Not just shades of grey: life is full of colour for the ocellate river stingray (Potamotrygon motoro)

Colour vision refers to the ability of the visual system to distinguish between and respond to different wavelengths of light. It is based on the presence of two or more types of photoreceptors that feature different spectral maxima that have absorbance profiles that overlap. Recent findings have shown that most elasmobranchs (sharks, skates and stingrays), i.e. all shark and skate species investigated to date, are monochromats and are indeed unable to perceive colour behaviourally (e.g. Dowling and Ripps, 1970; Hart et al., 2011; Lisney et al., 2012; Ripps and Dowling, 1990; Schieber et al., 2012). Conversely, several batoids (rays and skates) are both physiologically (Bedore et al., 2013; Hart et al., 2004; Theiss et al., 2007) and behaviourally (Van-Eyk et al., 2011) able to perceive and discriminate colour, using multiple cone visual photopigments. For example, Glaucostegus typus (giant shovelnose ray) possesses three spectrally distinct cone photopigments, with wavelengths of maximum absorbance (λ_max) at 477, 502 and 561 nm (Hart et al., 2004), and can discriminate between hues of blue and grey of similar brightness (Van-Eyk et al., 2011). Other batoids were found to possess cone types with very similar absorption maxima such as Urobatis jamaicensis at 475, 533 and 562 nm (Bedore et al., 2013); Neotrygon (Dasyatis) kuhlii at 476, 498 and 522 nm (Theiss et al., 2007); and Rhinoptera bonasus at 470 and 551 nm (Bedore et al., 2013).

All of the previously studied batoids are marine species, with most living in brightly lit, clear water habitats such as coral reefs or sandflats. These environmental conditions are reflected in the spectral sensitivity of their cones, which absorb light maximally in the blue to green range of the visible spectrum (Hart et al., 2004; Lisney et al., 2012; Theiss et al., 2007). However, the ambient light conditions in marine environments differ considerably from those of freshwater systems (Loew and McFarland, 1990; Lythgoe, 1979), where members of the family Potamotrygonidae exclusively live (Brooks et al., 1981; Costa et al., 2013). Being endemic to South American rivers and lakes, freshwater stingrays inhabit beach sands, flooded forests, as well as habitats featuring muddy or stony substrates (Araújo et al., 2004; Rosa, 1985). Accordingly, Potamotrygon motoro, which utilises both shallow- and deep-water habitats, encounters clear water as well as murky water environments. Thus, the spectral sensitivity of the visual photoreceptors is likely to reflect this riverine ecology, with spectral absorbance peaks being shifted towards longer wavelengths; however, the only visual photopigment assessed and measured so far was in the rods (with a λ_max of 499 nm) (Muntz et al., 1973). The rod to cone ratio is very low (6–7:1) and represents one of the highest proportions of cones of any elasmobranch retina studied to date (Ali and Anctil, 1974). Considering that freshwater stingrays are thought to have derived from a freshwater-invading ancestor that occurred along the South American coast (Pacific and Caribbean) prior to the emergence of the isthmus of Panama (Lovejoy, 1996), it is intriguing to see whether the ability for colour vision still exists in this family and how visual photopigments are spectrally tuned to match ambient light conditions of freshwater habitats.

Several studies have shown that elasmobranchs are cognitively on a par with other vertebrates and can solve many cognitive visual tasks (Schluessel, 2015). For example, P. motoro uses its visual system for orientation (Schluessel and Bleckmann, 2005; Schluessel et al., 2015a; Schluessel and Ober, 2018) and can discriminate successfully between different quantities presented in the form of two-dimensional geometric symbols on cards (Christofzik, 2016). Nonetheless, in behavioural experiments, its visual resolution was found to be low, around 0.23 cycles deg⁻¹, which was determined by discriminating between horizontally and vertically arranged black and white stripes (Daniel et al., 2020; Daniel and Schluessel, 2020). Potamotrygon motoro is probably a diurnal species, but also shows some activity during certain times at night (Velte et al., 2002; V.S., unpublished observation).

In the presented experiments of this study, the ability of five female ocellate river stingrays, Potamotrygon motoro (J. P. Müller and Henle 1841), to discriminate between coloured and grey stimuli of equal brightness, as well as between additional grey distractor stimuli of varying brightness, was tested. In addition, transcriptome sequencing and predicted spectral tuning analyses were performed to determine which opsin genes are expressed in the retina of P. motoro. Colour cues (red and green) were chosen according to potentially relevant colours found in a riverine system (Costa et al., 2013). Responses to blue and yellow were also tested for comparative purposes, as these colours were previously assessed in two marine species (Schluessel et al., 2015b; Van-Eyk et al., 2011). Additionally, we assessed whether and how well stingrays can discriminate between the training stimulus (red) and another shade of red featuring either a lower or a higher brightness.

All applicable international, national, and/or institutional guidelines for the care and use of animals were followed. All procedures performed in this study were in accordance with the ethical standards of the University of Bonn, Germany.

For opsin identification, one P. motoro individual was euthanised with an overdose of MS222 according to the local animal ethics guidelines of the University of Bonn, Germany. Both retinas were carefully dissected, placed in an RNA stabilisation solution (RNA-Later, Sigma-Aldrich), incubated overnight at 4°C to allow for passive diffusion of RNA-Later and stored at −20°C until needed. Subsequently, both samples were shipped on dry ice to BGI (Hong Kong, China) for next-generation sequencing (NGS; specifically, reference-based and de novo RNA-Seq) as well as transcriptome quantification.

