PANTHER IBA family review

Bottom line: every IBA annotation descends from a PAINT curator's IBD
judgment placed at an ancestral node of a PANTHER tree, so the place to test
an IBA is that node and the target's position below it. We rebuilt the
propagation behind all 160 IBAs on the 41 reviewed S. pombe genes (36 of
which carry IBAs) from cached repo data: source node, seed genes, subfamilies,
PAINT loss annotations, and our per-gene action. We did this to check whether
the per-gene calls hold up at the family level, and to find the patterns that
mark a real over-propagation. They held up. The per-gene reviews kept 148 of
the 160 IBAs (117 ACCEPT, 31 KEEP_AS_NON_CORE); the 36 cross-subfamily flags
turned out to be mostly conserved functions; and the family lens confirmed the
two localization REMOVEs (pom1 cytoskeleton, rqh1 cytoplasm) and recast the
third REMOVE (mid1 septin ring organization) as sub-functionalization between
the two pombe anillins. No new IBA errors were found among the accepted rows.

The same tooling also extracts PAINT's own loss annotations (IRD/IKR) as a
curation guard: 2,129 loss findings across 549 cached families (2,123 paired with a confirmed ancestral gain), of which 63 IKR losses fall on a
reviewed member and are ready for residue-level follow-up. The written review
is in REVIEW.md.

The rest of this page documents the scripts and tables.

• extract_iba_propagation.py — reproducible extractor: for each IBA, resolves
  the ancestral PANTHER node, the seed genes, and the subfamilies of our gene
  and its seeds (from the cached interpro/panther/<FAM>/ tables); flags
  cross-subfamily and localization propagations and joins our curation action.
• iba_propagation.tsv — the resulting per-IBA table (regenerate with the script).
  Carries inline node-level (PAINT) columns joined from IBD.gaf:
  node_seed_count (the authoritative canonical seed count curated at the source
  node, vs. the few n_seeds echoed into the leaf), node_evidence
  (IBD/IRD/IKR), and node_loss. New flags: SINGLE_NODE_SEED (≤1 canonical
  seed — a provenance count, not a measure of evidential strength), NODE_LOSS (an IRD/IKR loss at the source node), and
  NODE_NOT_IN_IBD; NONE denotes a row with no propagation flags.
• extract_node_annotations.py — pulls the PTN node-level (PAINT) annotations
  themselves from PANTHER's IBD.gaf (the IBD/IRD/IKR — plus a few IBA-on-node —
  layer that is the source of every IBA). For each ancestral node our genes
  derive from, it lists the full
  node annotation set with the canonical seed counts, marks loss (IRD/IKR NOT)
  annotations, and flags node annotations that did not propagate to our gene
  (propagated=false → candidate missing annotation or lineage loss).
• node_annotations.tsv — the resulting per-node-annotation table.
• extract_function_losses.py — flags families where a subfamily lost a
  function: pairs every PAINT loss (any NOT annotation — usually IRD/IKR,
  occasionally NOT|IBD) with its ancestral gain node
  (IBD), checks the loss is within a cached family's PTN node set, and attributes
  it to the family subfamilies whose members descend from the loss node
  (resolved from the leaf GAF + member tables). This is the strongest
  within-family neo-/sub-functionalization signal and a direct curation guard:
  do not propagate the ancestral function to members under the loss node.
• family_function_losses.tsv — per-(family, lost-GO) findings with
  ancestral_node, loss_node, gain_confirmed, and the affected
  subfamilies. (n_members_affected=0 means the loss is in an unsampled
  subfamily — the gain→loss pair is still confirmed.)
• prepare_loss_analysis.py — turns one loss finding into a ready-to-run input
  bundle for a bioinformatics agent to reconstruct the key-residue rationale
  (PANTHER never publishes which residues an IKR was based on — see the IKR note
  below). Emits YAML with seed_uniprot (where the function + its key residues
  are characterized), loss_clade (members predicted to have lost it), and
  retaining_clade (members that kept it). The agent fetches these sequences,
  aligns them, and compares the functional columns. Example:

bash uv run python projects/PANTHER_IBA_REVIEW/prepare_loss_analysis.py \ --family PTHR10443 --loss-node PTN000047776 --go GO:0016805

Note: loss_clade/retaining_clade are resolved from the reviewed member
tables + leaf GAF, so a finding with n_members_affected=0 yields an empty
loss_clade (the loss is in an unsampled subfamily); seeds are still provided.
Of the 403 IKR findings, 63 have ≥1 attributed reviewed member and are
immediately actionable.
- REVIEW.md — the written review and findings.

Regenerate:

just refresh-panther-iba-project

The three tables can also be refreshed independently with
just refresh-panther-iba-propagation,
just refresh-panther-iba-node-annotations, and
just refresh-panther-iba-function-losses.

The node-level source files (IBD.gaf, leaf GAF) are downloaded on demand into
a gitignored .cache/panther/ and are not committed. Per-family node slices can
be materialised under interpro/panther/<FAM>/<FAM>-paint.tsv with:

just fetch-panther-paint PTHR10177

Scope: the 160 IBAs in the 41 reviewed genes (39 PANTHER families, all cached
locally). Note the cross-subfamily flag is deliberately sensitive and
over-fires on broadly conserved functions — it is triage, not a verdict.
One well-characterized descendant can soundly ground an ancestral assertion.
Review its phylogenetic placement and relevant functional divergence; do not
infer weak support from a short seed list.

Slides

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