OpenScientist adjudication of borderline and disputed ProtNLM2 predictions
Status (2026-10-01): these verdicts are advisory evidence. The per-gene
*-protnlm-predictions-review.yamlfiles are the calls of record. A term-by-term
reconciliation (generated by
reconcile_openscientist.pyfrom a
transcription of the verdicts below
against the current reviews) finds, across 31 GO terms in 21 genes:
- 13 agree with the current YAML;
- 5 differ within the same polarity group: COR vs CNN for DROVI/B4MAQ2, CALMI/A0A4W3GVU1
and GADMO/A0A8C5FPT8 (whether the term is already in the target's cached GOA), and PLI vs NPI
for ORYSI/B8BAB0 pollen maturation and ARATH/F4JLB7 kinase activity;- 12 are decided differently: 9
NPI → UNC(four WHEAT/A0A3B6RKV1 light and hormone
terms, two PHATC/B7FXQ8 stress responses, two WHEAT/F6LAX4 chromosome terms and
XENTR/A0A8J0SCI2), F4JLB7 phosphorylationPLI → UNC, ORYSJ/Q6YYC5 K63-linked
ubiquitinationUNC → PLI, and F6LAX4 protein heterodimerizationNPI → LSP;- 1 has no verdict (ABRPR/A0A8B8L1Z3; the run failed).
Since the first reconciliation (2026-09-27), the reviews have been revised to incorporate these
reports, and 30 of the 31 reviewed terms now cite them, including all 12 that differ. Most
differences follow one rule in the project's evidence standards: missing evidence leads to
UNC unless there is contrary evidence. OpenScientist scored "not transferable by orthology"
or "over-specific" as NPI, while the reviews treat those as unestablished (UNC). Round 2 below
states only "6 animal/mitotic terms" for F6LAX4 without listing them, so that transcription
covers all six emitted terms and is the least certain row.
Bottom line
Every ProtNLM2 prediction the ARGO-ProtNLM-50 review had scored UNC (genuinely
on the fence — "cannot validate or refute from the evidence at hand") was re-tested as an
independent, blinded gene-function hypothesis by an OpenScientist agent (structure +
comparative genomics; the suspected answer was withheld from the prompt). 15 UNC terms
across 11 genes; 10 genes adjudicated (14 terms), 1 gene failed upstream.
The result splits cleanly by prediction type:
- Localization calls held up — 3/3 correct, and one (ARTAN) caught a mis-localized
curated annotation. But they systematically undersell a specific protein identity
(Exportin-5, ESRP1 flattened to a generic "cytoplasm"/"nucleus"). - Specific function / family-process refinements over-reached — 9/14 →
NPI(wrong
subfamily, over-specific substrate, or untransferable ortholog-derived developmental roles). - 2/14 were genuinely undecidable from sequence and were kept as
UNCleads.
This mirrors and sharpens the main evaluation's finding that ProtNLM2 "captures broad
functional categories but lacks resolution": where it names a compartment it is usually
right; where it names a specific mechanism or process it tends to over-generalize from
family/ortholog associations.
A second round then tested the opposite population — the 10 genes whose predictions the
review had disputed (NPI/PLI, "the prediction is wrong") — as the same blinded
hypotheses, to check our own rejections. 9 of 10 were independently confirmed as
misassignments; 1 (GADMO/tbc1d14) was partially overturned (see the dispute-confirmation
section). Together the two rounds show the adjudicator moves in both directions: it upholds
well-reasoned rejections and catches an over-harsh one.
