FLG (human, UniProt P20930) — review notes

Provenance note on deep research

Automated deep research could not be run for this gene: every configured provider
was unavailable in this environment on 2026-08-26.

• falcon (Edison): HTTP 402 — the EDISON_API_KEY is set but the account has no credits.
• perplexity-lite / asta / consensus: no API key present.
• openai: OPENAI_API_KEY present but rejected as invalid_api_key.
• openscientist: working; used for the focused molecular-function hypotheses below
  (FLG-hypotheses/).

Also note that the deep-research-client pin (0.2.7rc1) requires Python >=3.12 while
the repo venv is 3.11.15, so just deep-research-* fails at dependency resolution; the
runs here used uvx --python 3.12.

Per repository policy no self-authored file is named *-deep-research-*.md. Everything
below is my own literature work, with verbatim provenance.

Gene / protein overview

Filaggrin is synthesised as profilaggrin, a ~400 kDa, heavily phosphorylated
polyprotein that is the major structural component of keratohyalin granules (KGs) in the
epidermal stratum granulosum. UniProt P20930 (4061 aa) is annotated with:

• two N-terminal EF-hand domains (residues 6–43 and 49–84) inside an S100-like "A domain";
• Ca(2+) BINDING sites at residues 62, 64, 66, 68 and 73 — i.e. only in EF-hand 2;
• 23 tandem Filaggrin REPEAT features (residues 258–3872), the whole region 255–3971
  flagged Disordered by MobiDB-lite. These are not the filaggrin monomers. Each is a
  49–56 aa Pfam PF03516 match (DR Pfam; PF03516; Filaggrin; 23.), and they occur in
  pairs: the two matches within a pair start 119 aa apart (116 for the first pair), and
  consecutive pairs start ~324 aa apart (observed 321–325, modal 324). That is two
  signature matches per ~324-residue filaggrin unit, so the 23 matches are 11 complete
  pairs plus an unpaired tail (3821..3872) — i.e. the ~11–12 units the literature
  describes, not 23 monomers;
• family assignment "Belongs to the S100-fused protein family" (PANTHER PTHR22571:SF51).

The domain architecture matters for MF curation: the EF-hands and the filaggrin repeats
are on different proteolytic products. Profilaggrin is cleaved during terminal
differentiation, releasing (a) the N-terminal S100/A+B fragment and (b) ~10–12 free
filaggrin monomers. They have different molecular functions and different locations.

Molecular function

1. Intermediate-filament / keratin-filament binding (the aggregation activity)

This is the founding, directly-demonstrated activity, and it belongs to the repeat
unit, not the S100 domain.

PMID:6170061 and PMID:6170061

Mechanism (the "ionic zipper"):
PMID:7687298 The binding is
via the filament rod domains, not the end domains:
PMID:7687298 and
PMID:7687298

Two curation-relevant consequences:

• The in vitro activity is intermediate filament binding in general — the same paper
  reports binding of type III vimentin/desmin as well as type I/II keratin
  (PMID:7687298).
  In vivo the partner is keratin, so GO:1990254 keratin filament binding is the
  physiologically correct child term, with GO:0019215 intermediate filament binding
  as the demonstrated in-vitro superclass.
• Binding is charge/β-turn-driven and sequence-degenerate: a 20–26-residue synthetic
  peptide with ≥5 β-turns and net charge +2 was as effective as full-length filaggrin
  PMID:7687298.
  This is a binding activity of a disordered polycation, not the behaviour of an
  architectural subunit with a defined fold.

Secondary literature states the activity the same way:
PMID:19386895

2. Molecular condensate scaffold activity (keratohyalin granule assembly)

The modern mechanistic account of what KGs are and what filaggrin does to make them:

PMID:32165560 and PMID:32165560

Supporting detail from the same paper:

• PMID:32165560
• Repeat number sets the critical concentration — i.e. the tandem array is the functional
  unit: PMID:32165560. The wild-type baseline matters:
  the effect is a ~65- to 750-fold rise, not a rise from ~130 μM.
• The repeat array is also required for granule formation at all, on a separate assay:
  PMID:32165560
• The S100 domain contributes by dimerising: PMID:32165560, and PMID:32165560
• Condensates and keratin bundles are reciprocally organising, which is the modern
  restatement of "filaggrin aggregates keratin": PMID:32165560
• Dissolution is pH-triggered by the His-rich repeats at the granular→corneum transition:
  PMID:32165560

GO:0140693 molecular condensate scaffold activity ("Binding and bringing together two
or more macromolecules in contact, permitting those molecules to organize as a molecular
condensate") is a precise fit and is not currently annotated to FLG.

