hallucinated — is near-circular. KCTD14 now rests on the
  PubMed-verified KCTD family review (PMID:24268103) plus the repo's own GOA record (no experimental
  MF); AP3B2 on Assoum 2016 (PMID:27889060), Newell-Litwa 2009/2010 (PMID:19144828, PMID:20089890),
  Dilber 2022 (PMID:36356440) plus its GOA record (no experimental cargo annotation). The
  unlocatable KCTD14 CUL3-non-binding claim was dropped. General principle adopted: provenance for
  a gap must be a verified primary source or a checkable repo fact (GOA evidence codes), never an
  unverified deep-research summary — see the new "Reviewing deep-research output" section.