The visual ecology of Holocentridae, a nocturnal coral reef fish family with a deep-sea-like multibank retina

Vision in teleost fishes plays a crucial role in communication, prey detection, predator avoidance, habitat choice, and navigation (Marshall and Vorobyev, 2003; Cronin et al., 2014; Marshall et al., 2019). Because teleost fishes inhabit a broad range of environments with different light conditions and structural complexity (rivers, lakes, open ocean, deep-sea, coastal, coral reefs), have different activity patterns (nocturnal, diurnal, crepuscular) and diets (herbivory, carnivory, detrivory, planktivory), and display a variety of visually guided behaviours (courtship, communication, territorial, migratory), their visual systems had to adapt to meet these different visual demands (Walls, 1942; Collin and Shand, 2003; Warrant et al., 2003; Marshall et al., 2019; Carleton and Yourick, 2020).

These adaptations can be seen at many different levels of the visual system. At the ocular level, the shape and size of the eye and/or pupillary aperture affect the amount of light reaching the retina (Douglas and Djamgoz, 1990; Cronin et al., 2014), while filters present in the cornea or lens may modify the light spectrum before it reaches the light-sensitive opsin proteins located in the photoreceptor outer segments (Thorpe et al., 1993; Siebeck et al., 2003). At the retinal level, it is the type, size, number and distribution of the different neural cells that shape the visual system (Walls, 1942). The first level of visual processing is achieved by cone and rod photoreceptors (Lamb, 2013). Rods usually contain the highly sensitive rhodopsin protein (RH1) and mediate vision in dim-light conditions. Cones contain up to four different cone opsin proteins (short-, medium- and long-wavelength sensitive, SWS1/SWS2, RH2 and LWS, respectively) and mediate vision in bright-light conditions as well as colour vision (Yokoyama, 2008). In addition, rods and cones can vary in length and width (Ali and Anctil, 1976), and in the case of the cones, can be further divided into different morphological subtypes: single, double (two single cones fused together), triple or quadruple cones, although the last two types are relatively rare (Engström, 1963). The last level of retinal processing is performed by the ganglion cells, and their receptive field ultimately sets the upper limit of visual acuity as well as the optical sensitivity of the eye (Warrant and Locket, 2004). At the molecular level, it is the specific opsin gene repertoire, type of chromophore, as well as the level of expression of each opsin gene including the co-expression of multiple opsins within the same photoreceptor, that determines the spectral sensitivity of the photoreceptors and their capacity for colour vision (Hunt et al., 2014). All of these visual characteristics may differ between teleost species or even within the eye itself (intraocular and intraretinal variability) depending on the visual ecology, environment, and phylogenetic inertia of each species (Collin and Pettigrew, 1989; Cronin et al., 2014; Dalton et al., 2017; de Busserolles and Marshall, 2017; Stieb et al., 2019; Carleton et al., 2020).

Coral reefs are some of the most vibrant and colourful environments on the planet (McFarland, 1991). Consequently, coral reef inhabitants often have complex visual systems with well-developed colour vision capabilities. Colour vision in coral reef teleosts relies on the comparison of two to four spectrally different cone photoreceptors (di- to tetrachromatic), which may vary greatly in their spectral placement (Lythgoe, 1979; Marshall et al., 2019). This variability in spectral sensitivities seems to correlate, at least to a certain extent, with changes in the light environment due to season, habitat depth or ontogeny, and also due to differences in ecology and behaviour (Lythgoe, 1979; Shand, 1994b; Cortesi et al., 2016; Stieb et al., 2016, 2017; Tettamanti et al., 2019). Intraretinal variability in the distribution and density of photoreceptors and ganglion cells (i.e. retinal topography), has also been shown to vary with habitat structure (Collin and Pettigrew, 1988a,b), behavioural ecology (Stieb et al., 2019; Luehrmann et al., 2020) and ontogeny (Tettamanti et al., 2019) of coral reef fishes.

While vision in diurnal reef fishes has received substantial attention, the visual systems of nocturnal reef fishes remain understudied. From the few studies available, nocturnal reef fishes seem to have developed similar adaptations to other teleosts living in low-light environments (turbid/murky waters or the deep sea) in order to enhance the sensitivity of their eyes. These include: large eyes and pupillary apertures (Pankhurst, 1989; Schmitz and Wainwright, 2011), a smaller focal length (McFarland, 1991; Shand, 1994b), a tapetum lucidum (Nicol et al., 1973), a rod-dominated retina (Munz and McFarland, 1973; Luehrmann et al., 2020), longer and denser rods (Pankhurst, 1989; McFarland, 1991; Shand, 1994a), and an increase in the summation ratio of rods onto bipolar and ganglion cells (Shand, 1997). However, these observations are limited to few species and the opsin expression, retinal topography and most of the visual capabilities and visual ecology of nocturnal reef fishes remain understudied. A notable exception are members of the Apogonidae, which have recently been investigated in greater detail (Shand, 1997; Fishelson et al., 2004; Luehrmann et al., 2019, 2020).

