Genome-wide validation

Scoring a whole genome's annotation set for completeness, coherence and consistency

AI Gene Review · projects/GENOME_WIDE_VALIDATION · E. coli pilot rerun 2026-10-04

Bottom line

  • Per-gene review asks if one annotation is right; this asks if the whole set could coexist in a living cell.
  • Coherence is built: the E. coli EcoCyc GAF satisfies 112 of 129 activated GO has_part dependencies (86.8%).
  • The 17 violations are curation leads: likely pathway overreach, missing complex/MF terms, viral or heterochromatin over-annotations, and unresolved cases. Completeness and consistency are still roadmap items.

Why genome scale

  • Prediction and evaluation score one protein at a time, as if annotations were independent facts.
  • A method can be accurate per protein yet yield a genome no organism could have: no DNA replication, a late pathway step with no upstream enzymes, photosynthesis next to neuron development.
  • Framework: completeness / coherence / consistency (Tawfiq, Kulmanov & Hoehndorf 2026, Brief. Bioinform. bbag336), using constraints already in GO.

The coherence check

What the 17 violations are

Limits of the pilot

  • Asserted has_part only (743 pairs); the reference paper reports thousands of additional ELK-inferred pairs, so this is a lower bound.
  • Set-based, not sequence-based: a violation cannot tell "gene absent" from "gene present but unannotated". Confirm gaps with GapMind or Pathway Tools before any REMOVE.
  • The completeness probe in the output (DNA replication, transcription, translation present) is a sanity check, not the metric.

Status and next steps

  • Done: E. coli coherence pilot, from public data (pilot-ecoli/coherence_pilot.py).
  • Next: add inferred has_part pairs and MetaCyc routes.
  • Next: minimal-genome essential-function set for completeness; GO taxon constraints for consistency.
  • Next: predictor sweep (InterPro2GO, DeepGO-family); decide whether to reuse GAEF.

Read more: project page · E. coli pilot RESULTS.md