3. Calcium ion binding — real, but only in the N-terminal S100 domain

Experimentally demonstrated, not merely inferred from the signature:
PMID:8417356

Note that the 1992 gene paper that GOA/UniProt cite for the domain was explicitly
speculative about function — PMID:1429717 — so PMID:8417356 (and the structure below) are the load-bearing
references, not PMID:1429717.

Structure confirms it, and also shows the EF-hands are not equivalent
(PMID:25760235, PDB 4PCW, 2.2 Å; full text PMC4466033):
each monomer "binds two calcium ions", one at "an N-terminal pseudo/S100 (non-canonical)
calcium binding loop" and one at a canonical C-terminal EF-hand. That matches UniProt,
which places Ca(2+) BINDING residues only at 62–73 (EF-hand 2).
PMID:25760235

So GO:0005509 is correct, but it is a property of profilaggrin / the cleaved N-terminal
fragment, not of mature filaggrin, which contains no EF-hand.

4. Transition metal ion binding — not supported

GO:0046914 comes solely from InterPro2GO on IPR034325 (S-100 domain) via GO_REF:0000002.
Canonical S100 proteins bind Zn2+/Cu2+ at a secondary site at the dimer interface, but:

• the dedicated structural study of this exact domain reports only calcium — no zinc or
  other transition metal is discussed anywhere in PMID:25760235 / PMC4466033;
• UniProt P20930 lists no transition-metal BINDING site, only the five Ca(2+) sites;
• no primary experiment demonstrating metal binding by FLG was found.

This is a signature-level over-propagation. (OpenScientist hypothesis run below tests it
independently.)

5. "protein binding" (GO:0005515, IPI with KLK5)

The interaction is enzyme–substrate, not a functional partnership of FLG:
PMID:23629652 and
PMID:23629652 Per repo curation guidance, bare protein binding conveys no
functional information and should not stand as a core MF.

The N-terminal domain does have genuine, pocket-mediated protein partners
(PMID:25760235 — annexin II/p36, stratifin/14-3-3σ, HSP27), which would be the
informative thing to capture if anything.

6. Is "structural molecule activity" the right MF?

GO:0005198 is defined as "The action of a molecule that contributes to the structural
integrity of a complex"; GO:0030280 as "The action of a molecule that contributes to the
structural integrity of an epidermal cutaneous structure."

There are two distinct claims that must be kept apart:

• The keratin-aggregation role is a binding activity, not a structural one. Mature
  filaggrin is an intrinsically disordered polycation that transiently condenses
  pre-formed keratin filaments and is then completely proteolysed to free amino acids
  (PMID:8417356). It does not persist as an architectural
  subunit of the structure it organises.
• The cornified-envelope role genuinely is structural. Filaggrin is covalently
  isodipeptide cross-linked into the CE: PMID:7543090 That is the fact behind the IDA to GO:0030280, and it
  survives review — GO:0030280 (the specific epidermal child) is the better term than the
  bare GO:0005198 IEA.

7. Terminal catabolism to natural moisturising factor

Not an MF of FLG itself (FLG is the substrate) but essential context for why the barrier
phenotype exists:
PMID:21654840 and PMID:21654840 Deiminated (citrullinated) filaggrin is
broken down by bleomycin hydrolase PMID:19286660; FLG is citrullinated in vivo
PMID:8780679.

Localisation

• Keratohyalin granule (GO:0036457) — the defining location of profilaggrin.
  PMID:1429717
  UniProt: "In the stratum granulosum of the epidermis, localized within keratohyalin
  granules (PubMed:1429717)."
• Cornified envelope (GO:0001533) — covalently cross-linked component PMID:7543090.
• Nucleus (GO:0005634) — genuine, but for the cleaved N-terminal domain in
  keratinocytes, and not established by the sperm-nuclear-proteome dataset GOA cites.
  PMID:9800950 The NLS is functional:
  PMID:12230510
• HPA "cytoplasmic bodies" → GO:0036464 is a mapping artefact. HPA reports FLG
  localised to cytoplasmic bodies in HaCaT keratinocytes (Enhanced reliability,
  antibodies HPA027505/HPA030188). In HaCaTs those FLG-positive cytoplasmic puncta are
  KG-like condensates — this is the exact cell line and structure characterised in
  PMID:32165560 — not ribonucleoprotein granules. GO:0036464 requires RNP content;
  filaggrin has no RNA-binding function. The correct term is GO:0036457.
• Extracellular matrix (GO:0031012) from a multiple-myeloma bone-marrow ECM
  proteome PMID:28344315 is not credible for a 4061-aa intracellular epidermal protein;
  keratin/skin-derived contamination is the standard confounder in such datasets.