Briefly, total RNA was extracted, before mRNA enrichment using oligo-dT selection. The mRNA fraction was fragmented and subjected to reverse transcription to produce double-stranded cDNA (dscDNA) using random hexamers. The resultant dscDNAs were end-repaired and 3ʹ-adenylated; adapters were then ligated to the ends of the adenylated dscDNA fragments. The ligation products were purified and PCR amplified using primers to enrich the purified cDNA template. Subsequently, PCR products were denatured by heat and the single-stranded DNA (ssDNA) cyclised by splint oligo and DNA ligase, prior to sequencing on the BGISEQ-500 platform.

For the bioinformatics workflow, low quality reads (where base quality was less than 10% or greater than 20% per read), reads with adapters, and those with unknown bases (N bases more than 5%) were filtered out to obtain clean reads. These clean reads, stored as FASTQ files (Cock et al., 2010), were assembled using Trinity v2.0.6 (via Inchworm, Chrysalis and Butterfly software modules; Grabherr et al., 2011) and transcripts were clustered into Unigenes by TIGR Gene Indices clustering tools (TGICL v2.0.6; Pertea et al., 2003). This was followed by Unigene functional annotation using Blastn v.2.2.23 (Altschul et al., 1990), Blastx v.2.2.23 (Altschul et al., 1990) and Diamond v0.8.31 (Buchfink et al., 2015) to align Unigenes to various functional databases, namely NT, NR, Gene Ontology [GO; via Blast2GO v2.5.0 (Conesa et al., 2005) with NR annotation], EuKaryotic Orthologous Groups (KOG), Kyoto Encyclopedia of Genes and Genomes (KEGG), UniProtKB/SwissProt and InterPro (via InterProScan5 v5.11-51.0; Quevillon et al., 2005). Unigene coding domains were identified using TransDecoder v3.0.1 and the longest open reading frames (ORFs) were extracted. These ORFs were translated and Pfam v32 comparison analysis (El-Gebali et al., 2019) was conducted by blasting to SwissProt and HmmScan to predict protein sequences. Differential gene expression was conducted using hierarchical clustering analysis with hclust function software. Briefly, clean reads were mapped using Bowtie2 v2.2.5 (Langmead and Salzberg, 2012), before gene expression levels were calculated with RSEM v1.2.12 (Li and Dewey, 2011), where the relative expression of a transcript is linearly proportional to the number of cDNA fragments that originate from it. Thus, the calculation of fragments per kilobase of transcript per million (FPKM) mapped reads is an accepted proxy for relative gene expression (Mortazavi et al., 2008; Trapnell et al., 2010).

A codon-matched nucleotide sequence alignment of 36 gnathostome (jawed vertebrate) opsin coding regions, ranging from cartilaginous fishes to mammals, was generated by ClustalW (Higgins et al., 1996) and manually manipulated to refine the accuracy of cross-species comparison. Specifically, the alignment incorporated the opsin sequences of the three visual photopigments expressed in the retina of P. motoro (the ocellate river stingray) (GenBank accession numbers: MN954690–MN954692) compared with those species listed in the figure legend of the phylogenetic tree. All five opsin classes were included, with zebrafish (Danio rerio) vertebrate ancient (va) opsin sequences (Davies et al., 2010), specifically va1 and va2, used collectively as an outgroup given that this opsin type is a sister clade to all five visual photopigment classes.

Phylogenetic analysis of 1000 replicates was conducted in MEGA7 (Kumar et al., 2016), with evolutionary history being inferred by using the Maximum Likelihood method based on the General Time Reversible model (Nei and Kumar, 2000), resulting in a tree with a highest log likelihood of −17,075.70. Initial trees for the heuristic search were obtained by applying Neighbour-Joining and BioNJ algorithms (Saitou and Nei, 1987) to a matrix of pairwise distances estimated using the Maximum Composite Likelihood approach (Tamura and Nei, 1993), with a homogeneous pattern of nucleotide substitution among lineages, uniform rates, and bootstrapping with 1000 replicates. The tree was drawn to scale, with branch lengths measured in the number of substitutions per site. A total of 831 positions was present in the final dataset, with all positions with less than 95% site coverage being eliminated. That is, fewer than 5% alignment gaps, missing data and ambiguous bases were allowed at any position.

There are >40 known amino acid tuning sites that alter the spectral peak of absorbance (λ_max) of the five classes of opsin-based photopigments in vertebrates. As only lws, rh2 and rh1 opsin-based photopigment classes are relevant to cartilaginous fishes in this case, sws1 and sws2 spectral tuning will not be discussed further. Using conventional numbering based on the bovine rod opsin protein sequence (GenBank accession number NP001014890; Palczewski et al., 2000), a number of studies have determined that the following 27 amino acid sites are important for the spectral tuning of many rh1 photopigments, although for both rh1 and rh2 photopigments there are predominantly seven critical sites (underlined): 83, 90, 96, 102, 113, 118, 122, 124, 132, 164, 183, 194, 195, 207, 208, 211, 214, 253, 261, 265, 269, 289, 292, 295, 299, 300 and 317 (Chan et al., 1992; Davies et al., 2012, 2007; Hope et al., 1997; Hunt et al., 2001; Janz and Farrens, 2001; Sakmar et al., 1991; Yokoyama, 2000, 2008; Yokoyama et al., 2007, 2008, 1999). Similar studies applied to other photopigment classes have determined that five sites (namely 164, 181, 261, 269 and 292) are important for determining the λ_max values of lws photopigments (Davies et al., 2012; Yokoyama, 2000). Inspection of these sites was used to estimate the λ_max value for the three photopigments, lws, rh2 and rh1, expressed in the retina of P. motoro. The overall spectral effects of these sites and their relevance to elasmobranch vision are presented in the Results and Discussion. Dark spectra for these three visual photopigments were generated using a standard A₁-based rhodopsin template (Govardovskii et al., 2000).