To learn more about the visual world of nocturnal reef fishes, we focused our study on the Holocentridae, which comprises 91 recognised species divided into two subfamilies, the squirrelfish (Holocentrinae) and the soldierfish (Myripristinae) (Fricke et al., 2020). Holocentrids are found circumtropically and usually inhabit shallow coral reefs, although a few species, especially from the genus Ostichthys, occur in the deep-sea at depths of up to 640 m (Greenfield, 2002; Greenfield et al., 2017). While holocentrids are mainly active at night when they engage in feeding, they are also observed during the day hovering in or close to their refuges. Their large eyes (Schmitz and Wainwright, 2011) and ability to find their home after displacement (Demski, 2003) indicate that vision plays an important role in this family, although relatively little is known about their actual visual capabilities.

Genome mining in three species revealed that in addition to a single rod opsin, RH1, holocentrids possess several cone opsins: up to two SWS1 copies, two SWS2, up to eight RH2 paralogs and one LWS (Cortesi et al., 2015; Musilova et al., 2019). However, opsin gene expression and microspectrophotometry (MSP) in few species indicate that only a small subset of these opsins may be used in adult fishes (Losey et al., 2003; Musilova et al., 2019). Interestingly, their rod spectral sensitivity correlates with habitat depth, with deeper living holocentrids having shorter spectral sensitivities similar to those found in deep-sea fishes (λ_max=480–485 nm), shallower living species having longer sensitivities, comparable to those observed in other shallow-water fishes (λ_max=500–507 nm) and individuals living at intermediate depths having sensitivities somewhere in between (λ_max=490–495 nm) (Munz and McFarland, 1973; Toller, 1996; Yokoyama and Takenaka, 2004). Furthermore, the holocentrid ancestor is predicted to have had an RH1 sensitive to ∼493 nm λ_max suggesting that the family first emerged at intermediate depths (∼100 m or mesophotic depths; Yokoyama and Takenaka, 2004; Yokoyama et al., 2008; Musilova et al., 2019). This putative deeper origin, in addition to their nocturnal lifestyle on the reef and the few species inhabiting the deep sea, therefore make holocentrids particularly interesting for dim-light vision studies.

Using a range of techniques, including high-throughput RNA sequencing (RNA-seq), fluorescence in situ hybridization (FISH), photoreceptor spectral sensitivity estimates, and retinal anatomy and topography, we set out to scrutinise the visual system and visual ecology of several species of shallow water holocentrids, with the following two aims in mind: (1) to extend our knowledge about the visual ecology of nocturnal coral reef fishes; and (2) to assess if the holocentrid visual system differs from other nocturnal coral reef fishes because of their atypical ecological and evolutionary ties to deeper habitats.

Adult fishes from nine holocentrid species were investigated in the study. Most fishes were collected on the Great Barrier Reef around Lizard Island, Australia, under the Great Barrier Reef Marine Park Permit (G12/35005.1) and the Queensland General Fisheries Permit (140763) using spear guns on SCUBA in 2015 and 2016. Five specimens were obtained from the aquarium supplier Cairns Marine who collects fish from the Northern Great Barrier Reef (Cairns Marine Pty Ltd, Cairns, Australia). Each individual was anaesthetized with an overdose of clove oil (10% clove oil; 40% ethanol; 50% seawater) and killed by decapitation. Eyes were subsequently enucleated, the cornea and lens removed, and the eye cup preserved in different fixative solutions depending on the analysis (see below for details). All experimental procedures were approved by The University of Queensland Animal Ethics Committee (QBI/236/13/ARC/US AIRFORCE and QBI/192/13/ARC).

One eye of Sargocentron diadema, Neoniphon sammara and Myripristis murdjan was enucleated in daylight conditions and fixed in a solution of 2.5% glutaraldehyde and 2% paraformaldehyde in 0.1 mol l⁻¹ phosphate buffered saline (PBS). An extra individual of S. diadema was dark adapted for 2 h prior to euthanasia in the dark and one eye was enucleated and fixed as above. The retinas were dissected out of the eye cup and small pieces from different locations (dorsal, temporal, ventral, nasal, central) were post-fixed in 1–2% osmium tetroxide in 0.15 mol l⁻¹ PBS, dehydrated through an acetone series, and infiltrated with Epon resin (ProSciTech) using a Biowave tissue processor. Semi-thin transverse sections of the retinas (1 μm) were cut with a glass knife using a Leica EM UC7 Ultramicrotome and stained with an aqueous mixture of 0.5% Toluidine Blue and 0.5% borax. Sections were viewed with a Carl Zeiss Axio Imager compound light microscope and photographed using an Olympus DP70 digital camera. Retinal thickness, photoreceptor layer thickness and rod outer segment length were then measured from the photographs using ImageJ v1.52p (National Institutes of Health, USA). An average of three measurements per parameter were taken.

