Turning Red to Green Repeatedly: Ancestral LWS Duplications Drive Convergent Visual Tuning in Teleost Fishes

ABSTRACT

Photopigments, formed by an opsin protein bound to a light‐sensitive chromophore, underlie vertebrate vision. Long‐wavelength–sensitive (LWS) opsins mediate red‐light detection, and most teleosts retain a single functional LWS1. A shorter‐shifted green‐sensitive paralog (LWS2) is found only in a few lineages. By mining teleost genomes and sequencing retinal transcriptomes, we identify elopomorphs as an additional lineage retaining LWS2 (alongside characins and osteoglossomorphs). Surprisingly, we reveal a previously overlooked, shorter‐shifted paralog, LWS3, restricted to gobies (Percomorpha) and arising from an ancestral duplication either in the teleost or in the percomorph ancestor. Structural modelling of twelve LWS opsins reveals convergent evolution at four key amino acid sites (214, 259, 261, 269) in the retinal‐binding pocket, indicating convergent substitutions in human MWS (an old‐world primate‐specific duplicate of the vertebrate LWS1), teleost LWS2, and goby LWS3 relative to red‐sensitive counterparts, consistent with repeated spectral shifts towards greener wavelengths. In several lineages, including characins, mormyrids, gobies, and primates, these shorter‐shifted LWS opsins have at least partially replaced the canonical green‐sensitive RH2‐based visual pigment. Retinal transcriptomes revealed variable lws3 expression across adult gobies, either complementing lws1 in most species, or dominating the visual opsin expression profile as in the sicydiines. In situ hybridisation in the whitebarred goby ( Amblygobius phalaena; Valenciennes, 1837) showed that lws1 and lws3 are expressed in separate double‐cone members, analogous to rh2 expression in other fish species. Together, these results show that repeated convergence towards shorter LWS sensitivity in vertebrate evolution involves coordinated changes at the molecular, regulatory, and functional levels, providing a striking example of multilevel sensory adaptation.

1. Introduction

Animals rely on vision for a variety of fundamental tasks, including foraging, mating, avoiding predators and navigating their environment. At the molecular level, the visual process is initiated when light is absorbed by photopigments in the outer segments of retinal photoreceptors (Wald 1968; Hunt et al. 2014). Vertebrate photopigments consist of light‐sensitive vitamin‐A‐derived chromophores, which are covalently bound to one of five different visual opsin protein types. These five types were already present in the vertebrate ancestor and are thought to be the products of two rounds of whole‐genome duplication (2R) ~500 million years ago (Mya) (Lamb 2013). In the retina, they are expressed in an ancestral set of photoreceptor types (PR0—PR4) (Baden et al. 2025). They are the dim‐light active rod opsin (rhodopsin; RH1, expressed in PR0) and four cone‐based photopigments [SWS1 (PR4), SWS2 (PR3), RH2 (PR2) and MWS/LWS (PR1)], which are active during bright light and used for colour and motion vision (Hunt et al. 2014). Their spectral sensitivity ranges are a remarkable example of niche partitioning, in which each cone opsin pigment occupies a constricted visual space, spanning from the ultraviolet (< 400 nm; SWS1) to the red end of the spectrum (> 600 nm; LWS) (Marshall et al. 2024).

The evolution of vertebrate visual opsins is characterised by changes in opsin gene numbers, gene structure and gene expression (Hunt et al. 2014; Lamb 2019). Unlike in terrestrial vertebrates, where the basic opsin setup was either maintained [e.g., birds (Borges et al. 2015), reptiles (Katti et al. 2019)] or reduced [e.g., mammals (Jacobs 2013)], opsin genes have continued to proliferate in teleost fishes (Musilova et al. 2021; Policarpo et al. 2025). With over 35,000 species, teleosts are not only the most species‐rich clade of vertebrates (Fricke et al. 2021) but also exhibit extreme ecological, physiological, and morphological diversity, having conquered almost every aquatic and some terrestrial habitats (Cortesi et al. 2026). It is these differences in photic environments and life‐history traits that are the primary drivers of opsin gene diversity and evolution in fishes (Carleton, Conte, et al. 2020; Carleton, Escobar‐Camacho, et al. 2020).

Opsin gene evolution affects teleost vision at all taxonomic levels, from species to families and entire classes (Cortesi et al. 2015; Lin et al. 2017; Musilova et al. 2019; reviewed in Carleton, Conte, et al. 2020; Carleton, Escobar‐Camacho, et al. 2020; Musilova et al. 2021). Novel opsins commonly arise through tandem duplication, but, like the ancestral five types, may also be the product of whole‐genome duplications (Musilova et al. 2021; Policarpo et al. 2025). Moreover, in some rare cases, they arise from retrotransposition (Mano et al. 1999; Ward et al. 2008). While most genes are lost shortly after duplication and disappear or become non‐functional pseudogenes (Walsh 2003), some persist by acquiring novel functions (Lynch and Conery 2000). The neofunctionalisation of visual opsins typically involves changes at ‘key’ amino acid sites in the retinal binding pocket that shift the spectral sensitivity of the photopigment and are often associated with the colonisation of novel light habitats or visual niches (Yokoyama 2008; Carleton, Conte, et al. 2020; Carleton, Escobar‐Camacho, et al. 2020). However, neofunctionalisation can also occur due to differences in gene expression with ontogeny (Carleton et al. 2008; Cortesi et al. 2016), location in the retina or other tissues (Davies et al. 2015; Morrow et al. 2017), or by changing non‐visual characteristics of the protein (e.g., Castiglione et al. 2018; Luk et al. 2016). Finally, teleost opsins are prone to gene conversion, a form of reticulate evolution whereby sequence information is unidirectionally exchanged between sister genes (Cortesi et al. 2015; Sandkam et al. 2017). Gene conversion homogenises paralogs, which may repair and, in some cases, even restore gene function (Cortesi et al. 2015).

The genomes of extant teleosts contain a median of 7 visual opsin genes: 1 rod opsin and 6 cone opsins (Musilova et al. 2019). Most of the cone opsin expansions can be attributed to more recent lineage or species‐specific tandem duplications of the rh2 gene, which is sensitive to the most common, blue to green light spectrum underwater (peak spectral sensitivity [λ _max] range ~450–540 nm) (Carleton, Conte, et al. 2020; Carleton, Escobar‐Camacho, et al. 2020; Musilova and Cortesi 2023). Several older tandem duplications have also persisted, including those in the neoteleost and percomorph violet‐to‐blue‐sensitive sws2 duplicates (Cortesi et al. 2015). Interestingly, at their origin ~320 Mya, teleosts underwent a third round of whole‐genome duplication (3R) (Ravi and Venkatesh 2018), and two extant opsin gene duplicates, one rh1 (Chen et al. 2018) and one lws (Liu et al. 2019), have been attributed to this event. The ancestral lws2 duplicate has long been overlooked but has recently been described from bony tongues (Osteglossiformes) and characins (Characiformes) (Liu et al. 2019; Escobar‐Camacho et al. 2020).

