Genome-wide validation

Bottom line: Per-gene review asks whether one annotation is right; genome-wide
validation asks whether the whole set of functions annotated to an organism could
coexist in a living cell. Following the completeness / coherence / consistency
framework of Tawfiq, Kulmanov & Hoehndorf (2026), we score a genome's annotation
set against constraints already encoded in GO (essential functions present,
has_part dependencies satisfied, taxon constraints respected) and turn each
violation into a reviewable curation lead. So far only coherence has been built:
the E. coli pilot scored the EcoCyc
GAF against the 743 asserted has_part pairs in GO and found 17 of 129 activated
dependencies unsatisfied (86.8% coherence). The 17 include one plausible biological
gap (a denitrification pathway with no nitrous-oxide reductase), several
granularity gaps, and probable over-annotations such as viral and heterochromatin
terms on E. coli. Completeness, consistency, the minimal-genome pilot and the
predictor sweep are not started.

We are doing this because computational predictors can be accurate protein by
protein yet produce a genome that no viable organism could have, and a cheap
set-level screen would catch that across all the prediction sets this repo reviews.

Key ideas

• A new validation altitude. Existing repo QC works one gene, one term at a time. This
  project validates the annotation set of an entire genome/proteome as a system, catching
  errors that are invisible per-protein (a pathway missing its committed step; a function that
  can't belong to this organism's lineage).
• Constraints already encoded in GO, not hand-authored rules. Completeness draws on
  minimal-genome essential-function sets; coherence reuses GO has_part (RO:0000051) axioms
  and MetaCyc pathway structure; consistency reuses GO taxon constraints (only_in_taxon /
  never_in_taxon). The validator is a reasoner over existing axioms, so it scales across
  genomes without per-organism curation.
• Sharpest as a predictor QC. Curated model-organism annotations largely satisfy these
  constraints; genome-scale computational predictions systematically violate them. That makes
  this a natural, cheap screen over the prediction-set work already in this repo.
• Every violation is a lead, not just a score. A missing coherence dependency points at a
  specific unannotated step; a taxon-constraint violation flags a specific over-propagated
  annotation — output a reviewer can act on, in the same spirit as
  Pathway satisfiability.

Why genome-scale

GO annotation, prediction, and evaluation are overwhelmingly reductionist: they assign and
score functions one protein at a time, treating each annotation as an independent fact. But an
organism is not a random bag of proteins — it is an evolved system with functional
requirements and constraints. A prediction method can be accurate per-protein yet produce a
genome-scale annotation set that could not describe any viable organism (missing DNA
replication; a late pathway step with no upstream enzymes; photosynthesis and neuron
development in the same genome).

This gap is the motivation of Tawfiq, Kulmanov & Hoehndorf (2026), whose completeness /
coherence / consistency framework (GAEF) we adopt and extend here, and it is the same gap that
Pathway satisfiability attacks on the eukaryotic, context-resolved
side. Genome-wide validation is the umbrella that generalizes both to a genome-scale QC pass.

The three criteria

Criterion	Question	Axiom source	A violation means
Completeness	Are functions essential for life present?	minimal-genome essential-function set (mapped to GO)	the genome lacks a process no viable cell can omit — likely an annotation gap
Coherence	Are functional dependencies satisfied?	GO has_part (RO:0000051) + MetaCyc pathway structure	a step/function is annotated but a required part/precursor is annotated nowhere in the genome
Consistency	Are mutually exclusive functions absent together?	GO in_taxon / only_in_taxon / never_in_taxon constraints	two functions restricted to disjoint lineages are annotated to one genome — an over-propagation

Each has a protein-level form (both functions on one protein) and a genome-level form
(the dependent/exclusive function anywhere in the proteome); GO's true-path rule already
enforces the protein-level coherence, so the novel signal is genome-level.

Presence ≠ flux / capacity. As in Pathway satisfiability, presence is used asymmetrically:
absence of a required part is strong evidence of a gap; presence only permits a route. A
"complete and coherent" genome is a necessary, not sufficient, condition for viability.

How it connects to existing work

This project is an umbrella; it should reuse and link, not re-implement:

• Pathway satisfiability — the module-logic engine
  (src/ai_gene_review/module_logic.py) that already tests pathway coherence and emits
  abduction leads. Genome-wide coherence is this engine run over all pathway/has_part
  dependencies for a genome rather than a single curated module.
• Metabolic Model Analysis — GEM gene-protein-reaction
  associations as an independent EC/MF validation source; a complementary, quantitative
  coherence check.
• Validating E. coli predictions — the per-prediction
  validity taxonomy this framework screens in bulk at genome scale.
• Reactome gap filling — pathway-hole reasoning that overlaps with
  the coherence "missing required part" signal.

Pilots

Start bacterial, where the constraints are strongest and the gold standard is cleanest; the
engine is organism-agnostic.

1. E. coli K-12 (gold standard) — Pilot 1, first result.
   Coherence implemented end-to-end on the EcoCyc GAF using GO has_part axioms:
   86.8% coherence, and the 17 violations triage into a genuine biological gap
   (denitrification lacking nitrous-oxide reductase), annotation-granularity gaps (complex /
   MF sub-terms), and probable over-annotations (eukaryote/viral terms on E. coli) — every one
   a reviewable lead. Runs from public data; see the pilot README and its RESULTS.md.
2. A minimal genome (JCVI syn3.0 / Mycoplasma). Small, tractable, and the natural source
   of the completeness essential-function set — a strong signal for completeness/coherence.
3. Predictor sweep. Run one or more genome-scale prediction sets (e.g. InterPro2GO,
   DeepGO-family) through the same criteria and quantify the curated-vs-predicted gap.

Plan (scoping)

• [x] Extract the dependency set (asserted GO has_part pairs) and score a genome's GAF for
  coherence — done in Pilot 1. Next: add
  ELK/relation-graph inferred pairs and the MetaCyc-derived pathway routes.
• [ ] Define the essential-function set from a minimal genome and map to GO classes.
• [ ] Score a whole genome's annotation set for completeness / consistency too (coherence done);
  emit per-genome scores + per-violation leads.
• [x] Run the E. coli coherence pilot; sanity-check against expectations (86.8% coherence,
  violations interpretable). Next: the minimal-genome pilot.
• [ ] Feed violations back as curation leads / prediction-review inputs.
• [ ] Decide whether to interoperate with or reproduce GAEF
  (github.com/bio-ontology-research-group/GAEF) rather than reimplement.

Open questions

• Axiom coverage. GO has_part between processes is sparse and unevenly curated — a
  coherence "failure" can be a real gap or a missing axiom. How do we separate the two
  (cross-check with sequence-level tools like GapMind / Pathway Tools before calling a gap)?
• Annotation gap vs biological gap. A set-based check never looks at sequence, so it cannot
  by itself distinguish "gene absent" from "gene present but unannotated." Which follow-up
  (BLAST/HMM step search) resolves flagged coherence gaps?
• Taxon scope. The three criteria are cleanest for free-living bacteria/archaea/eukaryotes;
  how far do they hold for obligate symbionts, reduced genomes, and community-dependent
  auxotrophs?

References

• Tawfiq R, Kulmanov M, Hoehndorf R. Evaluating completeness, coherence, and consistency of
  genome-scale function annotations. Briefings in Bioinformatics, 2026, 27(3):bbag336.
  doi:10.1093/bib/bbag336. Software (GAEF):
  https://github.com/bio-ontology-research-group/GAEF.

Slides

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