For each behavioural experiment, five P. motoro (five females for the first, four males and a female for the second; disc width 15–30 cm) were obtained from Frankfurt Zoo, Germany. All animals were kept in a wooden aquarium (2.35 m×2.08 m×0.4 m) that was lined with black pond foil and filled with deionised, aerated, filtered water at 28–29°C (Fig. 1). The walls were 40 cm in height and the water level was at a height of around 25 cm. Reef Salt (Aqua Medic) was added to the water to maintain a conductivity between 380 and 420 µS. During training, food (e.g. shrimps, mussels, trout and red mosquito larvae) was only provided during the experimental trials. Experiments were conducted during the day, from morning until early afternoon. There was a natural light:dark cycle and animals were identified by phenotypic characteristics. The aquarium contained various pots, plants and other equipment, for sheltering and ecological enrichment. Water depth was approximately 30 cm.

The housing tank also served as the experimental tank (Fig. 1). The experimental set-up (53.5 cm×106 cm) was situated within the holding tank (2.08 m×2.35 m), and consisted of a starting compartment (including a start box) and a decision area, separated by a manually operated guillotine door (Fig. 1). When not engaged in trials, stingrays could freely swim throughout the entire set-up, whereas during experiments, the guillotine door was closed. All experiments were conducted as two alternative forced choice experiments, i.e. stingrays had to choose between two stimulus cards, presented simultaneously to the right and left of a divider in the decision area. There were eight overhead fluorescent lamps (Radium Spectralux plus, 58 W) at a height of about 4 m above water level. The experimental set-up was oriented towards the lamps in such a way to allow for equal illumination of the decision area. Stimulus cards were laminated and measured 21.0×29.7 cm in size (i.e. A4).

Stimuli used for the discrimination tests were selected by taking the river environment that P. motoro inhabits into account; accordingly, red and green hues were chosen. For comparison, blue and yellow hues were also tested. Spectra were originally selected according to human perception and were produced on a computer using the program Microsoft Paint^©.

Standardised reflectance scans of the experimental stimuli and the visual background (grey–brown plastic partitions) were recorded using an Avantes Avaspec 2048 spectrophotometer connected to a deuterium–halogen light source (Avantes AvaLight DHs) for illumination. Measurements were taken by holding a bifurcated 200 µm fibre-optic probe with unidirectional illumination and recording at a 45 deg angle to the respective surface. The probe end was inserted in a darkened pipette tip to exclude ambient light and to take measurements at a fixed distance of 0.3 cm from the measured surface. The reflectance intensity between 400 and 700 nm for each stimulus card and the visual background was determined relative to a 98% white reference (Spectralon WS-2) and a black standard (shut-off light source). Five measurements were made in succession without changing probe contact and were averaged for each stimulus. Data were recorded with Avasoft 7.7 (Avantes) and imported into Microsoft Excel. Downwelling irradiance was measured in air under experimental lighting conditions by placing an Avantes CC-UV/Vis cosine corrector at the position of the eye of the test animal. Conversion of the measured spectral irradiance (Fig. 2G) to illuminance was performed by multiplying the irradiance spectrum (in W m⁻² nm⁻¹) by the photopic luminosity curve for humans and by Δλ (Johnsen, 2012). The products were then summed and multiplied by 673 according to Johnsen (2012), which resulted in a light intensity of 417 lx, which lies in the photopic range of diurnal species such as humans (Davies et al., 2012). Hence, a potential contribution of rods to colour vision can be largely ruled out under the chosen experimental lighting conditions.

There were two experiments: the first featured a red (Red 120) and a green stimulus (Green 60) together with four grey stimuli (Grey 20, 60, 120, 170; Fig. 2A,B), as well as 10 red stimuli of varying brightness (Red 40, 50, 60, 70, 90, 120, 170, 180, 190, 200; Fig. 2C,D). The second one featured a blue (Blue 160) and a yellow (Yellow 200) stimulus card, and seven grey stimuli (Grey 60, 140, 160, 190, 200, 220, 230; Fig. 2E,F). Grey 0 would have been equal to black (according to human perception), and grey 240 equal to white; the numbers in between give the degree of shading.

At the beginning of each trial, a stingray was placed in the starting compartment. To start, the stingray had to enter the start box and push against the guillotine door with its snout. A choice was recorded as soon as the stingray crossed over a line that was perpendicular to the beginning of the partition that divided the right and left compartments. The two stimuli to be discriminated were displayed simultaneously (one in each division) and switched randomly between the two sides to prevent any side bias and direction conditioning. In addition, four alternating rotational schemes for stimulus presentation were used to vary the succession of stimuli shown on a particular side between sessions (i.e. in each scheme, the sequence of when the correct stimulus was presented on the right or left side was different) and thereby to prevent simple learning of stimulus presentation sequence by animals. A correct choice (i.e. choosing red/blue) was rewarded with food. There was an inter-trial interval (ITI) of 30 s. If a stingray did not choose a stimulus within 2 min, the trial was aborted; three such trials would terminate a session. Training sessions were carried out 6 days per week; each session consisted of 10 regular training trials plus two transfer test trials once the learning criterion was achieved. Training for each individual was considered successful or completed as soon as a learning criterion of ≥70% correct choices on three subsequent sessions was reached (χ²₁≤0.05; to show statistical significance). To prepare the stingrays for subsequent transfer test trials, which were interspersed with regular trials over the next sessions (but were unrewarded), food was only provided in 8 out of 10 correct trials in the three sessions following successful training. Prior to each session, it was randomly determined which trials remained unrewarded (regardless of the actual choice made by the stingray). The first and last trials within a session were always rewarded and there were never two unrewarded trials in a row. Not rewarding some of the regular trials was intended to prevent the stingrays from realising that transfer trials were unrewarded and, therefore, not worth participating in.