Most teleosts possess a single LWS visual pigment (LWS1), which is sensitive from the yellow to red spectrum (~560–620 nm λ _max) (Carleton, Conte, et al. 2020; Carleton, Escobar‐Camacho, et al. 2020; Schweikert et al. 2018), or they might have lost the gene altogether, as commonly seen in deep‐sea lineages (Musilova et al. 2019; Lupše et al. 2021). The second copy, LWS2, in extant species but likely already in the ancestor shortly after its duplication by 3R, has shifted its spectral sensitivity to shorter, greener wavelengths as measured by visual pigment in vitro reconstruction and microspectrophotometry (~501–530 nm λ _max) (Parry et al. 2003; Liu et al. 2019; Escobar‐Camacho et al. 2020). This is similar to the MWS/LWS duplicate in primates, which re‐enabled trichromatic vision after the loss of rh2 in the mammalian ancestor (Carvalho et al. 2017). Concurrently, both bony tongues and characins have either lost or severely reduced rh2 expression, suggesting a trade‐off between pigment spectral overlap and the potential for evolutionary retention (Liu et al. 2019; Escobar‐Camacho et al. 2020). Interestingly, a shift towards shorter spectral sensitivities can also be observed in some more recent LWS1 duplicates in several fish species, as well as between different alleles of the same gene (Seehausen et al. 2008; Sandkam et al. 2018). Given the repeated pattern of lws duplication followed by occupation of the green‐light niche, it is surprising that teleosts only possess a limited number of these genes. Moreover, the convergence of green‐sensitive LWS visual pigments across distantly related vertebrate groups presents a unique opportunity to examine how evolution repeats itself to achieve similar visual functions.

Against this background, we explore the evolutionary history of the lws opsins in teleost fishes from a phylogenetic representative sampling of more than 100 species. By combining next‐generation genome and transcriptome sequencing with open‐source data mining, we uncover a rich history of gene duplication, loss and gene conversion. Among these, we find two more teleost groups: tarpons (Elopomorpha), which also retain the lws2 duplicate, and gobies (Percomorpha), which have a third lws copy (labelled here as lws3) of ancestral origin. We provide evidence that the lws3 copy is functional, expressed, and used as a visual opsin in double cones within the retinal mosaic. We then focus on the (re)evolution of shorter‐shifted spectral sensitivities in teleost LWS2 and LWS3 paralogs. Finally, we investigate parallel protein‐level signatures of a blue shift (i.e., becoming more green‐wavelength sensitive) between the teleost LWS2/LWS3 visual pigments and the human MWS, uncovering remarkable signatures of repeated amino acid substitutions and functional convergence in gene expression patterns across the vertebrate tree of life.

2. Materials and Methods

2.1. Ethical Approval and Relevant Permits

All the research complies with applicable laws on sampling from natural populations and animal experimentation (including the ARRIVE guidelines). Experimental procedures involving live animals within Australia were approved by the University of Queensland Animal Ethics Committee (QBI/304/16). Fish were collected under the Great Barrier Reef Marine Parks Permit (G17/38160.1) and Queensland general fisheries permits (207976 and 199,434). Sample collection in the Czech Republic was conducted under a permit issued by the Czech Fisheries Association. Specimens collected for retinal transcriptomics were euthanised using MS222 or an ice slurry, in accordance with best practice standards.

2.2. Taxon Sampling

Samples of the whitebarred goby ( Amblygobius phalaena ; Valenciennes, 1837) and the sling‐jaw wrasse ( Epibulus insidiator ; Pallas, 1770) for gene expression analysis were previously published in Musilova et al. (2019) (see Table S1 for the sample list and Accession numbers). The sleepy cod ( Oxyeleotris lineolata ; Steindachner, 1867) (n = 4 sub‐adults) was collected from the Dawson River at Taroom, Queensland, Australia, in May 2020. A. phalaena specimens (n = 3 adults) for in situ hybridisation were collected on snorkel from shallow reefs (1–2 m) surrounding Lizard Island (14°40′ S, 145°27′ E), Great Barrier Reef, Australia, in March 2021 using an anaesthetic clove oil solution (10% clove oil, 40% ethanol, and 50% seawater) and hand nets. Eyes were enucleated, and the retinas were dissected out and either fixed in 4% PFA for in situ hybridisation or stored in RNAlater at −20°C until further processing for transcriptome sequencing. Samples of the round goby ( Neogobius melanostomus ; Pallas, 1814) were collected in the Elbe River (Děčín, Czech Republic) using an electrofishing device. Samples of the remaining seven species: spinycheek sleeper ( Eleotris pisonis ; Gmelin, 1789), marble goby ( Oxyeleotris marmorata ; Bleeker, 1852), peacock gudgeon ( Tateurndina ocellicauda ; Nichols, 1955), rainbow goby ( Stiphodon ornatus ; Meinken, 1974), red lipstick goby (Sicyopus rubicundus; Keith, Hadiaty, Busson & Hubert, 2014), bumblebee fish ( Brachygobius xanthozonus ; Bleeker, 1849), and knight goby ( Stigmatogobius sadanundio ; Hamilton, 1822) were obtained from the aquarium trade.

2.3. Transcriptome Sequencing

Retinal RNA was extracted using a Qiagen mini kit (https://www.qiagen.com) column extraction following the manufacturer's protocol, including a DNase step. Total RNA concentrations were measured on a fluorometer (Qubit, ThermoFisher), and integrity was verified using a Pico chip on an Agilent 2100 Bioanalyzer (https://www.agilent.com). Library preparation (unstranded, 300 bp insert) was conducted using an Illumina NEBNext Ultra II Directional RNA library preparation kit, NEBNext Multiplex Oligos and the NEBNext Poly(A) mRNA Magnetic Isolation Module (New England Biolabs) for all osteoglossiform, tarpon and goby species used for transcriptomics except for O. lineolata, N. melanostomus and A. phalaena , whose library preparations were outsourced to Novogene, Singapore (https://en.novogene.com/). Transcriptome sequencing (RNAseq, HiSeq, Paired‐End PE150) was outsourced to Novogene for all species.

2.4. Assembly and Gene Extraction

Transcriptome raw reads were filtered following the protocols outlined in Musilova et al. 2019, and mapped with Medium sensitivity settings against a general goby reference (based on the genome of N. melanostomus ; Adrian‐Kalchauser et al., 2020) in Geneious 9.0 (https://www.geneious.com). Species‐specific gene sequences were retrieved from the mapped alignments and used as a species‐specific reference for another round of mapping with the Lowest sensitivity settings. The number of reads was calculated as FPKM (fragments per kilobase per million reads), and the proportion of each opsin relative to total cone opsin gene expression was calculated following the protocols of de Busserolles et al. (2017), Musilova et al. (2019), and Tettamanti et al. (2019).

2.5. Whole Genome lws Gene Mining and Synteny Analysis

High‐quality teleost whole‐genome sequences were downloaded from GenBank (accession numbers as listed in Table S1). Most of the high‐quality genomes were sequenced by the Sanger Institute (Rhie et al. 2021). We further complemented this data set with lower‐quality genomic data (draft genomes, as per Cortesi et al. 2015, and Musilova et al. 2019). Assembled genomic scaffolds or chromosomes were mapped against a single‐exon reference for lws1 ( Oreochromis niloticus ), lws2 ( Astyanax mexicanus ), and lws3 ( N. melanostomus ). Mapping was performed in Geneious 9.1.8 using Medium to Medium‐High sensitivity settings; these settings were extensively trialled to capture all scaffolds containing lws opsin gene sequences (higher sensitivity yielded false hits). Single lws genes were then extracted from the scaffolds either according to the annotation (if available) or by mapping the single exons to the scaffold with the Highest sensitivity setting. Next, the synteny of the lws copies (±10 flanking genes) and how they are conserved across teleosts was investigated in the annotated genomes from selected teleost representatives: tarpon ( M. cyprinoides ), elephant fish (P. kingsleyae), cavefish ( A. mexicanus ), and two goby species [round goby ( N. melanostomus ) and mudskipper (P. magnuspinnatus)]. Finally, we performed a VISTA plot analysis (Frazer et al. 2004) focusing on a larger region of 1–2 Mbp around the lws2 gene in the elephant fish and a cavefish compared to the two tarpon chromosomes carrying lws2 copies.

