PANTHER family assignment discrepancies

Bottom line: the same PANTHER release can put a protein in different
families depending on where you look. UniProt's DR PANTHER cross-references
come from InterPro's best-hit HMM match. PANTHER's own sequence classification
files (and the PAINT trees that IBA annotations come from) come from PANTHER's
own pipeline. When the two disagree, a module or review that takes its family
id from UniProt can be grounded on a family that has no PAINT tree for the
protein, while the family that does carry that tree goes unused. This project
records such cases. So far it holds one, found while building the phagocytosis
modules.

Why this matters here

Module grounding. FamilyDescriptor.term ids are checked against the
PANTHER member index (.cache/panther/panther-members-<release>.tsv, built by
just refresh-panther-members), which uses PANTHER's own classification first
and falls back to UniProt. A curator who copies the family id from a UniProt
entry may therefore get a family that differs from the one the validator
checks against.
IBA interpretation. PAINT nodes (PTNs) belong to PANTHER's own trees. If
the UniProt cross-reference names another family, the IBAs a protein carries
appear to come from a family it does not belong to.
Orthology claims. Family ids are used as evidence of orthology between
representative members. Two sources that disagree make that evidence
ambiguous.
Gene fetching. fetch-gene picks the PANTHER family to download from
the UniProt cross-reference (_extract_panther_family_id in
src/ai_gene_review/etl/gene.py). For a discrepant gene it caches family
data for the wrong family, and fetch-panther-paint then cannot build the
PAINT slice for the right one, because that needs <FAMILY>-entries.csv.

Case log

#	Proteins	UniProt / InterPro family	PANTHER 19.0 classification	PAINT	Status	Found in
1	C. elegans ced-1 (UniProtKB:Q9XWD6), human MEGF10 (UniProtKB:Q96KG7)	ced-1 → PTHR24043:SF8 (SCAVENGER RECEPTOR CLASS F / EGF-LIKE DOMAIN-CONTAINING PROTEIN); MEGF10 → PTHR24052:SF13 (DELTA-RELATED / MULTIPLE EGF LIKE DOMAINS 11)	Both in PTHR24035 (MULTIPLE EPIDERMAL GROWTH FACTOR-LIKE DOMAINS PROTEIN): ced-1 → SF109 PROTEIN DRAPER; MEGF10 → SF136	MEGF10 IBAs come from PTN002372116 and PTN009076277, nodes whose seeds include ced-1, Draper and mouse Megf10	Resolved in module: grounded on PTHR24035 with PAINT nodes; upstream discrepancy still open	phagocytic engulfment module

Case 1: CED-1 / Draper / MEGF10 engulfment receptors

CED-1 (C. elegans), Draper (Drosophila) and MEGF10 (vertebrates) are the
standard orthologous engulfment receptors for apoptotic cells. They are
multiple-EGF-like-domain transmembrane proteins with an EMI domain and NPxY /
YxxL tail motifs.

What UniProt says (entry version 168 for Q9XWD6 and 173 for Q96KG7,
2026-06-10; InterPro 110.0, PANTHER 19.0):

ced-1 → PTHR24043 SCAVENGER RECEPTOR CLASS F, subfamily PTHR24043:SF8.
The InterPro match covers only residues 685–963 of 1111, which lie in the
EGF-like-repeat region.
MEGF10 → PTHR24052 DELTA-RELATED, subfamily PTHR24052:SF13
("MULTIPLE EGF LIKE DOMAINS 11"). The InterPro match covers residues 9–675
of 1140.

What PANTHER's own classification says (PTHR19.0_<organism> sequence
classification files, cached under .cache/panther/ by
refresh-panther-members):

PTHR24035:SF109 PROTEIN DRAPER: ced-1 (Q9XWD6), Drosophila drpr (Q9W0A0),
and Dictyostelium sibA–sibD (Q54KF7, Q54JE1, Q54JA5, Q54JA4).
PTHR24035:SF136 MULTIPLE EPIDERMAL GROWTH FACTOR-LIKE DOMAINS PROTEIN 10:
MEGF10 orthologs from human, mouse, rat, cow, chicken, Xenopus and
zebrafish.

What PAINT says (QuickGO IBA rows for Q96KG7, WITH/FROM column):

GO:0043652 engulfment of apoptotic cell, from PTN002372116. Seeds:
FB:FBgn0027594 (drpr), MGI:2685177 (Megf10), WB:WBGene00000415 (ced-1).
GO:0005044 scavenger receptor activity, from PTN009076277. Seeds:
Q96KG7 and WB:WBGene00000415.