After successful training, two transfer test trials were randomly interspersed with the 10 regular trials per session. There were several types of transfer test, as explained in the following sections. A total of 10–20 trials per type of transfer test were conducted per ray (N=5 in experiment 1 and N=4 in experiment 2). These trials were meant to elucidate whether the stingrays had based their choice on stimulus colour or brightness, so one of the original training stimuli was now presented in combination with various grey stimuli. Transfer tests were unrewarded; they were neither conducted at the very beginning nor at the end of a session, and never preceded or followed one of the unrewarded regular trials. Performance in the transfer tests was assessed and compared with that in regular trials.

The red (Red 120) and the green (Green 60) cards were presented during all training trials and in all regular trials thereafter; the choice of the red card was rewarded with food. The two stimuli varied in hue and brightness.

The first set of transfer tests (T1–T3, n=20 per type of test) was conducted to determine whether red was chosen over green based on colour or brightness differences. Accordingly, red was paired with a grey of either a greater brightness than red (T1=Grey 120, T2=Grey 170) or a lower brightness than red (T3=Grey 60).

The second set of transfer tests (T4–T6, n=10 per type of test) was designed to test whether the alternative stimulus was remembered (as being ‘incorrect’) independently of the rewarded training stimulus and, if that was the case, to determine whether green was remembered based on colour and/or brightness. Accordingly, green was paired with grey of either a greater brightness than green (T4=Grey 120) or a lower brightness than green (T5=Grey 60 and T6=Grey 20).

In a third set of transfer tests (T7–T15, n=20) it was investigated whether stingrays could discriminate between the red training stimulus (Red 120) and another shade of red, i.e. either with a lower brightness than Red 120 (T7=Red 40, T8=Red 50, T9=Red 60, T10=Red 70, T11=Red 90) or with a higher brightness than Red 120 (T12=Red 170, T13=Red 180, T14=Red 190, T15=Red 200). The visual model, based on the transcriptome data (which were collected after the behavioural experiments were conducted), showed that in some cases the varying shades of red did not just vary in brightness but also in colour, as can be seen in Fig. S1 and Table S1, which provided an additional means of differentiation.

The blue (Blue 160) and yellow (Yellow 200) cards were presented during all training trials and in all regular trials thereafter; the choice of the blue card was rewarded with food. The two stimuli varied only in hue but were of equal brightness.

Following successful training, the first set (T1–T5) of 10 different types of transfer test was conducted to test whether blue was chosen based on colour or brightness differences. Blue was paired with a grey of either a lower brightness than blue (T1=Grey 60, n=10; T2=Grey 140, n=10; T3=Grey 160, n=20;) or a greater brightness (T4=Grey 190, n=10; T5=Grey 220, n=10).

In a second set of transfer tests (T6–T10), yellow was paired with a grey of either a lower brightness (T6=Grey 140, n=10) or a greater brightness than yellow (T7, Grey 190, n=10; T8=Grey 200, n=20; T9=Grey 220, n=10; T10=Grey 230, n=10).

The average trial time and the percentage of choices for each stimulus card were recorded for each session per individual. To test for statistical significance of learning success, the learning criterion was established to be ≥70% correct choices in three consecutive sessions (P≤0.05). Generalised linear mixed models (GLMM) were run in R (using a binomial distribution with success as the response variable (0=success, 1=failure), individual as random variable and the distractor as a fixed factor to test whether stingrays as a group chose one stimulus significantly more often than another during the transfer tests. For all tests, P≤0.05 was considered significant and P≤0.001 highly significant.

To identify the visual photopigments found in the rods and cones of P. motoro, RNA-Seq or transcriptome profiling was used. Specifically, retinal tissues from two eyes were sent to BGI (Hong Kong), where raw transcript sequence data were generated using a BGISEQ-500 sequencing platform. These reads were filtered to generate clean reads and were subjected to quality control pipelines as described in Materials and Methods.

To obtain potentially useable assembled libraries, three different approaches were applied. Firstly, filtered clean reads expressed in the retina of P. motoro were compared and assembled using the genome of the elephant shark (Callorhinchus milii) as a reference. The results indicated that the average mapping ratios with the reference genome and gene databases were 0.04% and 0.45%, respectively, whereas there was shared identity of 17.49% when comparing protein sequences (via the NR protein database). Secondly, when compared with the whale shark (Rhincodon typus), with an NR protein database identity of 55.44%, genome and gene mapping ratios (i.e. at 0.20% and 1.90%, respectively) were higher than those obtained with the application of the elephant shark genome as a reference. In both cases, despite being cartilaginous fishes, the mapping ratios for obtaining meaningful sequence and quantification data for the P. motoro retinal transcriptome were too low. As such, a validated de novo strategy using Trinity (Grabherr et al., 2011) was employed to generate assembly data that showed higher sequence fidelity that could be quantified more accurately.

In total, the de novo RNA-Seq approach, using a BGISEQ-500 sequencing platform, generated 67.47 megabytes (MB) of raw transcript sequence data (Table S2). These reads were filtered to generate 66.47 MB of clean reads (i.e. 98.52% clean output reads) as described above with quality control metrics outlined in Table S2. When converted, the filtered sequences yielded 6.65 gigabases (Gb) of clean data. Unigenes were annotated by aligning to seven functional databases (see Materials and Methods), which resulted in the detection of 26,113 coding domain sequences with discernible ORFs by using TransDecoder software.