2.6. Phylogenetic Analysis

The coding sequences of the mined lws genes were aligned in Geneious 9.1.8 using the MAFFT Alignment plugin (Katoh et al. 2019) for 946‐bp‐long nucleotide sequences, excluding the short exons 1 and 6. Five independent runs (each with two parallel runs) were performed in MrBayes 3.1.2 (Ronquist and Huelsenbeck 2003) for 50 million generations. The output files were manually inspected in Tracer 1.5 (https://github.com/beast‐dev/tracer/releases) to ensure the runs reached the lnL plateau, identify the best ‐lnL score, and determine the burnin breakpoint. Three of the ten runs were then used to construct a consensus tree, with a burnin of 50%. We further analysed each of the four single‐exon datasets (2, 3, 4, and 5) separately, using the same analytical parameters as in the main analysis. We also performed the same analysis on the full nucleotide sequence after excluding the third codon position. To complement the Bayesian inference, we applied Maximum Likelihood reconstruction using the PHYML plugin in Geneious, with the GTR substitution model and the BEST topology search option (combining NNI and SPR algorithms). Branch support was assessed using 100 bootstrap replicates. Finally, we performed distance‐based Neighbour‐Joining analyses on both the full‐sequence and third‐codon‐position‐excluded datasets to obtain additional support values and compare the results with those from the Bayesian and Maximum Likelihood analyses.

2.7. Ancestral State Reconstruction

To assess the evolution of key‐tuning amino acid sites, we focused on three sites (164, 261, and 269) known to cause large shifts in the λ _max of LWS of 6, 10, and 16 nm, respectively (Yokoyama 2008). We reconstructed the ancestral states for each key site using the gene tree pruned to taxa with multiple LWS types: Elopomorpha+Osteoglossomorpha+Otomorpha with LWS1 and LWS2, and Gobiiformes with LWS1 and LWS3. Using amino acid variants that shift λ _max as character states and the equal transition rates (ER) model, we performed ancestral‐state analysis with the ace function in the APE package (Paradis and Schliep 2019). The ace function employs a maximum‐likelihood approach, in which the reconstructed ancestral states are presented as proportions of the total likelihood for each state at each node. The resulting trees were visualised and edited with the phytools package (Revell 2012).

2.8. Gene Conversion Analysis

GARD (Kosakovsky Pond et al. 2006) was run with general discrete site‐to‐site variation and three rate classes on the Datamonkey platform (www.datamonkey.org) (Weaver et al. 2018) using an alignment of lws1 and lws3 from goby species for which both copies were recovered. This was complemented by a sliding window analysis, following the methods described in Cortesi et al. 2015. In short, the intraspecific dS (neutral process) along the coding sequences of lws1 and lws3 was calculated using a window size of 30 and a step size of 1 in DNAsp v.6 (Rozas et al. 2017).

2.9. Regulatory Elements Analysis

The key regulatory elements controlling expression of the lws opsin gene were examined, including the locus control region (LCR), which is shared across vertebrates (Smallwood et al. 2002), and the associated miRNA (miR‐726) shared across fishes (O'Quin et al. 2011; Wakefield et al. 2008) (Table S2). In most fishes, the miRNA and LCR are within a 500 bp region. We aligned these regions from Oryzias latipes, Gasterosteus aculeatus, O. niloticus and Danio rerio, to identify the core LCR and miRNA regions. These were then blasted against the genomes of Astyanax, Paramormyrops, Megalops, and Neogobius to identify homologous regions upstream of the lws genes. Within the LCRs, we identified binding sites for key regulators of the lws gene (OTX2, THR or RAX; Ghinia Tegla et al. 2020; Emerson et al. 2013; Volkov et al. 2020; Mackin et al. 2019; Nelson et al. 2009; Furukawa et al. 1997) using JASPAR 2020 (Fornes et al. 2020). If there was a homologous sequence to miR‐726, we tested miRNA stability by calculating the miRNA folding energy with miRNAfold (https://evryrna.ibisc.univ‐evry.fr/miRNAFold). In this method, a more negative energy value indicates greater stability and thus a higher likelihood of being a functional miRNA (Table S2).

2.10. Spectral Sensitivity and Protein Structure Predictions, and Retinal Binding Pocket Analysis

We used the machine‐learning prediction tool OPTICS (Frazer and Oakley 2025) to perform spectral predictions (λ _max) for 15 goby species and three teleost species with lws2. The whole‐dataset parameters with and without the mine‐and‐match model (whole‐dataset, whole‐dataset‐mnm), and specifically considering amino‐acid properties (AA_prop), were used (Table 1).

TABLE 1.

Predicted LWS peak spectral sensitivities (λ _max) in gobies (LWS1, LWS3) and other teleosts (groups with LWS2).

Species	Gene	Five‐sites rule amino acids					λmax prediction—5‐sites rule	λmax prediction by OPTICS
		180	197	277	285	308		WDS‐mnm; AA prop; median + SD		par: WDS; AA prop; median + SD
	Bovine rhodopsin equivalent:	164	181	261	269	292
	Bovine rhodopsin amino acid:	A	E	F	A	A
Goby species with both types of LWS (LWS1, LWS3) found in the transcriptome/genome
Amblygobius phalaena	LWS1	S	H	Y	T	A	560 nm	560.3 nm	4.7	563.4 nm	10.8
	LWS3	A	H	F	T	A	546 nm	546.6 nm	3.4	548 nm	7.6
Boleophthalmus pectinirostris	LWS1	A	H	Y	T	A	553 nm	551.1 nm	4.3	551.5 nm	7
	LWS3	A	H	F	A	A	531 nm	531.4 nm	3.1	530.9 nm	5.8
Brachygobius xanthozona	LWS1	A	H	Y	T	A	553 nm	549.9 nm	4	551.9 nm	7.3
	LWS3	A	H	F	T	A	546 nm	543.9 nm	2.9	544.1 nm	6
Chaenogobius annularis	LWS1	S	H	Y	T	A	560 nm	560.4 nm	5.1	565.1 nm	11.3
	LWS3	S	H	F	T	A	551 nm‡	549.1 nm	4.4	555.7 nm	7.6
Neogobius melanostomus	LWS1	S	H	Y	T	A	560 nm	559.9 nm	4.9	562.7 nm	10.7
	LWS3	S	H	F	T	A	551 nm‡	544.9 nm	4.4	550.6 nm	6.8
Oxyeleotris lineolata	LWS1	A	H	Y	T	A	553 nm	552.8 nm	3.9	556.2 nm	8.8
	LWS3	A	H	F	T	A	546 nm	544.7 nm	4.8	542.4 nm	10.4
Periophthalmus magnuspinnatus	LWS1	S	H	Y	T	A	560 nm	559.6 nm	4.3	562.7 nm	8
	LWS3	A	H	F	T	A	546 nm	543.7 nm	4	543.4 nm	6.8
Stiphodon ornatus	LWS1	S	H	Y	T	A	560 nm	552.6 nm	5.4	558.6 nm	9
	LWS3	S	H	F	T	A	551 nm‡	551.5 nm	3.4	555.8 nm	8.8
Tateurndina ocellicauda	LWS1	A	H	Y	T	A	553 nm	552.9 nm	3.7	554.1 nm	8.4
	LWS3	S	H	F	T	A	551 nm‡	548.2 nm	3.5	552.7 nm	7.2
Goby species with only one type of LWS found in the transcriptome/genome
Butis butis (mislabelled as O. marmorata on GenBank)	LWS3‐1	A	H	F	T	A	546 nm	n/a	n/a	n/a	n/a
	LWS3‐2	P	H	F	T	A	546 nm	n/a	n/a	n/a	n/a
Eleotris pisonis	LWS1‐1	S	H	Y	T	A	560 nm	552.4 nm	4	557.4 nm	7.4
	LWS1‐2	S	H	Y	T	A	560 nm	552.5 nm	3.7	557.6 nm	7.1
	LWS1‐3	S	H	Y	T	A	560 nm	551.9 nm	3.7	557.1 nm	7
Oxyeleotris marmorata	LWS1	A	H	Y	T	A	553 nm	547 nm	3.4	552.3 nm	6.1
Syciopus rubicundus	LWS3	S	H	F	T	A	551 nm‡	550.8 nm	3.6	556 nm	9.2
Stigmatogobius sadanundio	LWS1	A	H	Y	T	A	553 nm	552.5 nm	3.8	556.9 nm	7.8
Lythrypnus dalli	LWS3	A	H	F	T	A	546 nm	n/a	n/a	n/a	n/a
Teleosts with LWS2 (Elopomorpha + Osteoglossomorpha + Characiformes) + Human
Megalops cyprinoides	LWS1a	S	H	F	T	A	544 nm	551 nm	1.6	551.6 nm	3.4
	LWS1b	S	H	Y	T	A	553 nm	561.1 nm	1.4	560.7 nm	3.4
	LWS2_715	A	H	F	A	A	528 nm	530.6 nm	3.4	530.3 nm	5.3
	LWS2_709	A	H	F	A	A	528 nm	514.4 nm	8.5	524.2 nm	8.1
Gymnarchus niloticus	LWS1	S	H	Y	T	A	553 nm	550.8 nm	3.4	554.7 nm	6.1
	LWS2	A	H	F	A	A	528 nm	529.3 nm	3.5	529.6 nm	6
Cyphocharax magdalenae	LWS1	S	H	Y	T	A	553 nm	553.3 nm	3.5	556.6 nm	7.5
	LWS2_1	A	H	F	A	A	528 nm	513.2 nm	8.7	524.9 nm	8.4
	LWS2_2	A	H	F	A	A	528 nm	528.6 nm	3.4	531.1 nm	3.9
Homo sapiens	LWS	A	H	Y	T	A	547 nm	557.9 nm	0.9	557 nm	1.1
	MWS	A	H	F	A	A	528 nm	529.7 nm	0.5	530.2 nm	0.5