Round 1 — borderline (UNC) verdicts
| Gene | UNC prediction(s) | Verdict | New assessment | Why |
|---|---|---|---|---|
| ARTAN/A0A2U1PS28 | chloroplast (GO:0009507) | Supported | UNC → COR |
~78% id to Arabidopsis chloroplastic paralog Q9FNM5 vs ~47% to the mitochondrial one, + chloroplast transit-peptide signature. The GOA/HAMAP mitochondrial annotation is the error. |
| DROVI/B4MAQ2 | cytoplasm (GO:0005737) | Supported (incomplete) | UNC → COR |
It is Exportin-5; cytoplasm is valid (obligatory cytoplasmic phase) but undersells a Ran-GTP nucleocytoplasmic shuttle. |
| CALMI/A0A4W3GVU1 | nucleus (GO:0005634) | Supported | UNC → COR |
It is ESRP1 (paralog ESRP2 ruled out); human ortholog is IDA-nuclear. |
| STRCO/Q9L243 | deoxyribonucleotide catabolism (GO:0009264) | Over-annotated | UNC → NPI |
A generic HAD phosphohydrolase (PF18143 RNA-repair family), not a dNMP-specific 5′-nucleotidase; the process term has no pathway support. |
| AQUCT/A0A2G9RZF1 | single fertilization (GO:0007338) | Refuted | UNC → NPI |
A 156-aa CUB-domain-only protein with no serine-protease domain or catalytic triad; real ovochymases are large multidomain S1 proteases. Family over-annotation. |
| ORYSI/B8BAB0 | pollen maturation (GO:0010152) | Refuted | UNC → NPI |
Belongs to the PG1β-like cell-wall/pectin BURP clade, not the anther-specific RAFTIN clade that supplies the reproductive precedent. Wrong-subfamily over-generalization. |
| WHEAT/A0A3B6RKV1 | photomorphogenesis, GA signaling, red light, seed germination (GO:0010099/0010476/0010114/0010030) | Split | UNC → NPI ×4 |
A catalytically competent JMJ22 co-ortholog (correcting the "KDM5/JARID" mislabel — it is a JMJD6-type histone arginine demethylase), but the four terms are Arabidopsis IMP redundant-pair phenotypes not transferable to wheat by orthology alone. |
| PHATC/B7FXQ8 | salt stress, H₂O₂ (GO:0009651, GO:0042542) | Over-annotated | UNC → NPI ×2 |
A bona fide small-HSP/ACD chaperone whose only defensible family process is response to heat; salt is weakly plausible only by cross-kingdom analogy and H₂O₂ is partly contradicted by sHSP precedent. |
| ORYSJ/Q6YYC5 | K63-linked ubiquitination (GO:0070534) | Undecidable | kept UNC (lead) |
A real RGLG-family RING E3 ligase, but chain-linkage specificity is E2-determined and not decidable from sequence/orthology — needs an in-vitro chain-linkage assay. |
| SOYBN/C6T1A2 | neg. reg. ABA signaling (GO:0009788) | Undecidable | kept UNC (lead) |
A real Q-type (QALGGH) C2H2 zinc-finger TF (MF scaffold solid), but the specific ABA-negative-regulation role is overexpression-only and should not be a direct annotation. |
| ABRPR/A0A8B8L1Z3 | endoplasmic reticulum (GO:0005783) | — | unchanged | Run failed 4× with a transient upstream error (Request ID: …); gene-specific, not attempted further. |
Round 2 — dispute confirmation (NPI / PLI)
The 10 genes whose predictions the review had disputed were re-tested the same way
(neutral hypothesis: "is this prediction supported or a misassignment?"). This checks whether
our own rejections hold up — and, as ARTAN showed in round 1, an independent agent can overturn
a call.