All resultant ORFs were compared and blasted to various protein databases [e.g. Pfam (25) and SwissProt], which resulted in the identification of three unique Unigenes (and their translated gene products) as being closely related to visual opsins of other vertebrates, especially those found in other cartilaginous fishes (Fig. S1). Indeed, closer inspection of these sequences by phylogenetic analysis, in comparison with elephant shark and whale shark visual photopigment genes, as well as those from other vertebrates, confirmed the visual opsins expressed in the P. motoro retina to consist of a single rod opsin (encoded by the rh1 gene), and two cone photopigments, namely the long-wavelength-sensitive (lws) and rod opsin-like 2 (rh2) opsin gene classes (Fig. 3; Fig. S2). Further analysis demonstrated the absence of the two short-wavelength-sensitive opsin classes, namely sws1 and sws2, a feature also found in the few cartilaginous fishes currently studied (Fig. 3; Fig. S2). Thus, this species of stingray, in concert with behavioural analyses shown above, exhibits a dichromatic colour visual system with a duplex retina. Critically, unlike in the small number of sharks and rays that have been molecularly investigated to date, this study is the first to genetically identify the presence of an rh2 photopigment expressed in the eye of elasmobranchs (Fig. 3; Fig. S2); the rh2 opsin gene has been lost in other monochromatic or colour-blind species, where only the lws cone opsin gene appears to be present [in addition to a single rod (rh1) opsin gene] (Delroisse et al., 2018; Fasick et al., 2019; Hara et al., 2018; Hart et al., 2011; Theiss et al., 2012). Currently, the only other cartilaginous fish to possess this opsin class is the elephant shark, C. milii, a species that is a deep-sea chimaera (Holocephali) and not an elasmobranch (Elasmobranchii) (Davies et al., 2009a). To quantify visual opsin mRNA levels found in the retina of P. motoro, differential gene expression was conducted using hierarchical clustering analysis with hclust, mapping of clean reads using Bowtie2 (Langmead and Salzberg, 2012), and RSEM (Li and Dewey, 2011). As a proxy for relative mRNA levels (Mortazavi et al., 2008; Trapnell et al., 2010), FPKM values were calculated, which demonstrated that while the two P. motoro cone photopigment genes were expressed at similar levels (FPKM_lws=170.44 versus FPKM_rh2=109.84; FPKM_lws/FPKM_rh2 ratio of 1.55), there was far greater expression of the rh1 opsin gene than the cone photopigment genes (FPKM_rh1=9431.79; FPKM_rh1/FPKM_lws ratio of 55.34 and FPKM_rh1/FPKM_rh2 ratio of 85.87) (Fig. 4). Collectively, when comparing rod versus cone gene expression, the FPKM_rod/FPKM_cone ratio was 33.65, suggesting a rod-dominant duplex retina in P. motoro (Fig. 4).

Only three P. motoro visual photopigment genes were identified by RNA-Seq analysis, namely lws, rh2 and rh1, so the absence of both sws1 and sws2 opsin genes in cartilaginous fishes (Davies et al., 2009a; and this study) will not be discussed further. It is important to note that transcriptome profiling failed to identify the expression of cytochrome P450 family 27 subfamily c member 1 (cyp27c1) in the retina of P. motoro. As this is the enzyme responsible for converting 11-cis retinal (present only in rhodopsins that utilise a vitamin A₁-based chromophore) into the longer wavelength-shifted 11-cis 3,4-didehydroretinal (found in porphyropsins that utilise a vitamin A₂-based chromophore) (Enright et al., 2015), all visual photopigments expressed in the retina of the ocellate freshwater river stingray studied here will possess vitamin A₁-based photopigments only.

In cases studied thus far, the spectral sensitivity of visual photopigments in elasmobranchs has been measured by microspectrophotometry (MSP) (Hart et al., 2011; Theiss et al., 2012) and, as such, the underlying opsin genes are unclear. However, by using a sensitive in vitro photopigment regeneration technique (Davies et al., 2009a), direct correlations between opsin gene identification, the protein sequence and spectral tuning sites, and the spectral peak of absorbance (λ_max) of the resultant photopigment have been demonstrated in another cartilaginous fish, namely the elephant shark (C. milii) (Davies et al., 2012, 2009a). Importantly, this evolutionarily related species also expresses lws, rh2 and rh1 photopigments that utilise a vitamin A₁-based chromophore, just like P. motoro. As such, the spectral sensitivity of visual photopigments identified in P. motoro was accurately predicted by direct comparison with those found to be expressed in the C. milii retina.

For lws photopigments, the λ_max values are determined by the residues present at five sites (namely 164, 181, 261, 269 and 292, using conventional numbering based on the bovine rod opsin protein sequence) (Davies et al., 2012; Yokoyama, 2000; Yokoyama and Radlwimmer, 1999). In the elephant shark, the lws2 photopigment possesses the lws tuning site complement of Ser164, His181, Phe261, Thr269 and Ala292; the spectral tuning of the P. motoro lws photopigment is identical to this, except for a Phe261Tyr substitution (Fig. S2). As a Tyr261Phe amino acid change causes a long-wavelength shift of 7 nm (Yokoyama and Radlwimmer, 1999), the λ_max value of the P. motoro lws photopigment was predicted to be 555 nm (i.e. C. milii, λ_max(lws2)=548 nm, plus 7 nm) (Fig. 5).