Note: For the gobies, we started with the predicted value (551 nm λ _max) for a mudskipper LWS opsin (You et al. 2014), based on the five‐site rule with S164‐H181‐F261‐T269‐A292 (marked with ‡). We added +6 nm at 164, +9 nm at 261, and + 13 nm at 269 when a polar residue was present (Sekharan et al. 2012; Yokoyama 2008). For species with LWS2 (Megalops, characin, Gymnarchus), we started with the value from bovine rhodopsin ~500 nm (+28 nm for E181H). OPTICS (Frazer and Oakley 2025) was used for machine‐learning predictions on the entire dataset (mine‐and‐match or heterologous), considering amino acid properties (AA_prop parameter). Generally, the spectral sensitivities of LWS2 and LWS3 are predicted to shift towards greener wavelengths compared to LWS1. The bovine rhodopsin sites 164, 181, 261, 269, and 292 correspond to sites 180, 197, 277, 285, and 308 in the human LWS (Hagen et al. 2023).

Three‐dimensional structures of LWS opsins were predicted using Boltz (Passaro et al. 2025), a deep learning‐based protein structure prediction method. Predicted structures were converted from mmCIF to PDB format for downstream analysis. Bovine rhodopsin (PDB: 1GZM, chain A) and human OPN1LW (NM_020061) and OPN1MW (NM_000513) were included as structural references.

The retinal chromophore binding pocket was characterised by identifying residues within 6 Å of 11‐cis‐retinal and its covalently bound lysine. Residues were categorised into three distance shells from the retinal‐lysine complex: close (< 4 Å), medium (4–5 Å), and far (5–6 Å). Analysis was performed using PyMOL (Schrödinger LLC) with custom Python scripts.

Opsin sequences were aligned to establish positional homology across species. Binding pocket residues identified by distance criteria were mapped to alignment positions, enabling cross‐species comparison of equivalent sites. Residues at positions corresponding to the ‘five‐sites rule’ for LWS/MWS spectral tuning [positions 180, 197, 277, 285, and 308 in human opsin numbering (Yokoyama and Radlwimmer 1998)] were extracted from the binding pocket analysis (Tables 1 and 2).

TABLE 2.

Variable position in the retinal binding pocket.

Bovine rhodopsin position	λ max/nm	88	118	164	201	208	214	259	261	269	270	275	293
human_NM_020061_LWS_OPN1LW	547	V	S	A	Q	M	I	F	Y	T	F	A	Y
human_NM_000513_MWS_OPN1MW	528	V	S	A	Q	M	T	L	F	A	F	A	F
Aphalaena_FZ1_LWSx1	560	L	S	S	L	M	I	F	Y	T	F	A	Y
Aphalaena_FZ3_LWSx2	546	V	S	A	Q	M	L	L	F	T	V	S	Y
Stiphodon_ornatus_LWS01	560	V	A	S	Q	M	I	F	Y	T	F	A	Y
Stiphodon_ornatus_LWS02	551	V	S	S	K	M	L	V	F	T	C	A	Y
Cmagdalenae_MT310731	553	V	A	S	Q	M	I	G	Y	T	F	A	Y
Cmagdalenae_MT310746	528	V	S	A	A	L	L	I	F	A	S	S	Y
Gymnarchus_niloticus_119C1_LWS1	553	L	A	S	Q	M	I	I	Y	T	L	A	Y
Gymnarchus_niloticus_119C1_LWS2	528	V	S	A	K	M	L	V	F	A	S	S	Y
Megalops_cyprinoides_CM023712_LWS01A	544	L	S	S	Q	M	I	F	F	T	F	A	Y
Megalops_cyprinoides_CM023709_LWS2	528	V	S	A	K	L	L	I	F	A	S	S	Y

Note: The five‐site rule amino acid positions, 164, 261, and 269, are highlighted in yellow. Predictions for peak spectral sensitivity (λ _max) values using the five‐site rule are provided. Colours indicate different amino acids, with shading indicating similar physicochemical properties.

2.11. Fluorescent In Situ Hybridisation

Dual‐labelling RNA‐FISH was performed on two wholemounted retinas (left and right) from one adult A. phalaena collected at Lizard Island on 16 March 2021 following standard protocols (e.g., Barthel and Raymond 2000; Raymond and Barthel 2004; Allison et al. 2010; Dalton et al. 2014; Dalton et al. 2015). Retinas were prepared as follows. After one hour of dark adaptation, the eyes were enucleated, the vitreous was enzymatically digested, and the retinas were extracted from the eyecups. The retinal pigment epithelium layer (RPE) was removed by PBS‐jetting using a single‐use syringe and needle (The). Retinas were flattened and pinned, photoreceptor side down, in Sylgard‐filled petri dishes, and fixed in 4% PFA at 4°C overnight. Fixed retinas were triple‐washed in PBS for 10 min, briefly rinsed in 70% methanol in PBS, and then transferred to 100% methanol and stored at −20°C.

