FA cohort — Affinage as a deep-research *input* to AIGR (22 Fanconi-anemia genes)

FA cohort — Affinage as a deep-research input to AIGR (22 Fanconi-anemia genes)

Earlier cohorts (pilot, batch2–4, narrative-vs-go)
evaluated Affinage's GO grounding against finished AIGR reviews. This cohort asks the
forward question the project page flagged as "weaker than it first appears": used as a
deep-research source, does Affinage's narrative actually improve a curated review?

The 22 Fanconi-anemia complementation-group genes (A–W) were freshly reviewed de-novo, then
each Affinage record was read against its review and its cited primary papers folded in
conservatively (see per-gene files in fa-cohort/). Unlike the earlier
cohorts this measures net curation value, not just exact-GO-id overlap.

Headline

So the two layers behave exactly as the narrative-vs-go analysis
predicted: the narrative is a strong, citation-dense research input; the GO layer is not
directly usable. The novelty here is quantifying the narrative's upside — it is real.

Per-gene outcome

Run dates 2026-06-09→11. "New" = new GO annotations added; "Papers" = Affinage-surfaced
primary papers folded into the review.

Gene FA Affinage self-eval Papers New What Affinage genuinely added (or the key disagreement)
FANCA A win (29) 2 2 Intrinsic ssDNA binding + RAD52-like single-strand-annealing (SSA) activity — a real biochemistry gap, now non-core (GO:0003697, GO:0045002)
FANCB B — (8) 2 0 Anchored HR/SCE annotations; MF mistagged "catalytic activity acting on protein" (FANCB is non-catalytic)
FANCC C win (34) 3 0 Its own cited separation-of-function mutants reinforce AIGR's non-core call on the redox/PKR roles
FANCD2 D2 win (60) 6 3 Fork protection/restart (→core), R-loop suppression, mitochondrial pool (GO:0031297, GO:0062176, GO:0005739)
FANCE E win (10) 4 0 4 primary adaptor/bridge-to-FANCD2 papers; MF over-reach "catalytic activity acting on protein"
FANCF F win (9) 2 0 The FANCF flexible-adaptor primary paper (PMID:15262960); blurs regulation-of / plant-meiosis into function
FANCG G win (30) 2 1 HR DSB-repair (Ser7 D1-D2-G-X3 complex) as a discrete annotation (GO:0000724)
FANCI I win (43) 3 0 Primary DNA-binding / FANCD2-monoub-stimulation papers; wrong-branches FANCI as a protein-acting enzyme
FANCL L win (22) 1 1 ELF-domain ubiquitin binding (GO:0043130); AIGR's RING-E3 term beats Affinage's generic ligase parent
FANCM M win (41) 3 1 Translocase-dependent ATR/Chk1 checkpoint signaling (GO:0000077)
BRIP1 J win (38) 1 0 No BACH1-bZIP collision this time; coarse layer mislabels the catalytic helicase "molecular adaptor"
PALB2 N win (35) 2 1 ChAM nucleosome binding (GO:0031491) + the primary KEAP1 paper; layer says histone-vs-nucleosome / catalytic-on-DNA
RAD51C O win (42) 5 1 ICL repair / FANCO founding paper (GO:0036297); does not claim RAD51C is itself an endonuclease — AIGR's over-annotation of GO:0008821 stands
SLX4 P win (57) 3 1 SLX4-complex SUMO E3 ligase (GO:0061665, non-core); narrative correctly frames SLX4 as nuclease-dead scaffold
ERCC4 Q tie / gate tripped (40) 1 0 The FANCQ disease-defining paper (PMID:23623386); flag warranted vs the GO layer, not the (sound) prose
RAD51 R win (45) 3 0 Foundational Gupta/Radding 1997 recombinase paper; RNA/TERRA left as documented candidate
UBE2T T win (35) 1 0 Anchored polyubiquitination; declines the large single-study cancer-substrate / RNF8 / NER set
XRCC2 U win (29) 5 1 ICL repair / FANCU (GO:0036297); narrative reinforces "not an ATP-dependent damage sensor"
RFWD3 W win (17) 2 0 Two RFWD3-focused fork-remodeling papers; p53/MDM2 correctly non-core
BRCA1 S tie / gate tripped (22) 3 0 Anchored HR/E3/H2AK127 terms; flag warranted vs GO layer (spurious RNA-catalytic term), not the prose; ~70 protein-binding REMOVEs untouched
BRCA2 D1 win (27) 3 0 3 canonical primary papers (Yang 2002 structure, fork protection, single-molecule RAD51 loading) replacing indirect support
MAD2L2 V win (47) 2 1 Shieldin-independent fork-resection protection (GO:0110027) + the previously-uncited FANCV disease paper

Cross-cutting findings

1. The GO layer's failure mode on this cohort is systematic: adaptors/scaffolds get typed as
enzymes. The FA pathway is rich in non-catalytic structural subunits, and Affinage's
mechanism_profile mislabels them almost every time — catalytic activity, acting on a protein for the non-catalytic adaptors FANCB / FANCE / FANCI, catalytic activity, acting on DNA for FANCA / PALB2 / SLX4, structural molecule activity for FANCA / FANCG, generic
ligase activity for the specific RING-E3s FANCL / RFWD3 / BRCA1, and — inverting the error —
molecular adaptor activity for the catalytic helicase BRIP1/FANCJ. None were importable;
AIGR's curated MF terms are correct in every case. This is the earlier cohorts' "collapses to a
general parent" finding, sharpened: on structural subunits it doesn't just generalize, it lands
on the wrong catalytic branch.

2. The narrative, by contrast, is a high-value research input. 59 primary papers and 13 new
annotations came out of it, including one genuine biochemistry gap (FANCA's intrinsic SSA
activity) and several functions the review had only in prose (FANCD2 fork protection, FANCM ATR
checkpoint, RAD51C/XRCC2 ICL-repair, MAD2L2 fork-resection control). Critically, several times
Affinage's own cited evidence reinforced an AIGR non-core/over-annotation call it superficially
seemed to challenge (FANCC redox mutants; RAD51C not-an-endonuclease; XRCC2 not-a-damage-sensor;
SLX4 nuclease-dead). This is the opposite of the redundancy the project page worried about — on a
genuinely hard, adaptor-heavy pathway the narrative pulled real, checkable literature the
GOA-seeded review had not surfaced.

3. The two tie records (ERCC4, BRCA1) were correctly gate-flagged, but for the GO layer,
not the prose. In both, the narrative was factually sound and PMID-dense with no fabrication or
mis-attribution; the tie tracked the coarse/wrong-branch GO grounding (a spurious RNA-catalytic
term on BRCA1; RNA-binding + parent-level catalytic terms on ERCC4) plus a breadth-over-focus
narrative. So the self-eval gate is a useful "scrutinize the GO layer" signal, not a
narrative-reliability verdict.

4. Net curation effect. Every FA review is now better sourced (mixed-maturity deep-research /
interactome support replaced by primary literature) and 10 are functionally more complete, with
no decision weakened. Used the way this project endorses — narrative-as-input, GO-layer-ignored —
Affinage delivered measurable value on the cohort where it should have been most redundant with
AIGR's own deep-research step.

Caveats