For both vertebrate rh1 (rod) and rh2 (cone) photopigments, there are seven major tuning sites, namely 83, 122, 207, 211, 265, 292 and 295 (Chan et al., 1992; Davies et al., 2012, 2007; Hope et al., 1997; Hunt et al., 2001; Janz and Farrens, 2001; Sakmar et al., 1991; Yokoyama, 2000, 2008; Yokoyama et al., 2007, 2008, 1999). In the elephant shark rh1 photopigment, these spectral tuning sites are Asp, Glu, Met, His, Trp, Ala and Ala, respectively, with a λ_max value of 496 nm (Davies et al., 2009a). As these sites are identical in the rh1 photopigment expressed in the retina of P. motoro (Fig. S2), the λ_max value of the stingray rod photopigment is also predicted to be 496 nm (Fig. 5). For rh2 photopigments, the complement of the seven spectral tuning sites in C. milii is Asn, Gln, Leu, His, Trp, Ser and Ser, respectively, with a λ_max value of 442 nm (Davies et al., 2009a). The P. motoro rh2 photopigment only differs from this at three residues, namely Glu122, Tyr211 and Ala292 (Fig. S2). Based on the literature, Gln122Glu, His211Tyr and Ser292Ala substitutions result in mean spectral shifts of +15 nm, −5 nm and +9 nm, respectively (Musilova et al., 2019; Nathans, 1990a,b; Yokoyama et al., 1999; Zhukovsky and Oprian, 1989). Thus, the overall shift in the spectral sensitivity of the rh2 photopigment of P. motoro compared with the elephant shark rh2 photopigment is +19 nm, which results in a predicted λ_max value of 461 nm (i.e. C. milii, λ_max(rh2)=442 nm, plus 19 nm) (Fig. 5).

Under the bright light experimental conditions presented here, where rods are bleached and cone-based colour vision would play a major role, only the two cone photoreceptors would be active. As shown in Fig. 5, the predicted spectral profiles for the lws and rh2 photopigments overlap at 505 nm with a normalised absorbance level at 44% (Fig. 5, line ii), a result that is consistent with photoreceptor opponency and the potential for dichromatic colour vision. It is possible, however, that under mesopic conditions in the wild, when both rods and cones might be active, the rh1 photopigment may provide a ‘green’ channel. The activity of rods and cones under these mid-illumination conditions is likely to be species specific, so further study is required to determine the potential for any mesopic-based activity of P. motoro photoreceptors. If deemed to be correct, however, mesopia may result in the potential for increased wavelength discrimination and the presence of ‘conditional trichromacy’ in P. motoro (as hypothesised for other vertebrates in Davies et al., 2012; Davies et al., 2009b; Katti et al., 2019; Katti et al., 2018;,Theiss et al., 2012), with an rh2/rh1 overlap at 479 nm at a normalised absorbance level at 92% (Fig. 5, line i) and an rh1/lws overlap at 529 nm at a normalised absorbance level at 71% (Fig. 5, line iii). Of course, overall spectral sensitivity is not only dependent on photopigment spectra but also influenced by a number of other factors, such as the number of photoreceptors, outer segment length, gene expression and opsin protein levels. As a proxy for spectral sensitivity, P. motoro photopigment spectral data, based on standard A₁ templates, were multiplied by the expression level of each opsin (Rennison et al., 2016) where the following assumptions were made: (1) that protein levels are relatively similar or reflect mRNA levels (especially when comparing between different opsins) and (2) that the predicted relative levels of opsin protein could be housed within a small number of cells with long outer segments, a higher number of cells with shorter outer segments, or both. Thus, mRNA expression was taken as a proxy parameter to account for these differences. As the expression of rh1 is much greater than for cones, relative opsin-based spectral sensitivity was presented on a semi-logarithmic scale (Fig. 6). Firstly, the figure shows that rod opsin spectral sensitivity dominates over that of the cones, which is consistent with a higher rod to cone ratio (Ali and Anctil, 1974). Similarly, the rh2 absorbance peak increases compared with that of the lws photopigment, such that the degree of spectral overlap between the cones changes from 44% to 80%, with an approximate short-wavelength shift at the point of cone overlap from 505 nm to around 490 nm (Fig. 6A). When considered together, the opsin-based spectral sensitivity of the two cones maximally plateaus with a spectral range from around 400 to 570 nm (Fig. 6B).

In the second set of transfer tests (T4–T6; n=50 per test), the (alternative) green stimulus was paired with a grey stimulus. Accordingly, there was no ‘correct’ colour stimulus, only the ‘incorrect’ green (which stingrays were expected to avoid if they had indeed learned that choosing green was incorrect) and a new stimulus (grey). In all three tests, stingrays either preferred the grey stimulus over the green stimulus or did not make a choice at all (28 trials, 18.7%; Table 1, Fig. 7). In all three tests, grey was chosen significantly more often than green or ‘no choice’ (T4; GLMM: d.f.=48, z=3.71, P<0.0001; T5; GLMM: d.f.=48, z=3.244, P=0.0012; T6; GLMM: d.f.=48, z=3.12, P=0.0018). Average trial times and statistical results are provided in Table 1. For individual stingray data, please refer to Table S3a.

In the third set of transfer tests (15 types, n=100 per test), the red training stimulus was always chosen significantly often over an alternative shade of red (Table 1, Fig. 8), with only one exception, i.e. Red 170. Table 1 shows that the greater the variation in brightness between the two shades, the better the training stimulus was distinguished from the alternatives. It also appears that darker shades of red were slightly better distinguished from the training stimulus (Red 120) than lighter shades of red. Shades between Red 90 and Red 120, and between Red 120 and Red 170, were not tested, as not a single individual chose significantly between Red 120 and Red 90 or between Red 120 and Red 170; however, group averages were still significantly above chance level for Red 120 versus Red 90. For Red 120 versus Red 180 or Red 80, two individuals chose significantly above chance level, while for Red 120 versus Red 190 or Red 70, all except one individual chose significantly above chance level. For individual stingray data, please refer to Table S3a.

Four rays reached the learning criterion within 3.5±0.9 (mean±s.d.) sessions. The fifth ray reached the learning criterion after 19 sessions, but ceased participation after that and was, therefore, excluded from further analysis. During the transfer test period, the blue stimulus was chosen in 98.5±1.0% of regular trials (Fig. 9). In the transfer tests (T1–T3; n=38–78 per test; Table 1, Fig. 9), the blue stimulus was chosen significantly more often than the alternative when the grey stimulus featured only a slightly lower brightness (Grey 160, T3) or greater brightness (Grey 190, T4; Grey 220, T5) than blue, but not in the two transfer tests in which the grey stimulus was of a much lower brightness than blue, i.e. Grey 60 (T1) and Grey 140 (T2). Actual number of choices, average trial times and statistical results are provided in Table 1. For individual stingray data, please refer to Table S3b.

