Zebrafish Genome Duplication and Paralog Pairs

Bottom line: Zebrafish keeps about 3,440 ohnolog pairs from the teleost-specific
genome duplication (TGD), roughly 320–350 Mya. This project will review such pairs
side by side and ask what happened to the ancestral function in each case:

Partition (subfunctionalization): the two copies split the ancestral functions.
Innovation (neofunctionalization): one copy gained a new function.
Backup (redundancy or dosage retention): both copies keep the same function.

The first phase is background research, summarized in
Background research.

Main finding so far: most retained pairs diverged in where and when they are
expressed rather than in what the protein does.
Clean cases are rare. Only a small minority are clear cases of
neofunctionalization or subfunctionalization, when judged against an unduplicated
outgroup (spotted gar).
Next step: choose pairs and run the first side-by-side review
(cryaba/cryabb, which are already both reviewed).

Motivation

A zebrafish gene often corresponds to one human gene through two "co-orthologs"
(the a/b pairs in ZFIN nomenclature). This matters for GO annotation in three
ways:

Experimental annotations are split between the copies. A zebrafish IMP or IDA
is made on one copy. Whether it transfers to the other copy, or to the human
ortholog, depends on whether the copies diverged in molecular function or only
in expression.
Phylogenetic propagation treats the copies as equivalent. IBA and ISO usually
give both copies the same terms. That is right for the backup and partition
fates, but may be wrong for neofunctionalized copies.
Mutant phenotypes can be masked. Zebrafish mutants can hide phenotypes when a
paralog is upregulated (transcriptional adaptation). This complicates any claim
that "gene X is not required for Y" or "X and Y are redundant".

Questions

For each pair, is the molecular function conserved in both copies? Or is there
protein-level divergence, such as lost catalytic residues, domain gain or loss,
or a changed interaction partner?
Are differences in biological process and location explained by expression
partition, as in the DDC model? Or by new expression domains absent from the
pre-duplication state (compared against gar or another non-teleost fish where
possible)?
Is there positive evidence of redundancy, such as double-mutant synergy,
cross-rescue, or paralog upregulation in mutants? Or is the "backup"
interpretation only an inference from a mild single-mutant phenotype?
Do existing GOA annotations (experimental, IBA, ISO, IEA) handle the pair
consistently? Are there annotations that should be copy-specific, or that should
be shared but are missing from one copy?

Approach

Pair selection criteria

TGD origin must be supported. Acceptable evidence is double-conserved synteny
(using the spotted gar "orthology bridge"), or a gene-tree duplication node at
the base of teleosts. Pairs from the older vertebrate duplications (e.g.
cyp26a1/cyp26b1) or younger tandem duplications are out of scope, or are used
only as contrasts.
Do not trust a/b suffixes. They are not a reliable guide to homeolog
identity (PMID:35961774).
Prefer informative pairs: pairs where at least one copy has experimental
annotations, or published comparative work on the pair (expression, rescue,
double mutants).
Mix the expected fates. Include pairs expected to be dosage or backup cases
(conserved enzymes), partition cases (developmental transcription factors), and
candidate neofunctionalization cases.

Per-pair workflow

just fetch-gene DANRE <gene> for both copies. Review each copy as usual.
Write a pair comparison note covering:
evidence of TGD origin;
protein-level comparison (domains, key residues, identity);
expression comparison (ZFIN, published in situs, PhyloFish or gar where
available);
mutant, morphant and compensation evidence;
annotation consistency across the pair.
Classify the pair as one of: partition / innovation / backup / dosage /
unresolved. Cite evidence for the classification.

Tools

scripts/panther_tgd_pairs.py finds zebrafish TGD pairs from PANTHER v19
gene trees. It uses PANTHER's AllParalogs and AllOrthologs downloads, plus
ZFIN and HGNC files for gene symbols; downloads are cached in .cache/, which
is gitignored. Outputs:
DANRE_DUPLICATION/panther_tgd_pairs.tsv: candidate pairs with the evidence
for each call.
DANRE_DUPLICATION/panther_paralog_branch_counts.tsv: zebrafish–zebrafish
paralog pairs per duplication branch.
scripts/check_quotes.py checks that every [PMID:N "quote"] in the project
notes is verbatim in the cached publication (ignoring whitespace differences).