Previously extracted retinal mRNA was reverse transcribed using a High‐Capacity RNA‐to‐cDNA Reverse Transcription Kit (Applied Biosystems). Riboprobe templates were synthesised from cDNA via standard PCR using opsin‐specific primers (Table S3) and Q5 High Fidelity DNA polymerase (New England Biolabs). Primers were designed to bind to the coding sequence of target opsins (sws2b, sws2aβ, lws1, lws3) and to contain a T3 or T7 RNA polymerase promoter sequence at their 5′‐ends (T3, reverse primer; T7, forward primer). Amplicons were isolated via gel electrophoresis and gel extraction (Qiagen Gel Extraction Kit) and subsequently used as templates in enrichment PCR with the same primers. Anti‐sense riboprobes labelled with either digoxigenin‐UTP (DIG) or fluorescein‐UTP (FL) were synthesised from cDNA templates via T3 RNA polymerase‐mediated strand‐specific RNA transcription using DIG/FL RNA labelling mix (Roche). Hybridised, labelled riboprobes were detected using anti‐digoxigenin or anti‐fluorescein antibodies conjugated to horseradish peroxidase (Roche). Fluorescent tagging was performed using Alexa Fluor 594 and 488 dyes with streptavidin Tyramide Signal Amplification (TSA, Invitrogen). Finally, retinas were mounted in 70% glycerol in PBS, photoreceptor side up, on microscopy slides and covered with a coverslip.

Fluorescent‐tagged photoreceptor cells were imaged using a CFI Apo Lambda 60×/1.4 NA oil immersion objective on a spinning disk confocal microscope (Diskovery, Andor Technologies, UK) built around a Nikon Ti‐E body (Nikon Corporation, Japan) equipped with two Zyla 4.2 sCMOS cameras (Andor Technology), and controlled by Nikon NIS Elements software (Nikon Corporation, Japan). Exported images were further processed (maximum‐intensity z‐stack projection, channel merging) in ImageJ v.1.52p (National Institutes of Health, USA).

3. Results and Discussion

To reconstruct the evolutionary history of LWS opsins, we analysed available whole‐genome data and sequenced the retinal transcriptomes from 20 selected species that carry a second, short‐wavelength–shifted lws paralog, primarily gobies, elopomorphs, characins, and osteoglossiforms (Figure 1, Table S1). Combined with structural protein modelling, these data allowed us to focus on the repeated evolution of shorter‐shifted LWS visual pigments and the convergent emergence of green–red LWS opsin pairs in teleost fishes and humans (Figure 2).

FIGURE 1.

Lws opsin gene repertoire in teleost fishes. (A) Putative gene duplications (star symbols) and gene losses (X symbols) mapped onto the teleost phylogeny (simplified after Betancur et al. 2017). Most duplications are species‐ or genus‐specific, except in the Esociformes, Salmoniformes, and Cyprinodontiformes, in which they are older. Two ancient gene duplications are marked at the onset of teleosts (grey stars). (B) The number and type of lws genes found in the whole genomes of teleost fishes. Red—lws1 (red‐sensitive), orange—lws2, brown—lws3 (both shifted towards shorter, green wavelengths). Note that only a few lineages, including Elopomorpha, Osteoglossomorpha and Characiformes, retain lws2, and only gobies (Gobiiformes) have lws3. (C) Proportional gene expression in representative species with multiple lws paralogs: Tarpon (Elopomorpha), elephant fishes (Osteoglossiformes), characins (Ostariophysi), and gobies (Percomorpha). The position of these groups within the teleost phylogeny is highlighted in the inset. Refer to Figure S1 for the lws gene tree, Table S1 for the list of analysed genomes, and Figure 5 for the details on the goby lws3 evolutionary history.

FIGURE 2.

Convergent evolution towards green‐sensitive LWS vertebrate opsins. (A) Proposed model of the lws gene dynamics in teleost fishes. Lws1 is present in most fish lineages and the vsiual pigment it forms is commonly sensitive to longer (red) wavelengths > 560 nm. The lws2 copy is retained in Osteoglossomorpha, Elopomorpha and Characiformes (and is absent from Euteleostei). Lws3 occurs in gobies (Gobiiformes) and has presumably been lost in other fish lineages. Both LWS2 and LWS3 have been shown (e.g., Liu et al. 2019) or predicted in this study to be sensitive to shorter, greener wavelengths ~501 to 550 nm. (B) LWS opsin protein models by Boltz (Passaro et al. 2025). Across vertebrates, green sensitivity evolved convergently from a red‐sensitive ancestor via changes in key amino acids. Depicted are three amino acid sites that show the same mutations in human MWS (green) opsin and in both teleost LWS2 and LWS3 opsins (red arrows). Further sites have evolved convergently among the teleost LWS2 and LWS3 paralogs (blue arrows). * marks the mutation T269A, which is found in humans, teleost LWS2, and in LWS3 of mudskippers, but not in the two goby species selected for the modelling. For all 12 structural models performed in this study, please refer to Figure S8. A. p. = Amphligobius phalaena; S. o. = Stiphodon ornatus.

3.1. Lws Gene Duplications and Synteny

Reconstructing a comprehensive lws gene tree based on 109 teleost genomes and adding non‐teleost ray‐finned (Actinopterygian), lobe‐finned (Sarcopterygii), and several ancestrally derived fishes as outgroups, we confirm that most teleosts have one red‐sensitive lws gene (lws1). In addition, we find that tarpons (Elopomorpha) retain the lws2 copy, previously reported only in bony tongues and characins, and that it emerged during 3R (Liu et al. 2019; Escobar‐Camacho et al. 2020). Surprisingly, we also uncover a third ancestral lws (lws3) gene in gobies (Gobiiformes) (Figures 1 and 2; Figure S1).

The 3R‐associated event that gave rise to lws2 is currently the earliest described cone opsin duplication in teleosts (reviewed in Musilova et al. 2021). We describe three possibilities for the ancestral origin of lws3 (detailed below and in Table S4). Two of the scenarios suggest that lws3 is of a similar, or even more ancient, origin than lws2, while the third suggests an origin in the percomorph ancestor. However, based on our dataset, there is insufficient resolution to favour any of the scenarios. This is because dating the lws3 duplication is challenging as extant gene copies lack a sufficiently strong phylogenetic signal (Figure S1).

In detail, the three proposed scenarios are:

1. Lws3 is an ancient duplicate predating the teleost lws1–lws2 duplication, which is supported by the nucleotide tree (Figure S1). It also aligns with the convergent evolution of shorter‐shifted, green(er) spectral sensitivity between the fish LWS2 and LWS3 opsins. However, scenario 1 lacks strong node support and may, to some extent, be influenced by saturation at third‐codon positions (Breinholt and Kawahara 2013), rendering the phylogenetic signal questionable. Moreover, as gobies are a crown‐percomorph group, such an ancestral origin would require at least 15 independent losses of lws3 across teleosts (Figure 1).

2. Lws3 is homologous to lws2 (i.e., derived from the same ancestral duplicate). However, there is no phylogenetic support for this scenario besides for exon 4, which groups the goby lws3 with the exon 4 of lws2 (Figure S2). This hypothesis also lacks any syntenic support (Figure S3) and, similar to scenario 1, implies the independent loss of lws2/lws3 in all non‐goby euteleosts. Moreover, although LWS2 and LWS3 opsin proteins share some identical amino acid pairs in the retinal‐binding pocket (Figure 2B), this does not include all key tuning sites [e.g., position 269 (Yokoyama 2008)]; hence, the pattern may also have arisen convergently.