In the second set of transfer tests (T6–T10; n=36–80 per test, Table 1), the alternative stimulus, i.e. the yellow stimulus, was paired with a grey stimulus. Accordingly, there was no ‘correct’ colour stimulus to choose, just the ‘incorrect’ yellow (which rays were expected to avoid if they had indeed learned that choosing yellow was incorrect). In all five tests (T6–T10), stingrays preferred any grey over yellow, regardless of whether the grey was of lower or greater brightness than yellow (Table 1, Fig. 9). For individual stingray data, please refer to Table S3a.

The transcriptome results presented here demonstrate that P. motoro expresses three retinal visual photopigment genes, a single rod opsin (encoded by the rh1 gene) and two cone photopigments (encoded by the lws and rh2 opsin genes). Relative transcript levels were compared (i.e. rod versus cone opsins) and the significant level of rh1 expression in comparison with total cone expression was shown to be consistent with the high rod:cone ratio for P. motoro (Ali and Anctil, 1974). By analysing the opsin protein sequences in detail, and in particular the known tuning sites for all three photopigments, it was possible to predict accurately the spectral peak of absorbance (λ_max) for each opsin class. As such, λ_max values for lws, rh1 and rh2 photopigments were determined to be 461, 496 and 555 nm, respectively.

Akin to other cartilaginous fishes studied recently (Delroisse et al., 2018; Fasick et al., 2019; Hara et al., 2018; Hart et al., 2011; Theiss et al., 2012), neither the sws1 nor the sws2 opsin gene was detected in the retina of P. motoro, which suggests that these genes have been lost. This is also the case in the elephant shark chimaera (C. millii; Holocephali) (Davies et al., 2009a), as well as multiple shark species [e.g. the lanternshark, Etmopterus spinax (Delroisse et al., 2018), the whale shark, Rhincodon typus (Fasick et al., 2019), and the wobbegong sharks, Orectolobus sp. (Theiss et al., 2012)]. MSP also confirmed the lack of sws cones in both shark and ray species (Hart et al., 2004, 2011; Theiss et al., 2007). Taken together, these data strongly support the hypothesis that both sws1 and sws2 genes were lost very early in the Chondrichytes lineage, but after the split between the ancestor to all cartilaginous fishes and the rest of the gnathostomes ∼450–460 million years ago (Mya) (Inoue et al., 2010; Sansom et al., 1996).

Interestingly, the elephant sharks [via a speciation event that occurred around 4–6 Mya (Inoue et al., 2010), the origins of which derive from the evolution of the holocephalans about 374–420 Mya (Cappetta et al., 1993; Inoue et al., 2010)] and P. motoro [a relatively recent myliobatiform that evolved within the Batoidea (rays and skates), which diverged from the Selachii (sharks) ∼280 Mya (Inoue et al., 2010)] both retain two cone opsin genes (namely lws and rh2), as well as a rod rh1 photopigment gene, suggesting that the ancestral condrichthyan was likely to be dichromatic (Davies et al., 2009a). Despite similarities in the repertoire of visual opsins, the elephant shark is primarily a deep-sea species (200–500 m; Last and Stevens, 1994), whereas P. motoro dwells mostly in shallow freshwater rivers (Brooks et al., 1981; Costa et al., 2013). The two cone photopigments of P. motoro are predicted to exhibit spectral peaks of absorbance at 461 nm (for rh2) and 555 nm (for lws). In response to the filtering effects of greater depths of water on downwelling light, especially at the short-wavelength and long-wavelength extremes of the visible spectrum (e.g. Loew and McFarland, 1990), the photopigments in the elephant shark have short-wavelength shifted to 442 and 548 nm for rh2- and lws-based opsins, respectively (Davies et al., 2009a). It appears, therefore, that the cone photopigments of P. motoro have retained sensitivity to longer wavelengths, which are likely to be spectrally tuned for adaptation to a riverine habitat (Brooks et al., 1981; Costa et al., 2013) containing sand, mud or stones (Araújo et al., 2004; Rosa, 1985) that result in more turbid water compared with clear marine environments. Typically, freshwater species (e.g. many teleosts) will utilise porphyropsins that are based on the presence of a vitamin A₂-based chromophore that further shifts photopigment spectral maxima to longer wavelengths (reviewed in Davies et al., 2012; Yokoyama, 2000). However, the lack of cyp27c1 expression (which is required for the conversion of vitamin A₁-based 11-cis retinal into vitamin A₂-based 11-cis retinal; Enright et al., 2015) in the retina of P. motoro supports the presence of vitamin A₁-based photopigments only. The positive selection of rhodopsins over porphyropsins is unclear; however, it is also likely to be part of an overall adaptive strategy for living in freshwater. Indeed, as P. motoro is capable of living in clear and murky water environments, perhaps the presence of non-long-wavelength-shifted (rhodopsin) photopigments (but which are not further short-wavelength shifted as in the elephant shark) might be a compromise to maximise survival in a mixed photo-habitat.