How TGD pairs are called from PANTHER

PANTHER trees contain zebrafish, medaka, and spotted gar (the unduplicated
outgroup). For every paralog pair, PANTHER reports the tree branch that carries
the duplication, as parent|child. The calls, from strongest to weakest:

Call	Rule	1:1 pairs	All pairs
`TGD_tree`	Duplication on `Neopterygii\|Teleostei`: after the gar split, before the zebrafish–medaka split. This is the TGD	778	3,210
`TGD_tree_no_gar`	Duplication on another branch ending at `Teleostei` (mostly `Euteleostomi\|Teleostei`); no gar gene in the tree	93	610
`TGD_likely_parallel`	Duplication on `Teleostei\|DANRE`, but both copies share one gar co-ortholog, and their medaka co-orthologs were also duplicated (`Teleostei\|ORYLA`)	1,062	(1:1 only)
`TGD_or_lineage`	Duplication on `Teleostei\|DANRE`, one shared gar co-ortholog, a single medaka co-ortholog. Either TGD with loss of one medaka copy, or a zebrafish-only duplication	1,107	(1:1 only)
`unresolved`	Duplication on `Teleostei\|DANRE` without that gar/medaka pattern	363	(1:1 only)

1:1 means each gene has exactly one partner on that branch, i.e. a clean
ohnolog pair. From Teleostei|DANRE only 1:1 pairs are kept, because that branch
also holds large zebrafish-specific family expansions (40,822 pairs in all).

Why Teleostei|DANRE pairs are included. Many textbook TGD pairs are placed
there, not on the TGD branch: mitfa/b and pax6a/b come out as
TGD_likely_parallel, and sox9a/b as TGD_or_lineage. This happens because the
zebrafish and medaka copies do not group as ((zfA, medA), (zfB, medB)). That is
the gene-tree signature expected from lineage-specific resolution after delayed
rediploidization (PMID:35961774), or from poorly resolved trees. So strict
TGD_tree placement is conservative: it has high specificity and low
sensitivity.

Total. About 1,900 clean 1:1 pairs have tree-level support (TGD_tree,
TGD_tree_no_gar and TGD_likely_parallel), and 1,100 more are possible. That is
the same order as the 3,440 ohnolog pairs from synteny in Howe et al. 2013
(PMID:23594743).

Caveats

multi pairs on the TGD branch include some implausible ones. For example,
vegfaa is paired with vegfba and vegfbb, which come from an older vertebrate
duplication. Treat multi pairs as family-level signals and check the tree.
Some known pairs are absent. alcama/alcamb is placed at
Euteleostomi|DANRE, and fabp1b does not appear in the PANTHER tables. The
table is a PANTHER view, not a complete ohnolog catalogue.
PANTHER places duplications by tree reconciliation. Synteny (ZFIN/Ensembl, or
the gar bridge) remains the independent check for TGD_or_lineage and
unresolved pairs.

Candidate pairs

Pairs involving genes already reviewed in `genes/DANRE/`

From panther_tgd_pairs.tsv:

Pair	PANTHER call	Pair class	Partner reviewed?	Notes
cryaba / cryabb	TGD_or_lineage	1:1	yes	Pilot pair: both copies already reviewed; medaka has one co-ortholog
mfsd2aa / mfsd2ab	TGD_likely_parallel	1:1	no	LPC transporter (human MFSD2A)
flvcr2a / flvcr2b	TGD_likely_parallel	1:1	no	Human FLVCR2
hs6st3a / hs6st3b	TGD_likely_parallel	1:1	no	Heparan sulfate 6-O-sulfotransferase 3
he1.2 / he1.3	TGD_likely_parallel	1:1	no	Hatching enzymes; also he1.2 / npsn as TGD_tree (multi)
spns1 / spns3	TGD_tree_no_gar	1:1	no	Check: human has both SPNS1 and SPNS3, so this may be an older duplication
crppa / ispd	TGD_or_lineage	1:1	no	Check the symbols: ispd is a former name of CRPPA
rpe65a / rpe65b, rpe65c	TGD_tree	multi	no	Three copies
sult1st2 / sult1st5, sult1st6	TGD_tree_no_gar	multi	no	Family expansion; weak candidate

Pairs with published comparative work

These pairs come from the background research.