3. Lws3 is a more recent duplication originating in the percomorph ancestor. Supporting this scenario, to recover the monophyly between goby lws1 and lws3 always involves other percomorphs based on amino‐acid sequences or reduced nucleotide trees (i.e., excluding the third codon position). In other words, lws1 and lws3 sequences consistently cluster with other percomorph orthologs, with no evidence of a distinct Gobiiformes‐specific cluster (Table S4; Figure S4). As in scenario 1, scenario 3 is consistent with the convergent evolution of shorter‐shifted sensitivity between the fish LWS2 and LWS3 opsins. However, the phylogenetic trees for this scenario remain largely unresolved (Figure S4), and percomorph ancestry would also imply an unusually rapid sequence evolution in goby lws3.

Across all three scenarios, our results consistently support an ancestral origin of lws3 in either the teleost or percomorph ancestor. Moreover, regardless of the evolutionary scenario, the teleost LWS1 and LWS2/3 paralogs show convergence on the mammalian (primate) LWS and MWS opsin pair (Figure 2).

An additional and unexpected finding from our phylogenetic analyses is that lws genes from some tetrapods (amphibians, snakes, lizards, and mammals) and the non‐tetrapod lungfish cluster within the teleost lws clade. In contrast, lws genes from other tetrapods (birds and turtles) follow the species tree and branch outside the ray‐finned fishes (Figure S1). This pattern suggests the possibility of yet another ancestral lws duplication, potentially in the common ancestor of ray‐finned and lobe‐finned fishes. Under this scenario, birds and turtles would have retained a different lws paralog than other tetrapods. However, the support values for an additional ancestral lws duplication are low, and more extensive sampling across tetrapod lineages will be required to fully resolve this aspect of the lws evolutionary history.

A detailed analysis of the synteny (±10 flanking genes) around teleost lws loci revealed a highly conserved genomic context for lws1, which is consistently located between hcfc1 or sws2 upstream and gnl3l downstream (Figure S3). In contrast, lws2 shows reduced syntenic conservation (Figure 3) and occurs on distinct chromosomal regions in bony tongues and characins. Notably, we detected interspecific synteny between two lws2 genomic regions in tarpon (Elopomorpha) and the lws2 regions of Paramormyrops (Osteoglossomorpha) and cavefish (Characiformes). One tarpon lws2 region closely matches the Paramormyrops cluster, whereas the second corresponds more closely to the characiform cluster (Figure 3; Figures S3 and S5). This pattern implies that the teleost ancestor possessed two lws2 paralogs located in distinct genomic regions, both of which are retained in tarpons, while only one copy persists in osteoglossomorphs and the other in ostariophysans.

FIGURE 3.

Lws2 gene synteny. Two clusters of the lws2 gene, located on different chromosomes, are found in the Indopacific tarpon (Elopomorpha). One cluster is present in each of the elephant fish (Osteoglossomorpha) and the blind cavefish (Characiformes, Otomoropha). One of the tarpon clusters (NC_050591) shares more syntenic features with the elephant fish cluster (genes connected by blue shading). The second cluster (NC_050585) is more like the cavefish cluster (connected by orange bars). Note that the lws2 genes themselves no longer carry this phylogenetic signal (see gene trees in Figure 4; Figure S1). For the synteny of a longer genomic fragment (up to 2 Mbp), refer to Figure S5. See Figure S3 for the synteny of lws1 and lws3.

In gobies, lws3 is located in a distinct genomic region, flanked by rab32 upstream and stxbp5 or androglobin downstream (Figure S3). This region lacks any opsin gene in other teleosts, and no conserved synteny is detectable among the lws1, lws2, and lws3 loci (Figure S3). One possibility is that both lws2 and lws3 have undergone genomic relocation over evolutionary time. Transposable elements are known to play a significant role in opsin gene dynamics in teleosts, as demonstrated in African cichlids (Carleton, Escobar‐Camacho, et al. 2020), and both teleost rh1 and one lws1 copy in guppies and other livebearers originated via retrotransposition (Mano et al. 1999; Ward et al. 2008; Sandkam et al. 2017). Alternatively, the ancestral synteny may have been obscured by lineage‐specific gene losses or extensive chromosomal rearrangements over deep evolutionary time.

A close inspection of the lws cis‐regulatory elements revealed that the canonical vertebrate lws1 transcription factor binding sites (OTX, THR, RAX; Ghinia Tegla et al. 2020; Emerson et al. 2013; Volkov et al. 2020; Mackin et al. 2019; Nelson et al. 2009; Furukawa et al. 1997) are absent from the upstream regions of lws2 and lws3 (Table S2; Figure S6). Moreover, putative miRNA signals detected upstream of one lws2 copy in Megalops and lws3 in Neogobius exhibit low folding energies (miRNAfold; Tav et al. 2016) and limited similarity to vertebrate miR726, indicating distinct regulatory architectures for these paralogs, whose functional elements remain to be identified (Table S2; Figure S6).

3.2. The Dynamics of the Lws Evolutionary History

Most shallow marine and freshwater fishes retain at least one copy of lws1 (Figure 1). In contrast, many deep‐living taxa, including anglerfishes, cusk‐eels, toadfishes, and the deep‐water goby Lesueurigobius sanzi (de Buen, 1918), have lost lws genes entirely. This pattern is consistent with the rapid attenuation of long wavelengths with depth, Jerlov et al. 1976; which likely renders lws functionally obsolete in deep‐water environments. In contrast, lws1 duplications are common in lineages inhabiting red‐shifted or visually complex environments [e.g., some characins (Escobar‐Camacho et al. 2020)] or that use orange and yellow colouration, such as wrasses (Phillips et al. 2016) and livebearers (Windsor and Owens 2009). These multiple lws1 copies arose through lineage‐specific whole‐genome duplications (e.g., Cypriniformes, Salmoniformes, Esociformes) or tandem duplications, reaching up to 5 lws1 paralogs in species such as brown trout ( Salmo trutta ; Linnaeus, 1758), Siamese fighting fish ( Betta splendens ; Regan, 1910), and several wrasses (Figure 1).

Gene conversion between lws paralogs has been documented for lws1 copies in guppies and between lws1 and lws2 in characins (Escobar‐Camacho et al. 2020; Sandkam et al. 2017). Similarly, we found evidence for gene conversion in exon 3 of the lws1 and lws2 opsins in tarpons (Figure S5). Furthermore, GARD (Kosakovsky Pond et al. 2006) revealed conversion between lws1 and lws3 in the peacock gudgeon T. ocellicauda (Figure S7), resulting in two genes with highly similar functional properties (Figure 5). However, for the remaining gobies, patterns of gene conversion were less pronounced than those reported for other opsin duplicates, such as sws2 (Cortesi et al. 2015). This suggests that gene conversion occurred earlier and that the lws3 gene diverged rapidly after duplication, with subsequent gene conversion acting as a partial homogenising force, primarily affecting the 5′ portion of the coding sequence (Figure S7). At the same time, exons containing spectral tuning sites remained comparatively unaffected.

FIGURE 5.

Lws evolution in gobies (Gobiiformes). (A) Phylogenetic relationship between studied gobies based on the genomic and transcriptomic data set (three families, n = 15 species). (B) Data source: Genomic (G), transcriptomic (T), both (GT). (C) Mined lws genes, where each polygon represents one copy (lws1, red polygon; lws3, brown polygon) and an X marks gene loss. Note that no conclusion about gene losses can be drawn based on transcriptomic data alone (marked by?). (D) Proportional cone opsin gene expression in the gobies. The colour code in the horizontal bars represents the spectral sensitivity range of the corresponding visual pigment: Blue = sws2, green = rh2, red = lws1, and brown = lws3. Note that either lws1 (most gobies) or lws3 (Sycidiinae) dominate the cone opsin gene expression profiles. (E) The spectral sensitivity (λ_max) of the LWS1 (red) and LWS3 (brown) opsins based on the predictions in Table 1. Larger versus smaller circles represent the dominant versus rare transcripts, and an empty circle represents no expression, based on data in (D). Note the functional proximity of LWS1 and LWS3 in the peacock goby ( Tateurndina ocellicauda ), in which gene conversion occurred (see also Figure S7). (F–H) Double‐labelling fluorescent in situ hybridisation of expressed opsin mRNAs in retinal double cone photoreceptors in adult Amblygobius phalaena . Lws1 (magenta) and lws3 (yellow) mRNAs were expressed in opposite members of double cones—scale bars: 10 μm. For the in situ of the single cones, refer to Figure S9.