Taken together, the data presented in this study strongly suggest that stingrays possess the potential for a dichromatic colour visual system, a hypothesis that was extensively tested by an array of behavioural approaches. The ability to discriminate between red and green was confirmed for all five individuals. Within an average of seven sessions, stingrays reached the first learning criterion (red versus green). These were followed by several types of transfer tests which together showed that learning was independent of brightness. Not only was red chosen over grey of various brightness, but any grey was also preferred over green (alternative or non-rewarded stimulus). The training stimulus red was lower in brightness than green. If animals had chosen according to brightness levels in the transfer tests, they should have picked any stimulus featuring a lower brightness level, which was not the case (see T3 and T4). In the second experiment (blue versus yellow), training stimuli did not vary in brightness and were still successfully discriminated. The results also indicate that the meaning of both the positive and the alternative stimulus had been learned. When faced without a ‘correct’ choice to be made in the transfer tests involving ‘grey’ versus ‘green’, for example, stingrays either chose the unknown stimulus (grey) or did not make a choice at all – a reaction that had previously been observed in other discrimination studies on cichlids and bamboo sharks (e.g. Schluessel and Duengen, 2015; Schluessel et al., 2012). It was also investigated whether and how well stingrays could distinguish the training stimulus red (featuring a characteristic brightness) from various other shades of the same or slightly varying colour. Here, group results varied from individual results. While group results clearly showed that the training stimulus was chosen significantly often over the alternative stimulus in all transfer tests, individual results showed a preference for the training stimulus in all but one instance, but this preference was not always significant. In fact, while all individuals chose Red 120 over Red 170, or Red 90, in most instances, none of them did to a significant degree. In case of Red 180, performance was variable, while Red 190 and 200, as well as Red 40–70, were always chosen significantly less often than Red 120. Obviously, the greater the difference between the two shades, the easier the differentiation was for the stingrays. It appears that contrast (aided in some cases by colour differences, see individual quantum catches) discrimination is worse in stingrays than in bamboo sharks, which were able to discriminate between shades of grey varying to a much lesser extent than the shades of red presented here (Schluessel et al., 2014). Considering that bamboo sharks are colour-blind, differentiation based on contrast may be of higher importance in this species. Nonetheless, comparison between the two studies may be compromised by the different colours that were used. In two studies by Ryan et al. (2016; 2017), contrast sensitivity was assessed in several shark species and was generally found to be high; nonetheless, interspecific differences were found. Overall, the current results show that stingrays easily differentiate between two stimuli of similar brightness, whilst shades of the same (similar) colour obviously need a certain amount of ‘variation’ to be recognised as different.

Only four animals participated in the second part (experiment 2) of the study. The learning criterion was reached after only 3.5 sessions (with 3 being the minimum possible). The drop-in sessions needed to reach criterion can be attributed to the stingrays now being familiar with the training procedure (two-alternative forced choice method) and/or to the task, i.e. the colour blue being more easily distinguished from yellow than red is from green. This is unlikely as blue and yellow had nearly identical brightness levels and transfer test results were less clear than in the first experiment. Surprisingly, in the transfer tests, blue was chosen significantly more often over a grey of slightly lower or greater brightness, but not over a grey of much lower brightness compared with blue (with blue and yellow having almost the same brightness). Possibly, animals remembered from the first experiment that the correct training stimulus (red) had a lower brightness than green and when confronted with the new transfer tests in experiment 2 were unsure whether to choose according to colour or brightness. Animals seemed to be ‘unsure’, which is reflected in the rays not choosing at all in 11 out of 78 trials (15%). In the second set of transfer tests, the (alternative) yellow stimulus was clearly identified as ‘incorrect’ in all tests and any grey stimulus was chosen significantly more often than yellow. As in experiment 1, this suggests that stingrays learned not only that blue was the correct stimulus but also that yellow was incorrect.

These results show that P. motoro can detect colour and are able to distinguish red from green, as well as blue from yellow, irrespective of brightness. Nonetheless, brightness levels are also perceived and used for differentiation. These findings are interesting from an ecological perspective, as they confirm the initial expectation that colour vision is advantageous for a diurnal species that spends at least some of its lifecycle in brightly lit and most probably clear water. These data are also consistent with the hypothesis that freshwater stingrays split some time during the Miocene from a marine ancestor (Lovejoy, 1996), which most certainly also possessed colour vision. Many fish species that dwell in bright light habitats possess multiple cone types and, therefore, complex colour visual systems (Marshall and Vorobyev, 2003). As such, photoreception has evolved for the execution of an entire range of behaviours related to mate choice and sexual displays, territorial defence, predator avoidance, and individual or species recognition. Some of these behaviours might be enhanced by colour vision in freshwater stingrays, e.g. freshwater stingrays possess ornamental colour pattern displays on their dorsal surfaces, which could potentially be used for courtship or individual recognition purposes.

All except one of the shark species (Heterodontus portusjacksoni) assessed by Hart et al. (2011) possessed duplex retinas. In the batoids, all ray species investigated were found to have two or even three cone types, while skates possess none; instead, they have all-rod retinas (Hart et al., 2004; Ripps and Dowling, 1990; Theiss et al., 2007). While the available data give reason to believe that sharks and skates are colour-blind and stingrays are not, it should be remembered that there is a large diversity of species within this group. Out of roughly 1200 species of elasmobranchs, only about 2% have been studied in any detail with regard to colour vision. So far, only three species have been assessed behaviourally, while lifestyle, habitat and ecology vary tremendously between different species. Despite the large biodiversity found within this group, the small number of species assessed so far means that generalised statements regarding the potential for colour vision in sharks and rays remains to be fully appreciated; however, our findings in stingrays presented in this study represents a significant progression towards this ambitious goal.

We would like to thank Slawa Braun for animal caretaking, maintenance and repairs, as well as Frankfurt Zoo for supplying the animals used for this study. The research reported here was performed under the guidelines established by the current German animal protection law. The experimental work presented herein was conducted by Friederike Seifert as part of her Bachelor of Science Project and by Christina Baumann as part of the Master of Science Project at the University of Bonn, Germany.