COSCIENTIST

The gap

Most AI curation help is literature synthesis: read papers, summarize, cite.

It cannot answer questions the literature never asked:

• Does this orphan fold imply a function?
• Are the catalytic residues actually present?
• Is this annotation an over-propagated electronic inference?

These need running analyses, not just reading.

The approach

Treat OpenScientist as a blinded, independent bioinformatics scientist
testing one hypothesis: gene G has function F.

• Executes code over iterations: structure fetches, Foldseek, sequence /
  active-site analysis, plotting — with provenance artifacts
• Holdout discipline: local *-bioinformatics/RESULTS.md is withheld from
  the prompt, then compared after the run
• Verdict + verbatim-checked quotes wired into GENE-ai-review.yaml

just gene-hypothesis-research openscientist <ORG> <GENE> --hypothesis "…"

Where it adds the most value

Steer it at questions the papers can't decide:

• uncharacterized / orphan proteins (predicted-but-unverified function)
• suspected over-annotations (IEA / IBA / ISS propagation)
• uninformative "protein binding" molecular functions
• pseudoenzymes / missing catalytic residues
• paralog discrimination

Highest-value pattern: catching a systematic mis-annotation that propagated
across a family or paralog.

Completed runs — structure / sequence

Completed runs — core-vs-non-core / over-annotation

Systematic mis-annotation catches

The highest-value outcomes were family-level errors, not single genes:

• MJ1511 → MJ0742: same erroneous oxidoreductase IBA on a catalytically
  dead paralog (crystal structure available)
• CLCN7: transepithelial-transport term traced to a ComplexPortal
  family-level sentence → propagated by PANTHER IBA to ~1,198 ortholog
  annotations
• RvY_17310: independently flagged the degenerate paralogs
  (RvY_13070 / RvY_13431) while clearing the target itself

Finding: execution depends on the question

• Structural questions → it executes code (AlphaFold fetch, Foldseek,
  geometry) with provenance artifacts
• Topology / regulatory / motif questions → it mostly reasons over domain
  architecture, sorting motifs, ChIP/E-box, InterPro/PANTHER families

Verdicts were strong either way — but the provenance differed.

Closing the gap: template tweak + A/B test

Tweaked the prompt to execute hypothesis-matched analyses (and save
provenance), with a hard never-fabricate / say-so-if-web-only clause.

CLCN7 A/B — only the template changed:

Topology behaviour is promptable; ChIP/expression remain tool-access-limited.

Proteostasis Network use cases

The PN reviewers assigned families by InterPro / CDD domain — not by
structure or catalytic-residue analysis. That layer is exactly the gap.

• HSPA12A / HSPA12B — HSP70 by architecture; do they retain folding
  machinery? (their InterPro set lacks HSP70-specific domains HSPA8 has)
• NPLOC4 — DUB > MPN: real JAMM metalloprotease, or pseudo-DUB?
• AARSD1 — AlaX editing residues present vs HSP90-cochaperone artifact?
• GABARAPL3 — ATG8 ortholog flagged "pseudogene": functional?

Evidence integrity & operational lessons

• No hallucinated citations across runs; every wired supporting_text is a
  verbatim substring of its source
• Give jobs time: ~50–90 min; inject timeout=7200 (API max) + subprocess
  wall 8100 s; 3600 s default cancels structural runs mid-analysis
• Empty output ≠ done: timeout vs transient req_… LLM error; clear the dir
  and re-run
• Holdout local bioinformatics for post-run comparison