3.3. Convergent Red to Green(Er)‐Sensitivity Shifts in Vertebrate LWS Visual Pigments

Similar to the repeated shift from red‐ to green‐sensitivity in mammals (Carvalho et al. 2017; Melin et al. 2016), we find that LWS opsins with shorter wavelength sensitivity have evolved repeatedly in teleost fishes, with functional validation available for the LWS1 and LWS2 paralogs (Parry et al. 2003; Liu et al. 2019) (Table 1). Importantly, amino acid sites involved in this shift, including those for the human MWS‐LWS pair, have undergone convergent changes across vertebrates (Hagen et al. 2023). In teleosts, this has previously been described for several derived teleost LWS1 copies and LWS2 in characins and bony tongues (Liu et al. 2019; Escobar‐Camacho et al. 2020; Sandkam et al. 2018). Here, we describe convergent amino acid changes leading to a predicted shorter‐wavelength shift in LWS2 of the tarpons and the goby LWS3 (Figure 2A; Table 1).

The molecular evolution of these copies was assessed using structural models for one opsin pair from each of six species: human ( Homo sapiens , MWS and LWS), the osteoglossomorph aba ( Gymnarchus niloticus , LWS1 and LWS2), the elopomorph tarpon ( M. cyprinoides , LWS1 and LWS2), the ostariophysan characin ( Cyphocharax magdalenae , LWS1 and LWS2), and two percomorph gobies ( A. phalaena and S. ornatus , both LWS1 and LWS3). The models focused on amino acid variation within the retinal‐binding pocket, encompassing ten variable sites (Figure 2B; Figure S8; Table 2).

We identified three key positions: Y261F, T269A and F259L (bovine rhodopsin numbering) that show clear signatures of convergence between human and teleost LWS opsins; in addition, site 214 consistently distinguishes green‐ and red‐sensitive opsins, albeit via different amino acid substitutions (Figure 2B). Sites 261 and 269 are well‐established contributors to vertebrate LWS spectral tuning and form part of the ‘five‐sites rule’ or the ‘OH‐sites rule’ (Yokoyama and Radlwimmer 1998; Yokoyama 2008; Sekharan et al. 2012; Chi et al. 2020). In addition, an extra five amino acid substitutions (L88V, A118S, Q201K, 270 variable, and A275S), and changes at site 164 (member of the five‐sites rule), were common between shorter‐wavelength shifted LWS2 and LWS3 opsins (Figure 2B). These shared substitutions support the convergent evolution between these two teleost LWS paralogs (i.e., lws2 and lws3; unless they are homologous; see scenario 2, above; Table S4).

Despite convergently shorter shifting of the LWS duplicates, ancestral‐state reconstruction of the three most variable tuning sites (164, 261, and 269) revealed functional divergence among the paralogs (Figure 4). LWS1 consistently retains red sensitivity by mainly preserving longer‐wavelength–shifted residues, whereas LWS2 predominantly carries shorter‐shifted variants at these three sites. LWS3 exhibits an intermediate pattern: site 261 is shorter‐shifted in all gobies, site 164 is variably shorter‐shifted among species, and site 269 is shorter‐shifted only in mudskippers, with other gobies retaining the longer‐wavelength variant (Figure 4; Table 1). This was supported by a machine‐learning approach (OPTICS; Frazer and Oakley 2025) to predict the spectral sensitivities of the LWS opsins from 15 goby species and three teleost species with LWS2 (Table 1). Estimated λ _max values for goby LWS1 ranged from 551.9 to 567.3 nm, whereas green‐shifted LWS2 ranged from 517 to 529 nm, and LWS3 was intermediate, ranging from 523 to 557 nm. Our study did not include any physiological measurements, but our predicted values largely overlap with measured photoreceptor spectral sensitivities reported for other goby species, including the shorter‐shifted double cone with λ _max = 527–531 nm (Pierotti et al. 2020) or ~531 nm (Utne‐Palm and Bowmaker 2006). As these studies had no transcriptomic data available, it remains to be seen whether LWS3‐based visual pigments were solely responsible for the formation of said spectral sensitivity.

FIGURE 4.

Ancestral state reconstruction of the three most variable key‐tuning sites of the teleost LWS paralogs. (A) Sites 164, 261 and 269 are known to have a large effect on LWS spectral sensitivity (Musilova et al. 2021). LWS1 has mostly longer‐shifted amino acid variants at all three sites (red circles). LWS2 has shorter‐shifted amino acid variants (orange circles), resulting in a notable shift towards green spectral sensitivity. LWS3 lies in between, with a shorter‐shifted variant (F) at site 261, a predominantly longer‐shifted variant (T) at site 269, and a combination of both variants at site 164 (S and A) or an even shorter variant (P) encoded in yellow. Pie charts at the nodes indicate the scaled likelihoods of each variant, calculated using the ace function in APE (Paradis and Schliep 2019). (B) Association of amino acid variants in the teleost LWS evolution (325 haplotypes). Most copies are long‐wavelength‐sensitive (red) at most sites, reflecting the dominance of LWS1 in teleosts.

3.4. Neofunctionalization of LWS3 in Gobies

We confirmed the use of multiple lws copies across teleost lineages by examining their retinal expression patterns. Consistent with their predicted distinct spectral sensitivities, lws1 and lws2 were found to be expressed in the retinas of the tarpon (Megalops; Elopomorpha), gymnarchids (Gymnarchidae) and elephantfishes (Mormyridae; both Osteoglossomorpha), and in characins (Characiformes; Otomorpha) (Figure 1C). Similarly, both lws1 and lws3 were expressed in gobies (Gobiiformes), although the expression patterns varied markedly between species (Figures 1C and 5).

Comparing the retinal transcriptomes of ten goby species representing nine genera and most major subfamilies revealed that lws1 was the most highly expressed paralog in most taxa, with five species expressing lws3 to varying degrees (Figure 5). An exception was found in the subfamily Sycidiinae: the rainbow goby ( S. ornatus ) predominantly and the red lipstick goby ( S. rubicundus ) exclusively expressed lws3, respectively (Figure 5C). In addition to lws, all gobies expressed at least one copy of sws2, albeit at low levels in the Sycidiinae and in the spinycheek sleeper ( E. pisonis ). In contrast, the green‐sensitive opsin RH2 was generally absent or expressed only at trace levels, except for the Sycidiinae, where rh2 expression was comparable to sws2 levels in other gobies (Figure 5C). The rh2 gene is still present in the genome of some goby species, while pseudogenised in others (Musilova and Cortesi 2023). Taken together with the modelled, green‐shifted sensitivity of LWS3 (Figure 2; Table 1), the reduced reliance on rh2 supports a functional replacement of RH2‐mediated green vision by LWS3 in the adult gobies tested here. Although we cannot rule out that rh2 might be expressed at earlier developmental stages in some species, our findings are consistent with the neofunctionalisation of the lws3 paralog in gobies.