Pair	Published fate	PANTHER call	Reference
mitfa / mitfb	Expression partition; proteins interchangeable in rescue	TGD_likely_parallel	PMID:11543618
pax6a / pax6b	cis-regulatory subfunctionalization (pancreas enhancer kept only at pax6b)	TGD_likely_parallel	PMID:18282108
sox9a / sox9b	Subfunctionalization, partly lineage-specific	TGD_or_lineage	PMID:14579386
elna / elnb	Subfunctionalization then neofunctionalization (bulbus arteriosus)	unresolved (no gar eln in tree)	PMID:26783159
fabp1a / fabp1b	Hierarchical subfunctionalization of expression	not in PANTHER table	PMID:16857010
vegfaa / vegfab	Paralog upregulation in PTC mutants (compensation)	TGD_tree (multi)	PMID:30944477
hbegfa / hbegfb	Paralog upregulation in PTC mutants (compensation)	TGD_or_lineage	PMID:30944477
vcla / vclb	Paralog upregulation in PTC mutants (compensation)	TGD_tree	PMID:30944477
alcama / alcamb	Paralog upregulation in PTC mutants (compensation)	not called (Euteleostomi	DANRE)
gpr22a / gpr22b	Clear subfunctionalization (brain vs heart), judged against gar	TGD_likely_parallel	PMID:28944589
grk7a / grk7b	Not yet reviewed; cone photoreceptor kinases	TGD_tree

STATUS

Last updated: 2026-09-27

[x] Create project page and folder
[x] Background research on the TGD and fates of duplicates
(background; 36 publications
cached, 85 quotes verified)
[x] Genome-wide TGD pair calls from PANTHER v19 trees
(panther_tgd_pairs.tsv; replaces the earlier Ensembl Compara script)
[ ] Decide the pair-selection strategy: a targeted list, or a random sample of
1:1 TGD_tree / TGD_likely_parallel pairs from panther_tgd_pairs.tsv
[ ] Synteny check for TGD_or_lineage pairs chosen for review (including
cryaba/cryabb)
[x] Define a pair-comparison template (pairs/README.md)
and scripts/compare_pair.py
[x] Batch 1, literature-supported pairs, all reviewed with pair pages:
cryaba/cryabb, mitfa/mitfb, pax6a/pax6b, sox9a/sox9b, elna/elnb
[x] Accession audit (scripts/accession_audit.py, results in
accession_audit.md). 8 of 10 genes are on the
accession holding all their experimental GOA rows. pax6a is missing 9 ZFIN
experimental rows and pax6b 3 UniProt rows, which sit on sibling accessions.
[ ] Decide how to handle genes whose GOA annotations are split across UniProt
accessions (option: let fetch-gene merge GOA rows from secondary accessions)
[x] Batch 2, literature-supported pairs, all reviewed with pair pages:
fgf8a/fgf8b, tbx5a/tbx5b, col1a1a/col1a1b, grk7a/grk7b (grk7b updated), vcla/vclb
[x] Batch 3: random sample of 8 TGD_tree 1:1 pairs (seed 20260928;
scripts/sample_pairs.py, batch3_sample.tsv), all reviewed with pair pages
[x] Batch 4: random sample enlarged to 18 Compara-confirmed pairs (35 draws); all reviewed.
Tabulated in random_sample_summary.md
(scripts/tabulate_sample.py)
[ ] Report upstream: PANTHER TGD_tree false positives (e.g. avp/oxt, myh10/myh14),
PANTHER subfamily/ortholog errors (exoc3l2 as EXOC3L4, sh3glb2a in a drebrin-like
subfamily, COL3A1 for col1a1a/b), and ZFIN attribution issues (col1a1a IMP genotypes;
gad1 probe identity)
[ ] Decide whether to fix PANTHER-derived symbol gaps in panther_tgd_pairs.tsv
(e.g. tmc2a has no ZFIN id)

