What is IBA?

IBA = Inferred from Biological Aspect of Ancestor

• Part of the GO Consortium PAINT / PANTHER phylogenetic annotation pipeline
• Curators annotate characterized proteins, then place those functions on the ancestral node of a phylogenetic tree (an IBD — Inferred from Biological aspect of Descendant)
• Those ancestral annotations propagate down to all descendant leaves as IBA annotations
• The WITH/FROM column records the exact source proteins a function was transferred from

IBA is a powerful, scalable way to annotate uncharacterized orthologs — but propagation can carry function where it no longer belongs.

The problem: failure modes

A function correct for an ancestor may be wrong for a given leaf when:

1. Functional divergence — orthologs evolved different functions
2. Pseudo-enzymes — catalytic function lost despite retained domain fold
3. Context-specific function — function differs between organisms / tissues
4. Over-generalization — broad terms transferred when specific functions differ

Root-level annotations are especially risky: they propagate to every descendant, including subfamilies that have diverged.

Two directions of IBA quality

This project examines both ways IBA fails:

Over-annotation (IBA is wrong) — patterns 1–15, the bulk of this work.

Incompleteness (IBA under-calls established biology) — phylogenetic propagation is conservative by construction, so much experimentally defined function never reaches the leaf.

Quantified loss: 511 curated human core molecular functions across 423 genes have no IBA support at all (401 grounded by experimental/traceable evidence).

The discovery method for both: the AI gene review framework.

The approach

Findings emerged from AI-assisted gene review, then verified against primary evidence:

• Mined all genes/*/*/*-ai-review.yaml — 2,732 reviews; IBA actions distributed as
  ACCEPT 4651 · KEEP_AS_NON_CORE 943 · MODIFY 321 · MARK_AS_OVER_ANNOTATED 189 · REMOVE 190 · UNDECIDED 39 · NEW 33
• Each REMOVE candidate cross-checked against: UniProt FUNCTION/CAUTION, the GO term definition (QuickGO/OLS), GOA WITH/FROM provenance + PANTHER family composition, cached publications, and a reproducible MSA of catalytic residues
• Incompleteness quantified with a generic evidence-subtraction tool (ai-gene-review subtraction-report)

Finding 1: Pseudo-enzyme propagation

Catalytic activity transferred to proteins that lost the active site but kept the fold — the most defensible REMOVE class (deficiency documented in UniProt).

• Epe1 (pombe) — GO:0032452 histone demethylase: HVD (not HXD) motif, no detectable activity
• DPYSL2 / CRMP1 / DPYSL3 / DPYSL4 (human) — GO:0016812 metallo-hydrolase: UniProt CAUTION "Lacks most of the conserved residues … for binding the metal cofactor"; MSA confirms loss of catalytic Lys159
• AGO4 (human) — GO:0004521 RNA endonuclease: "Lacks endonuclease activity"; MSA shows tetrad substitutions D669G, H807R (AGO3 retains the tetrad)
• AKTIP — GO:0061631 E2: lacks catalytic Cys · CASP12 — "Inactive caspase-12" · HSP47/Serpinh1 (mouse) — non-inhibitory serpin

Finding 2: Directional & sign errors

When a subfamily evolved the opposite reaction or a family mixes opposite-sign members.

• Cds1 (M. tuberculosis, V. cholerae; PTHR10314 SF135) — GO:0019344 cysteine biosynthesis propagated from the family root, but Cds1 catalyzes cysteine catabolism (EC 4.4.1.1: L-Cys → H2S + pyruvate). The most severe error type — directionally opposite. SF135 has the longest branch length (0.528) and only ~24% identity to synthases. (PMID:34439535, PMID:34283874)
• BCL2 (human, mouse) — GO:0043065 positive regulation of apoptosis, but BCL2 is the prototypical anti-apoptotic guardian. The IBA node (PTN000135648) mixes pro-apoptotic BAX/BAK1 with anti-apoptotic members. (Caveat: a separate NAS annotation + context-dependent pro-apoptotic roles exist → "sign unreliable", not "impossible".)

Finding 3: WITH/FROM reveals mis-grouping

The single most useful diagnostic — reading the source proteins exposes the error directly.

Tier A — wrong family / over-broad superfamily:

• NTN1 / NTN3 (human) — secreted Netrins given POU-domain TF activity (GO:0000981 etc.); WITH/FROM = POU2F1, POU1F1, POU4F1, POU4F3
• NOTCH1 — GO:0007411 axon guidance from SLIT1/2/3
• IL23R — GO:0004925 prolactin-receptor activity from PRLR

Tier B — wrong paralog:

• ABRAXAS1 — spindle/MT terms trace to ABRAXAS2
• opa1 / eat-3 — GO:0016559 peroxisome fission on mitochondrial-fusion OPA1 (DRP1 branch does fission)

Finding 4: Substrate & sub-activity over-propagation

Family nodes that lump enzymes of different specificities leak substrate terms.

