BIOINFORMATICS Case Studies

Bottom line: when a GO function rests on family membership, a reproducible
sequence and structure analysis can say whether the target still has the
parts the function needs. This page records two such cases in S. pombe and
the workflow that grew out of them. For pmp20, a scripted comparison against
three active peroxiredoxins, a 27-protein Prx5 panel and AlphaFold models found
a single cysteine (C43) and no resolving cysteine; the review removes the IBA
thioredoxin peroxidase activity row and keeps the curated NOT peroxidase activity. For Epe1, the review removes all seven JmjC catalytic and
metal-binding rows. Both genes were then re-tested as blinded OpenScientist
hypotheses (July 2026), which the page below still describes as a planned run:
the Epe1 run pinned the defect to Tyr370 at the third Fe(II) ligand, and one
of the two pmp20 runs cautioned that a missing resolving cysteine is normal
for 1-Cys peroxiredoxins, so the pmp20 case rests mainly on the negative assay
(PMID:20356456). The same hypothesis workflow now has 264 -hypotheses/
folders across genes/.

We did this because an analysis that runs once on one gene is weak evidence;
pmp20-bioinformatics is built to be rerun (just all), uses active
controls, and was tested on a second target (tpx1), so it serves as the
template for new cases.

Purpose

This document records concrete cases where custom bioinformatics analysis was used to resolve annotation uncertainty in ai-gene-review.

Each case links:
- biological question
- reproducible analysis workflow
- key outputs
- impact on curation decisions

Hypothesis-Linked OpenScientist Workflow

Gene function assignments can be treated as focused hypotheses before they are
accepted, removed, or converted into suggested experiments. The current CLI path
is:

just gene-hypothesis-list <organism> <gene>
just gene-hypothesis-research <provider> <organism> <gene> --annotation-term-id <GO_ID> --as-function-hypothesis --dry-run

Hypothesis reports are written under:

genes/<organism>/<gene>/<gene>-hypotheses/<hypothesis-slug>/<provider>.md

When the source review cites a local bioinformatics result such as
file:<organism>/<gene>/<gene>-bioinformatics/RESULTS.md, the hypothesis prompt
withholds that local analysis from OpenScientist. This keeps the run blinded so
OpenScientist can independently explore the hypothesis that gene G has
function F. After the run, compare the OpenScientist report against the local
RESULTS.md to see whether the independent search converged with, refined, or
contradicted the older analysis.

Use this distinction:

• OpenScientist-only hypothesis search: blinded exploration of whether gene
  G has function F, using seed review context but not local holdout
  bioinformatics results or prior curation rationales.
• Bioinformatics workflow plus OpenScientist: sequence, structure, domain,
  phylogeny, or localization claims where local computation is part of the
  post-run comparison and final evidence trail.

Case Studies

BIO-001: SCHPO/pmp20 (thioredoxin-dependent peroxidase activity)

Context

pmp20 is in the peroxiredoxin family, but automated/domain-based annotations suggested peroxidase activity while literature and curated UniProt comments indicated loss of thioredoxin-dependent activity.

Curation question

Should GO molecular function terms related to peroxidase/thioredoxin peroxidase activity be retained or removed for SCHPO/pmp20?

Analysis workflow

• Analysis directory: genes/SCHPO/pmp20/pmp20-bioinformatics
• Run command: just all
• Main run dataset: genes/SCHPO/pmp20/pmp20-bioinformatics/inputs/proteins.tsv
• Alternate test dataset (script generality): genes/SCHPO/pmp20/pmp20-bioinformatics/inputs/proteins-test-tpx1.tsv

Methods implemented:
1. Collect target + active controls from local UniProt TXT and UniProt REST.
2. Compare cysteine topology and catalytic-site correspondence at sequence level.
3. Map target residues against active-control resolving-cysteine positions by global alignment.
4. Fetch AlphaFold monomer models and compute CYS SG pair geometry summaries.
5. Validate that the same scripts run on an alternate target (tpx1).