NOTES

2026-09-27

Project created. Background research compiled from 36 cached publications. Key
takeaways (details and quotes on the background page):
About 26% of zebrafish genes are TGD ohnologs (3,440 pairs; PMID:23594743).
Loss was rapid in the first ~60 My (PMID:26578810).
Using gar as the outgroup, clear neofunctionalization is 6.6% and clear
subfunctionalization 0.8% of pairs. About 20% keep near-identical expression;
about two-thirds are uncorrelated (PMID:28944589).
Regulatory divergence (93% of pairs) is far more common than protein-level
divergence (24%) (PMID:19439512).
Methodological flag. Transcriptional adaptation (PMID:30944477) means zebrafish
PTC mutants can hide phenotypes by upregulating the paralog. When reviewing IMP
annotations for pair members, check the allele type.
Ensembl Compara caveat. Most candidate pairs are placed at the
Osteoglossocephalai node rather than Clupeocephala. Both are plausibly TGD
nodes, but hes6/her13 and gpat3/agpat9l look like older or unrelated paralogs.
Synteny confirmation is needed before a pair is used.
Switched pair calling to PANTHER (v19 AllParalogs/AllOrthologs). The
trees include medaka and spotted gar, so the TGD branch
(Neopterygii|Teleostei) can be read directly. Result: 778 clean 1:1 pairs on
that branch, 93 more on teleost-stem branches in trees without gar, and 1,062
"parallel duplication" pairs. Most literature pairs (mitfa/b, pax6a/b, gpr22a/b)
fall in the last group, so strict branch placement alone would miss them. The
Ensembl Compara script and its table were removed.
Batch 1 results (details on each pair page).
Three pairs are PARTITION at the expression level with the protein function
conserved: mitfa/mitfb, sox9a/sox9b and pax6a/pax6b. For pax6 the partition
sits on a dose-sensitive shared core (the eye).
Two pairs are MIXED. elna/elnb adds protein-level innovation in elnb, the one
clear neofunctionalization case. In cryaba/cryabb both copies keep chaperone
activity at different strengths, with an uneven expression split.
No pair shows pure backup; overlap is limited to shared expression domains.
This matches the genome-wide picture: regulatory divergence is common and
protein divergence is rare.
Accession issue. For pax6a, ZFIN's experimental annotations sit on
RefSeq-derived TrEMBL accessions (e.g. A0A8M3AP00), not the Swiss-Prot entry
(P26630) that was reviewed. This makes pax6a look less studied than pax6b in
GOA. Audit the other genes before interpreting annotation gaps.
Deep research. falcon ran for mitfa, sox9a/b and elna/b. It failed for mitfb
and pax6a/b (Edison 402 Payment Required; OpenAI key invalid), so those
literature searches were done by hand, as recorded in the gene notes.

2026-09-28

Accession audit. For zebrafish, GOA spreads a gene's annotations over several
UniProt entries:
IBA and UniProt-curated rows go to the reference-proteome or Swiss-Prot entry.
ZFIN's experimental rows go to RefSeq-derived TrEMBL entries.
For pax6a these do not overlap. P26630 has the IBAs and UniProt's IMPs;
A0A8M9P6C7 (identical sequence) has ZFIN's 12 experimental rows. Switching
accession would only move the gap, so pax6a stays on P26630. The missing rows
are listed in the audit and on the pax6 pair page.
The 8 other genes are unaffected.
Batch 2 genes are audited before fetching, and each is fetched on the accession
carrying the most experimental rows.
Batch 2 results. Each gene was audited before fetching, and all ten were on
complete accessions. Deep research failed for all nine new genes (Edison 402), so
the literature searches were done by hand.
fgf8a/fgf8b: PARTITION, lopsided. fgf8a keeps most domains.
col1a1a/col1a1b: DOSAGE. alpha3(I) is an extra chain of the same trimer;
double heterozygotes are fragile.
tbx5a/tbx5b: MIXED. Both copies are essential; the fin role is split; protein
divergence is inferred.
grk7a/grk7b: MIXED. Asymmetric cone expression; grk7b is reportedly a slower
kinase.
vcla/vclb: MIXED. Conserved protein with an expression-biased partition.
Across both batches (10 pairs).
Clear protein-level innovation in only one pair (elnb). It is suspected but
unestablished in tbx5b and grk7b.
No pair behaves as pure backup. Where copies overlap, the double heterozygote
or double mutant shows a dosage requirement (pax6, col1a1, vcl).
The best-known "compensation" case (vcla to vclb) is mRNA-decay-triggered
transcriptional adaptation. No protein-level substitution has been shown.
The partition is usually lopsided: one copy keeps most ancestral domains
(fgf8a, tbx5a, cryabb, grk7a).
Curation findings worth reporting upstream.
Two col1a1a "skeletal system development" IMP rows (PMID:30082390) are backed
by col1a1b and col1a2 genotypes. This is a ZFIN gene-attribution error.
The PANTHER human-ortholog column lists COL3A1 for both col1a1 paralogs.
Batch 3 results (random sample, 8 pairs). Each pair page carries a machine-readable
Sample record line; the table is in the pairs index.