| Gene | Disputed prediction (our call) | OpenScientist finding | Outcome |
|---|---|---|---|
| ARATH/F4JLB7 (RIC7) | kinase activity + phosphorylation (PLI) |
no kinase catalytic domain (LRR-RLP; no VAIK/HRD/DFG) | confirmed |
| MYTGA/A0A8B6GS20 | PI3P phosphatase (NPI) |
catalytically dead pseudophosphatase (MTMR9-like) | confirmed |
| ASPOR/Q2U1U6 | O-glycosyl hydrolase (NPI) |
polysaccharide lyase, not GH (and too small for either) | confirmed |
| WHEAT/F6LAX4 | 6 animal/mitotic terms (NPI×6) |
PP2A A scaffold; terms taxon-inappropriate or B56-subunit (not A) | confirmed |
| CAEEL/mcm-4 (A0A061AL94) | transcription initiation (NPI) |
74-aa MCM4 replicative-helicase fragment | confirmed |
| XENTR/A0A8J0SCI2 | DNA-binding TF activator (NPI) |
naked C2H2 array, no effector domain → direction not sequence-determinable | confirmed |
| DANRE/dnajc6 (A0A8M9QG43) | dephosphorylation (NPI) |
PTEN-fold pseudophosphatase, no intact catalytic site | confirmed |
| DROPS/A0A6I8W8A2 | ligase activity (NPI) |
169-aa RCC1 β-propeller, no HECT/RING (HERC3 name-transfer) | confirmed |
| XENNA/D3VIU4 | ligand-gated ion channel (NPI) |
periplasmic SBP; iGluR fold homology but no TM pore | confirmed |
| GADMO/tbc1d14 (A0A8C5FPT8) | autophagosome (NPI) |
defensible — human ortholog has a curated IDA | overturned → CNN |
9 confirmed, 1 overturned. The overturn is the informative case: our review flatly rejected
autophagosome for tbc1d14, but human TBC1D14 (Q9P2M4) carries a direct experimental IDA
autophagosome annotation (Longatti et al. 2012, PMID:22613832) plus a phylogenetic IBA
(verified independently via QuickGO). The prediction is correct-by-orthology, so it was
reassessed NPI → CNN — with the caveat, on which both analyses agree, that autophagosome is a
secondary, condition-dependent localization (the recycling endosome is primary and the role is
regulatory). The other nine confirmations recapitulate the main evaluation's failure modes:
paralog/pseudoenzyme over-annotation, cross-kingdom / taxon-inappropriate transfer, over-specific
directional child terms, and RefSeq-name-transfer artifacts.
Curation leads beyond the predictions
Three findings warrant fixes to the genes' own reviews, independent of the ProtNLM assessment:
- ARTAN/A0A2U1PS28 — the existing GOA/UniRule mitochondrial localization (GO:0005743,
GO:0005759 via HAMAP-Rule MF_03137) is very likely wrong; the protein is a chloroplastic
cpLEPA/EF4. Candidate for correction inA0A2U1PS28-ai-review.yaml. - WHEAT/A0A3B6RKV1 — the review text repeatedly labels the enzyme "KDM5/JARID1"
(H3K4 lysine demethylase); it is actually a JMJD6-type histone arginine demethylase
(F-box + Cupin_8/JmjC), which changes the H3K4-directionality reasoning. - ARATH/F4JLB7 — a gene-symbol collision: F4JLB7 is labelled "RIC7" but appears to be an
LRR receptor-like protein, whereas the characterized CRIB-motif RIC7 (ROP2 effector) is the
adjacent locus At4g28556 (Q1G3K8). Worth disentangling in the gene review.
Method & provenance
Each hypothesis was run with just gene-hypothesis-research openscientist <ORG> <GENE>
(3 iterations, 7200 s job timeout), framed neutrally as computational_prediction, with the
suspected answer withheld. No local *-bioinformatics/RESULTS.md was fed to the agent (there
was none for these genes). Full reports and provenance artifacts live under each gene's
genes/<ORG>/<GENE>/<GENE>-hypotheses/<slug>/openscientist.md. Verdicts are wired into the
per-gene *-protnlm-predictions-review.yaml files with verbatim supporting_text quotes.
(Correction, 2026-09-27: this was not carried out. Only ARATH/F4JLB7 cites the investigation, and
most verdicts differ from the current reviews; see the reconciliation.)
Caveat: all verdicts rest on computational evidence (sequence, structure, orthology,
comparative genomics) without new wet-lab data; the two COR localization upgrades and the
ARTAN curation lead are strong but should be curator-confirmed.