Neofunctionalisation of lws3 was further supported in adult A. phalaena , where, using fluorescent in situ hybridisation, we found that lws1 and lws3 are expressed exclusively in opposite members of the double cones (Figure 5F–H; Figure S9D–F). On the other hand, the sws2 opsins (sws2b and sws2aβ) were expressed exclusively in single cones (Figure S9A,B). Sws2aβ was expressed in nearly all photoreceptors, whereas sws2b showed variable and often lower expression, resulting in partial coexpression within the same single cones (Figure S9A–C). This cellular segregation demonstrates that lws3 has acquired an independent photoreceptor role in adult A. phalaena and, together with sws2‐ and lws1‐expressing photoreceptors, likely underpins trichromatic vision in this species.

4. Conclusion

By reconstructing the evolutionary history of lws opsins across ray‐finned fishes, we show that repeated gene duplication, divergence, and differential retention have generated three ancestral lws paralogs—lws1, lws2, and lws3—that remain functional in extant lineages despite extensive genomic rearrangement and regulatory divergence. Our discovery of lws3 reveals an unrecognised layer of sensory diversity and demonstrates that ancestral duplicates (originating in the teleost or percomorph ancestors) can be repeatedly co‐opted to expand the visual function long after their origin. In gobies, LWS3 has undergone neofunctionalization, and its sensitivity is predicted to be shorter‐wavelength‐shifted compared to LWS1, which, at least in adult specimens, functionally replaces the canonical green‐sensitive opsin RH2. This role likely parallels that of LWS2 in osteoglossiforms, elopomorphs, and characins, and mirrors primate colour vision, in which the duplication of lws enabled the evolution of green‐sensitive cones following the loss of rh2 in the mammalian ancestor (Carvalho et al. 2017). Across these systems, convergent amino acid substitutions at key spectral‐tuning sites underpin parallel functional shifts (between fishes and primates, and between two fish paralogs), highlighting common molecular solutions to similar selective pressures. Together, our results show that opsin gene evolution in teleost fishes is exceptionally dynamic, with ancestral gene duplicates serving as a recurrent substrate for sensory innovation. This evolutionary flexibility contrasts with the more constrained opsin gene repertoires of other vertebrates, underscoring the power of comparative genomic and functional approaches to reveal hidden aspects of visual system evolution.

Author Contributions

F.C. and Z.M. designed the research and analysed data; D.E.‐C. performed the ancestral state reconstruction, K.C. performed the regulatory regions tests and contributed to the overall interpretations. G.M.S. performed RNA extractions and transcriptome analyses. F.H. and X.D. performed the structural protein modelling. M.L. performed the fluorescent in situ hybridisation. F.C. and Z.M. wrote the initial version of the manuscript. All authors approve of the final version of the manuscript.

Funding

This work was supported by an Australian Research Council (ARC) Discovery Early Career Award (DE200100620), an ARC Future Fellowship (FT240100725), and a University of Queensland Amplify Fellowship to FC, a GAUK (1524119; Charles University) and SVV 260–792/2025 to GMS, and a Swiss National Science Foundation (SNF grant PROMYS, 166550), Charles University (Primus), Czech Science Foundation (21–31712S), and an ERC Consolidator grant (SensingDeep) to Z.M.

Disclosure

Benefit‐sharing: Benefits from this research arise from sharing our data and results in public databases, as described above.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: mec70592‐sup‐0001‐DataS1.zip.

Table S1: Samples and accession numbers used in this study.

Table S2: The LWS regulatory elements found in the lws1, lws2 and lws3 upstream gene regions of cavefish ( A. mexicanus ), elephant fish (P. kingsleyae), tarpon ( M. cyprinoides ) and goby ( N. melanostomus ). The miRNA folding energy was determined with https://evryrna.ibisc.univ‐evry.fr/miRNAFold. More negative energy means a more stable miRNA and is therefore more likely to be functional. The LCR binding sites (OTX2, THR or RAX) were identified using JASPAR (Fornes et al. 2020). Grey boxes mark genes not present. For more details on miRNA detection, please check Figure S5.

Table S3: Primers used for probe template synthesis for the fluorescent in situ hybridisation in Amblygobius phalaena . T7 (forward primer) and T3 (reverse primer) RNA polymerase promoter sequences are shown in small letters.

Table S4: The pros and cons of the three most likely scenarios for the duplication of lws3.

Figure S1: Phylogenetic gene tree and key tuning amino acids of LWS. Lws2 and lws3 genes are highlighted in orange and brown, respectively. Lws1 genes of the groups with multiple lws types (elopomorphs, osteoglossomorphs, characiforms and gobies) highlighted in red. Non‐ray‐finned fish clades highlighted in grey with labels (lampreys, sharks + rays, lungfish, tetrapods).

Figure S2: Single‐exon gene tree (MrBayes). Orange highlights the lws2 of osteoglossiform, characin, and elopomorph fishes; blue marks the goby lws3 opsins. Note that in exon 3, Megalops lws1 and lws2 are similar, most likely due to gene conversion. In exon 4, the lws3 and lws2 form a grouping, although not corresponding to the species tree.

Figure S3: Synteny of the lws1, lws2 and lws3 genes and the flanking regions in the round goby ( N. melanostomus ), elephant fish (P. kingsleyae), blind cavefish ( A. mexicanus ) and tarpon ( Megalops cyprinoides ).

Figure S4: Phylogenetic analysis excluding the third codon position yields an unresolved topology that includes percomorph fishes (lws1 no colour) and gobies (lws1 in rose, lws3 in vanilla).

Figure S5: Vista plot of the elephant fish (P. kingsleyae) and a cavefish ( A. mexicanus ) lws2 regions compared to two tarpon ( M. cyprinoides ) lws2 regions located on different chromosomes.

Figure S6: The miRNA detected in the upstream region of lws1 (all species), lws2 (Megalops, one copy) and lws3 (Neogobius).

Figure S7: Gene conversion among the two lws copies (lws1 and lws3) of the goby, Tateurndina ocellicauda . Large‐scale, more ancient gene conversion impacted exons 2–3 and exons 5–6. These two sequence stretches compromised the phylogenetic signal within the overall gene tree and had to be removed from the final analysis. (A) Identity plot of the two nucleotide sequences generated in Geneious 9.0 (https://www.geneious.com). Green bars mark identical stretches of the coding region, white ‘gaps’ indicate differences in sequence. Orange boxes highlight the putative regions of gene conversion, i.e., regions of high sequence similarity. (B) GARD (Kosakovsky Pond et al. 2006) results for the comparison of all goby lws1 and lws3 sequences. The most likely breakpoints for T. ocellicauda are highlighted in orange. (C) Drops in the dS ratio along the gene correspond to the detected converted regions. (D) A phylogenetic tree based on the longest converted region of the coding region, i.e., exon5‐6 (position 726–1074 of the genes), shows that the two T. ocellicauda sequences cluster together.

Figure S8: Structural models of LWS opsins in teleost fishes and in human.

Figure S9: Double‐labelling in situ hybridisation of opsin mRNAs in retinal single cone (A–C) and double cone (D–F) photoreceptors of adult whitebarred goby, Amblygobius phalaena . (A–C) sws2b (magenta) and sws2aβ (yellow) mRNAs were co‐expressed in some single cones. (D–F) lws1 (magenta) and lws3 (yellow) mRNAs were expressed in opposite members of double cones. Scale bars: 10 μm.

Data Availability Statement

Newly generated transcriptomic data have been submitted to NCBI's SRA database and are available under the accession numbers listed in Table S1. All other data are available in the Supporting Information, and raw data can be accessed from the authors upon request.