Fate	Pairs
PARTITION (expression)	tmc2a/tmc2b, lhfpl5a/lhfpl5b
BACKUP (provisional)	smad3a/smad3b
INNOVATION	none
UNRESOLVED	eef1da/eef1db, abi1a/abi1b, olfm3a/olfm3b, si:dkey-283b1.7/vwc2, magi3b/wu:fi36a10

Evidence is scarce for a typical pair. Only 4 of 8 pairs have experimental
data on both copies; 3 have expression data only. The literature-chosen batches
(1-2) therefore give a much richer picture than a typical TGD pair supports.
PANTHER placement is not sufficient on its own. For 2 of 8 pairs
(si:dkey-283b1.7/vwc2, magi3b/wu:fi36a10), Ensembl Compara and gar synteny do not
support TGD origin: the duplication is placed deeper, or the partner is a
divergent family member. For olfm3 the two sources disagree, although synteny
supports duplicated segments. Pair selection should add a synteny or Compara
check to the TGD_tree call.
Where there is evidence, the pattern matches batches 1-2. Proteins are
conserved and the copies differ in where they are expressed (tmc2, lhfpl5; also
the RNA-seq-only splits in abi1 and olfm3). The one backup call (smad3) is
provisional: the single-mutant data were not accessible.
Caution. With n=8, none of these proportions is precise, and the absence of
innovation in the sample is not evidence that it is rare beyond what the
genome-wide studies already show (about 7% of pairs against gar, PMID:28944589).
Batch 4 and the confirmed random sample (18 pairs). Details are in
random_sample_summary.md.
PANTHER placement alone is unreliable. Of 34 drawn pairs with annotations, only 18
(53%) were confirmed by Ensembl Compara. Rejected pairs include ancient paralogs such
as avp/oxt, myh10/myh14 and hnrnpm/nucleolin.
Fates: UNRESOLVED 12, PARTITION 4, BACKUP 1 (smad3, provisional), MIXED 1 (vcl),
INNOVATION 0.
- Three partitions are at the expression level (tmc2, lhfpl5, prom1).
- One is at the protein level: agxt split the ancestral dual peroxisomal and
  mitochondrial targeting signals between the copies, inferred from sequence.
Evidence depth is the limiting factor. Only 7 of 18 pairs have experimental data
on both copies; 8 have expression data only, and 1 sequence only. Most UNRESOLVED
calls reflect missing data, not ambiguous data.
Where divergence is seen, it is mostly in expression (11 pairs), often a
maternal-versus-zygotic or broad-versus-restricted split with one copy gar-like.
Protein-level signals are rare and mostly unverified:
- agxt targeting signals;
- guca1c/d Ca2+ sensitivity;
- gria1b regulatory serines;
- faster evolution of one copy in exoc3l2 and gria1.
No INNOVATION in 18 random pairs. By the rule of three this bounds its frequency
at roughly 17% or less (95%). That is consistent with the genome-wide estimate of
about 7% against gar (PMID:28944589).
Overall (literature-chosen batches 1-2 plus the random sample).
The typical retained TGD pair keeps the same molecular function in both copies and
differs in where and when each copy is expressed. The split is often lopsided, with
one copy keeping the ancestral, gar-like pattern.
Clear protein-level innovation was found only in the literature-chosen elnb.
Pure backup was not demonstrated for any pair.
For GO curation this means:
- MF annotations can usually be shared across a pair, while BP and CC need
  copy-specific evidence.
- IBA propagation of tissue-specific processes to the non-expressing copy is the most
  common error found. Examples: sox9a heart, pax6a pancreas, lhfpl5b hearing, prom1a
  photoreceptor differentiation, agxtb peroxisome.

Warnings (1)