PAINT therefore treats CED-1, Draper and MEGF10 as a single clade inside
PANTHER's tree. This agrees with the PANTHER classification and the
literature, and disagrees with the UniProt cross-references.

InterPro's own family membership is consistent with its cross-references.
The member list InterPro returns for PTHR24052
(interpro/panther/PTHR24052/PTHR24052-entries.csv, fetched by fetch-gene)
places human, mouse and zebrafish MEGF10 in PTHR24052:SF13, together with
human MEGF11. The PTHR24035 member list contains no MEGF10. The disagreement
therefore covers MEGF10 across vertebrates, not just one entry, and it lies
between the InterPro and PANTHER assignment pipelines.

Caveats.

PTHR24035 is a broad family. Of its 11 human members, 9 sit in a
laminin-dominated subfamily (SF127 LAMININ SUBUNIT ALPHA-5-RELATED, which
also holds MEGF11) and MEGF6 has its own subfamily (SF137). A family-level
id alone therefore says little. No single subfamily contains both
representatives.
The name of the UniProt-assigned subfamily (PTHR24052:SF13, "MULTIPLE EGF
LIKE DOMAINS 11") does not match PANTHER's own placement of human MEGF11,
which is PTHR24035:SF127. In the cached classification files the only
member of PTHR24052:SF13 is bovine MEGF11 (F1MRL6). MEGF11 may be a second
case; it has not been checked.
The likely mechanism is an assumption, not a verified fact. InterPro keeps
the single best-scoring family HMM, and EGF-repeat-rich proteins score well
against several neighbouring EGF-rich families. This needs confirming, for
example by comparing full HMM score tables or asking the PANTHER and
InterPro teams.

Current handling (resolved locally, 2026-10-03).

fetch-gene human MEGF10 followed the UniProt cross-reference and cached
interpro/panther/PTHR24052/.
The PTHR24035 family data was fetched explicitly with the same helper
(_fetch_panther_family_data). fetch-panther-paint PTHR24035 then produced
interpro/panther/PTHR24035/PTHR24035-paint.tsv. That slice contains
PTN002372116 (P:GO:0043652, seeds drpr, Megf10 and ced-1) and PTN009076277
(F:GO:0005044 scavenger receptor activity, seeds MEGF10 and ced-1;
C:GO:0005886).
modules/phagocytic_engulfment.yaml now grounds the CED-1/MEGF10 annoton on
PTHR24035, with both nodes as ancestral nodes and scavenger receptor
activity as its function.
The new MEGF10 gene review (genes/human/MEGF10/) accepts the IBAs from
both nodes and records the family question under suggested_questions.

Reproduce

# UniProt cross-reference (InterPro-derived)
curl -s https://rest.uniprot.org/uniprotkb/Q96KG7.txt | grep '^DR   PANTHER'
# InterPro match coordinates
curl -s https://www.ebi.ac.uk/interpro/api/entry/panther/protein/uniprot/Q96KG7
# PANTHER's own classification (after `just refresh-panther-members`)
grep -h Q96KG7 .cache/panther/PTHR19.0_human
# PAINT provenance of the IBAs
curl -s -H 'Accept: application/json' \
  'https://www.ebi.ac.uk/QuickGO/services/annotation/search?geneProductId=UniProtKB:Q96KG7&evidenceCode=ECO:0000318'

Next steps

[x] MEGF10 gene review.
[x] ced-1 gene review (genes/worm/ced-1/). Note that fetch-gene cached PTHR24043, the UniProt-assigned family, for ced-1 as well.
[ ] Check MEGF11 (possible case 2).
[ ] Triage the systematic scan. Since the member index moved to
.cache/panther/, refresh-panther-members records UniProt's family
next to PANTHER's (uniprot_panther_family_sf column) and reports
disagreements. A run on 2026-10-05 over every accession cited in
modules/ and family reviews found 137 disagreements: 119 at the
family level and 18 at the subfamily level only. MEGF10 (Q96KG7) and
ced-1 (Q9XWD6) are among them. Remaining tooling question: should
fetch-gene prefer the PANTHER classification over the UniProt
cross-reference when choosing which family to cache?
[ ] Decide whether disagreements of this kind should be reported to
PANTHER/InterPro.