Key outputs

• Summary report: genes/SCHPO/pmp20/pmp20-bioinformatics/RESULTS.md
• Sequence cysteine summary: genes/SCHPO/pmp20/pmp20-bioinformatics/results/main/sequence/sequence_cysteine_summary.tsv
• Target vs active-control mapping: genes/SCHPO/pmp20/pmp20-bioinformatics/results/main/sequence/target_vs_active_alignment.tsv
• AlphaFold download manifest: genes/SCHPO/pmp20/pmp20-bioinformatics/results/main/structures/alphafold_manifest.tsv
• Structure cysteine summary: genes/SCHPO/pmp20/pmp20-bioinformatics/results/main/structure/structure_cys_summary.tsv

Main findings

• pmp20 has one cysteine (C43) and no candidate resolving cysteine.
• Active controls show expected multi-cysteine topology with resolving-cysteine candidates.
• Active-control resolving positions do not map to cysteine in pmp20.
• AlphaFold monomer model for pmp20 contains only one cysteine, so no Cys-Cys pair geometry is possible in-model.

Curation impact

The bioinformatics case study was incorporated into the main review as a file: reference:

• file:SCHPO/pmp20/pmp20-bioinformatics/RESULTS.md

Updated review file:

• genes/SCHPO/pmp20/pmp20-ai-review.yaml

Validation status:

• just validate SCHPO pmp20 passed (warnings only, no validation errors).

Caveats

• AlphaFold was used as monomer structure context only; no multimer interface modeling was performed.
• Structural distance metrics are supportive context and not stand-alone activity evidence.

BIO-002: SCHPO/Epe1 (JmjC demethylase activity)

Context

Epe1 contains a JmjC-family domain, which can trigger automated histone
demethylase, dioxygenase, oxidoreductase, and metal-binding annotations. The
review instead treats Epe1 as a non-catalytic anti-silencing factor that acts in
heterochromatin regulation through Swi6/HP1-associated chromatin mechanisms.

Curation question

Should JmjC-derived GO molecular-function annotations such as histone demethylase
activity be removed for SCHPO/Epe1?

Existing workflow

• Analysis directory: genes/SCHPO/Epe1/Epe1-bioinformatics
• Summary report: genes/SCHPO/Epe1/Epe1-bioinformatics/RESULTS.md
• Review file: genes/SCHPO/Epe1/Epe1-ai-review.yaml

Hypothesis search candidates can be listed with:

just gene-hypothesis-list SCHPO Epe1

The histone-demethylase function assignment hypothesis can be staged with:

just gene-hypothesis-research openscientist SCHPO Epe1 \
  --annotation-term-id GO:0032452 \
  --as-function-hypothesis \
  --dry-run \
  -- --param use_hypotheses=true

Expected output path for a real run:

genes/SCHPO/Epe1/Epe1-hypotheses/function-hypothesis-go-0032452/openscientist.md

Current findings

• Epe1 is a JmjC-domain protein, but the accessible literature and local
  bioinformatics report support a pseudo-demethylase interpretation.
• Existing review decisions remove JmjC-derived catalytic activity annotations
  and retain chromatin/heterochromatin regulatory annotations.
• This case is the biological pilot for treating a function assignment as an
  explicit OpenScientist hypothesis.

Caveats

• The Epe1 bioinformatics folder is older and less reusable than the pmp20
  workflow. Use pmp20-bioinformatics as the engineering template for new
  reproducible analyses.
• The OpenScientist report should be run blind to the existing local
  RESULTS.md; compare the report against the older bioinformatics result after
  generation.

Template For Future Cases

For each new case, add:
1. Case ID, gene, species, and curation question.
2. Workflow path and exact run commands.
3. Key result files (especially RESULTS.md).
4. Clear statement of how outcomes changed curation decisions.
5. Validation status after integrating results into *-ai-review.yaml.
6. If applicable, the gene-hypothesis-research command and output path for the
OpenScientist hypothesis report.

Slides

• Slides (Marp source: BIOINFORMATICS-slides.md) — AI generated