• AGK (human) — GO:0001729 ceramide kinase: three lines refute it ("only … monoacylglycerols and diacylglycerols … not ceramide and sphingosine", PMID:15939762; PMID:16269826). The dedicated ceramide kinase is CERK. PANTHER PTHR12358 mixes acylglycerol + sphingosine kinases.
• SAMD8/SMSr — GO:0033188 sphingomyelin synthase: actually makes ceramide phosphoethanolamine ("larger PC prevents an efficient fit")
• CPT1C — RecName "Palmitoyl thioesterase"; lost carnitine transferase activity

Sub-activity loss: CAPG caps but does not sever actin (PMID:1322908); CRYAA is a holdase not a foldase — IBA contradicts a curated NOT(refolding); worm hsp-12.3/hsp-12.6 have no chaperone activity (PMID:9744800).

Finding 5: Compartment & lineage errors

Compartment over-transfer — but read the GO hierarchy first (two-sided):

• Valid REMOVE (mutually-exclusive): nucleus on cytoplasmic PIWI/Argonaute (PIWIL1, prg-1, wago-1); nucleus on ER-kinase EIF2AK3 / BIRC6; PM on yeast SSB2/SSZ1
• Anti-pattern (do NOT flag): "cytoplasm" on a mitochondrial/ER/lysosomal protein — GO:0005737 subsumes those organelles (e.g. Aga/GLA, DHCR24, ISCA1)

Lineage-inappropriate (cross-kingdom) process transfer:

• TOLL9 (mosquito) — GO:0006954 inflammatory response from human TLR4 · ndhA/D/K (poplar) — aerobic respiration on chloroplast NDH · che-3 (worm) — cilium motility on non-motile sensory cilia · sta-2 (worm) — JAK-STAT (no JAK in worms)

Summary: over-annotation patterns

The other direction: IBA incompleteness

Phylogenetic propagation only transfers what a curated ancestor already carries, at the ancestor's granularity — so much established biology never reaches the leaf.

Using evidence-subtraction over 1015 reviewed human genes (with ontology closure, so an IBA call to a general parent still counts):

• 62% of annotation-grounded core_functions terms (4516 / 7278) would be lost if IBA were the only evidence
• Restricting to molecular function (excluding low-info binding): 511 core MFs across 423 genes have no IBA support, 401 experimentally grounded (IDA/IMP/IPI/EXP/TAS)

These sit in core_functions — the curator's distilled judgement of what the protein does — so they are central activities, not noise.

Incompleteness: two mechanisms

A. Activity absent from IBA entirely (no IBA touches that branch):

B. IBA stops at a general parent (loss of resolution):

Caveat: IBA isn't always the laggard — for PTEN, IBA carries the textbook PIP3 phosphatase; and IBA is the sole support for 66 human core MFs.

Lessons learned (from the project log)

The recurring failure mode is acting on one line of evidence. Flagging a curated IBA is a strong claim — treat it like one.

1. No single source is sufficient — synthesize. Need ≥2 independent lines before REMOVE.
2. Read the GO definition, not the label. Worst miss: ATP7B "copper ion import" — definition covers movement into an organelle, so the Golgi-loading exporter is fine.
3. "By similarity" is weak; direct experiment beats it. AGK ceramide was a propagated tag refuted by two papers.
4. Read the WITH/FROM column first — wrong family / wrong paralog is near-mechanical evidence.
5. Place both compartments in the GO hierarchy before a localization REMOVE.
6. Check taxon/lineage appropriateness for process terms; respect the curator — uncertain ≠ wrong (UNDECIDED, e.g. UQCRC1).

Recommendations for the GO / PAINT pipeline

1. Flag pseudo-enzyme candidates — enzyme domains missing catalytic residues
2. Flag neo-functionalized subfamilies — long branch lengths + different EC numbers
3. Validate annotations at the family root — root terms propagate everywhere; ensure truly universal
4. Read the WITH/FROM — wrong family or single paralog source is strong evidence of error
5. Distinguish sub-activities — capping≠severing, holdase≠foldase, slicer≠non-slicer
6. Don't inherit a paralog's compartment/complex/pathway
7. Synthesize multiple lines of evidence — never a single keyword
8. Treat an IBA-only leaf as likely under-annotated — prompt to seek the specific experimental activity