Phylogenetic context (the IBAs)

The three IBAs (GO:0001533, GO:0036457, GO:0061436) come from PAINT node
PANTHER:PTN002503122 in PTHR22571 ("S100-fused Epidermal Structural Protein"). Donors
resolve to the S100-fused-type protein (SFTP) clade: human FLG (P20930, the target
itself), human HRNR/hornerin (Q86YZ3), human FLG2 (Q5D862), mouse Hrnr (MGI:3046938)
and mouse Flg2 (MGI:3645678). This is a coherent, tight clade with shared
architecture (S100 A-domain + tandem LC repeats) and shared epidermal-barrier biology, and
FLG appearing in its own WITH/FROM is the expected marker that experimental grounding
exists on the target. All three IBAs are consistent with the direct evidence above.

OpenScientist hypothesis runs

Three focused MF hypotheses were dispatched (see FLG-hypotheses/):

1. function-hypothesis-go-0005509 — does the N-terminal S100 domain really bind Ca2+,
   and is EF-hand 1 a degenerate pseudo-EF-hand?
   Verdict: SUPPORTED. Convergent structural, biochemical, sequence and evolutionary
   evidence; EF-hand 2 (62–73) is a textbook canonical loop with the hallmark bidentate
   Glu73, EF-hand 1 is the S100-type pseudo-EF-hand. Recommends retain + upgrade IEA→IDA
   on PMID:25760235, with a scope note that this is the N-terminal domain's property, not
   the repeat's. Surfaced PMID:32893105, which shows the A+B domains bind annexin II and
   keratin IFs calcium-dependently — now cited in the review.
2. mf-keratin-binding-vs-structural — is the mature repeat unit's MF keratin-filament
   binding rather than structural molecule activity?
   Verdict: SUPPORTED, with a qualifier. Independent sequence biophysics found every
   repeat region intrinsically disordered (Uversky charge–hydropathy, Chou–Fasman below
   average for both helix and sheet, TOP-IDP disorder fraction 0.92–0.98), strongly
   hydrophilic, and cationic only at acidic skin pH. Recommends GO:0019215/GO:1990254 as
   the primary MF and reframing GO:0030280 as a distinct, secondary, cornified-envelope
   role — which is exactly how this review wires it. Surfaced PMID:6195345, the direct
   in vitro demonstration of keratin bundling, now cited.
3. function-hypothesis-go-0046914 — is the InterPro2GO transition-metal-binding call
   warranted for this protein?
   Verdict: over-annotated / refuted as a direct function. A HETATM audit of 4PCW
   finds 8 calcium atoms and zero transition-metal atoms; residues 1–91 contain only two
   histidines (His60, His65) and no cysteines, lacking the His3Asp/His4–6 or Cys-cluster
   chemistry that characterised metal-binding S100 proteins use, and those histidines sit
   8–15 Å from any calcium site. A comparative audit shows all six human SFTPs (FLG, FLG2,
   HRNR, RPTN, TCHH, CRNN) carry GO:0046914 as IEA:InterPro with no member-specific
   evidence, while genuinely Zn-binding S100s (S100A7/A8/A9/B) carry the specific child
   GO:0008270 with experimental evidence. Caveat recorded in the review: no direct
   Zn/Cu binding assay has been done on this domain, so the case is structural, sequence
   and comparative rather than a direct negative experiment.

All three runs were launched blind to this notes file and agreed with the independent
reading above; where they diverged from my initial framing it was to sharpen it (the
GO:0030280 "secondary structural role" split), not to overturn it.

Operational note for the skill: --as-function-hypothesis silently discards
--context (build_function_hypothesis_record in
scripts/gene_hypothesis_deep_research.py rebuilds term_context from the annotation
alone). To pass scoping context, use a free-text --hypothesis with --focus-type function-support, which does honour --context.