Ortholog Conjecture Project

Why it matters

• The ortholog conjecture (OC) posits that orthologs (separated by speciation) tend to retain more similar functions than paralogs (separated by duplication) at the same evolutionary distance.
• This assumption underpins many automated functional annotations and cross-species inference pipelines.
• The OC remains debated: different data types, bias controls, and evaluation metrics can yield opposite conclusions.
• Consequence: the OC is not just theoretical — it affects annotation propagation, evidence reliability, and the risk of over-annotation.

How orthology appears in GO annotation

GO annotations explicitly encode how an inference was made:

• ISO — evidence code for inferences from sequence orthology.
• Phylogenetic annotations (PAINT/PANTHER, IBA/IBD) — explicit ancestral inference with traceability to experimental evidence. [GO_REF:0000033]
• Automated ortholog transfers documented in GO_REFs:
  • OrthoMCL-based ISO transfers [GO_REF:0000101]
  • Ensembl Compara orthology transfers yielding IEA [GO_REF:0000107]

These mechanisms make the OC a practical concern, not just a theoretical one.

The approach (science-first)

• This project is about the science of orthology vs. functional divergence and how to measure it without circularity or annotation bias.
• It is not primarily a curation rulebook — repo-level guidelines cover that.
• GO context and evidence codes are inputs; the emphasis is a reproducible, unbiased analysis of functional conservation/divergence.
• Built on the AI gene review framework to assemble literature-backed cases and test how evidence filters change conclusions.

Project goals

• Assemble literature-backed examples of functional divergence among orthologs (neofunctionalization, subfunctionalization, domain loss, subcellular relocalization).
• Build an "open world" curation lens: treat missing annotations as unknown, not absence; separate lack of evidence from evidence of loss.
• Develop unbiased similarity metrics controlling for annotation bias (authorship, term frequency, propagated annotation bias); compare GO-based vs. expression/phenotype signals.
• Identify families where orthology-based transfer overstates conservation, and quantify how often under different evidence filters.

The debate: conflicting conclusions

Same question, opposite answers — depending on data and bias controls.

Open-world metrics (draft spec)

Inputs: ortholog pairs/groups (1:1 → 1:many → many:many); GO annotations w/ evidence codes & dates; optional expression/phenotype matrices.

Evidence tiers (configurable):

• experimental_only — EXP, IDA, IMP, IGI, IEP
• curated_plus — + IBA/IBD/IC/ISS/ISO
• all_including_iea — + electronic, to quantify propagation effects

Goal: avoid penalizing missing annotations (unknowns) while still capturing divergence signals.

Similarity metrics & bias controls

Open-world similarity metrics:

• Asymmetric containment |A ∩ B| / |A| — do not penalize missing in B.
• Symmetric overlap (Jaccard) |A ∩ B| / |A ∪ B| — conservative baseline.
• IC-weighted overlap — weight by information content to reduce generic-term bias.

Bias controls:

• Author/publication bias — downweight same-source annotations across pairs.
• Term-frequency bias — inverse frequency or ontology depth.
• Propagation bias — compare experimental_only vs. curated_plus vs. all_including_iea.

Reporting: separate MF/BP/CC; stratify by orthology class; show "delta" when ISO/IEA added.

Curated divergence cases (1/2)

Curated divergence cases (2/2)

Large-scale datasets to mine

• Yeast–human replacement assays: only ~47% of essential yeast genes are functionally replaceable by their human orthologs; replaceability clusters by biological module → a divergence filter for ortholog transfer. [PMID:25999509]
• Expressolog dataset (Arabidopsis vs. A. lyrata): one-to-one groups show higher expression correlation than one-to-many or many-to-many; includes gene-level validation via complementation. [PMID:27303025]

These provide ground-truth signals beyond GO for benchmarking metrics.

Challenges

• Circularity & annotation bias: GO is incomplete and biased by study design and annotation practice — negative results can be misleading. [PMID:22359495; PMID:23209392]
• Conflicting conclusions depend on data type, bias controls, and evaluation metrics (microarray vs. RNA-seq; GO vs. expression).
• Open-world problem: missing annotation ≠ absence of function; must separate lack of evidence from evidence of loss.
• Boundary cases (lineage-specific LOF: UOX, GULOP) show orthology does not guarantee retention — but are excluded from subtle-divergence metrics.

Conclusions, status & future directions

Status (last updated 2026-02-09):

• [x] Seed list of divergence examples with citations (v0.1)
• [ ] Define open-world metric set and data filters
• [ ] Select ortholog datasets and clades for pilot analysis

Workplan ahead:

1. Build ortholog sets across well-studied clades; tag dominant evidence codes (ISO/IEA vs. IBA/IBD vs. experimental).
2. Compute similarity with experimental-only, then add ISO/IBA to measure transfer impact.
3. Cross-check GO-based similarity against RNA-seq / phenotype signals.
4. Summarize divergence rates + a "bias checklist" for correct interpretation.