Human genes re-review

A second reader for every curation action in 1,323 human reviews

AI Gene Review · projects/HUMAN_GENES_RE_REVIEW · 2026

Bottom line

  • We re-read the action on every existing annotation in 1,323 human reviews (AAAS → ZSWIM8, ~55k actions) and asked "do I agree?"
  • ~99.7% agreement: only 3 actions changed (ABL1, ADRM1, BRCA2); all 143 NOT annotations adjudicated.
  • Sweep complete (July 2026). The human set has since grown to 2,048 files, so ~725 newer reviews have not had a second pass.

Why a second pass

  • Actions (ACCEPT, REMOVE, MODIFY …) are what summaries and modules are built on.
  • A single reviewer can be wrong in systematic ways: over-removal, removing experimental rows it cannot see, mishandling NOT rows.
  • Rules: judge from the YAML and cached papers only; edit only on confident disagreement; log everything else as a fence case or literature question.

What changed

The NOT audit

What the first pass got right

  • Pseudo-enzymes: ancestral activity removed when catalytic residues are lost (CRMP1, ILK, ROR1, UBAC2).
  • Wrong enzyme or direction: phosphatase ≠ kinase (CDC25B, PTPN6); TXN vs TXNRD; CPS1 vs CPS2.
  • Name collisions: HAP1 (huntingtin-associated ≠ APE1), GNAS locus, AARSD1 vs AARS1.
  • Legacy TF labels removed from ESCRT/Golgi machinery: TSG101, TMF1, USO1.

Flagged for a curator, not edited

Case Tension
GAPDH ~43 REMOVEs include documented moonlighting (GAIT complex)
HSPA1B removes own IDA chaperone rows that twin HSPA1A keeps
ATP23 DSB-repair rows from its former KUB3 identity removed
AGO3 / AGR2 NOT rows removed as superseded
ASCL1, ATF3 REMOVE on IDA; need the source papers

Status and next steps

  • ✅ A→Z sweep over 1,323 files; 3 edits, all present in the current YAMLs.
  • ✅ TRA2B, flagged as never reviewed, has since been reviewed (no PENDING rows).
  • ⬜ Resolve the fence cases and the ASCL1 / ATF3 literature questions.
  • ⬜ Second pass for the ~725 human reviews added after July 2026.

Read more: projects/HUMAN_GENES_RE_REVIEW.md