One retina from two Myripristinae (M. murdjan and Myripristisviolacea) and three Holocentrinae species (S.diadema, Sargocentronrubrum and Sargocentronspiniferum) were dissected out of the eye cup and preserved in RNAlater (Thermo Fisher Scientific) at −20°C until further processing. RNA extraction, library preparation and sequencing at the Queensland Brain Institute's sequencing facility followed the protocol outlined in Tettamanti et al. (2019) and Musilova et al. (2019).

Newly sequenced transcriptomes were combined with previously acquired holocentrid transcriptomes from three species: Myripristisberndti (n=4), Myripristisjacobus (n=2) and N. sammara (n=3) (Musilova et al., 2019), to complete our dataset for opsin gene expression analysis. Transcriptome filtering and de novo assembly followed the protocol described in de Busserolles et al. (2017). In brief, raw reads were uploaded to the Genomic Virtual Laboratory (v.4.0.0) (Afgan et al., 2015) on the Galaxy Australia platform (https://usegalaxy.org.au/). The quality of sequences was assessed using FastQC (Galaxy v.0.53) and sequences were filtered using Trimmomatic (Galaxy v.0.32.2) (Bolger et al., 2014) before being de novo assembled using Trinity (Galaxy v.0.0.2) (Haas et al., 2013).

Two different strategies were used to mine visual opsin genes from the transcriptomes. First, assembled transcripts were mapped against the opsin gene coding sequences extracted from the genomes of N. sammara and M. jacobus (Musilova et al., 2019) using the medium sensitivity settings (30% max. mismatch between transcripts) in Geneious v.9.1.5 and v.11.0.2 (www.geneious.com). Because assemblies based on short-read sequences tend to overlook lowly expressed genes and/or may result in hybrid transcripts, a second, raw-read mapping approach was also taken, as described in detail in de Busserolles et al. (2017) and Musilova et al. (2019).

Holocentrid opsin genes were scored for similarity to publicly available opsin sequences using BLASTN (https://blast.ncbi.nlm.nih.gov/Blast.cgi). Their phylogenetic relationship was then confirmed using a reference dataset obtained from GenBank (https://www.ncbi.nlm.nih.gov/genbank/). The combined opsin gene dataset was first aligned in MAFFT v.7.388 (Katoh and Standley, 2013) using the L-INS-I algorithm and default settings in Geneious. jModeltest v.2.1.10 (using AIC for model selection; Ronquist et al., 2012) and MrBayes v.3.2.6 (Ronquist et al., 2012) were then run on the CIPRES platform (Miller et al., 2010) to select the most appropriate model of sequence evolution and to infer the phylogenetic relationship between genes, respectively. We used the GTR+I+γ model, with two independent MCMC searches (four chains each), 10 million generations per run, a tree sampling frequency of 1000 generations, and a burn in of 25% to generate the holocentrid opsin gene consensus tree.

Opsin gene expression was subsequently calculated by mapping the unassembled filtered reads against the extracted coding sequences of each species-specific opsin repertoire (threshold of 2–3% maximum mismatch between reads; read count was normalised to the length of the coding sequence of each opsin), as detailed in de Busserolles et al. (2017) and Tettamanti et al. (2019). The expression of each cone opsin was calculated as the proportion of the total cone opsins expressed or, in the case of the rod opsin, as the proportion of RH1 compared to the total opsin expression. Because RH2 paralogs in the Holocentrinae showed high sequence similarity (>96% pairwise identity), RH2-specific reads were extracted and sub-mapped against high variability areas (100–200 bp in length). The proportional gene expression of RH2 paralogs was then re-calculated using normalised read counts from the sub-mapping approach.

Maximal absorbance (λ_max) of holocentrid visual pigments were estimated by translating opsin gene sequences into amino acid sequences using Geneious and assuming an A1-based chromophore as found in Sargoncentron spinosissimum (Toyama et al., 2008). Amino acid sequences were then aligned with bovine rhodopsin (NP_001014890.1) and reference sequences of well-studied model systems with known visual pigment spectral sensitivities using MAFFT (v.7.222) (Katoh et al., 2002). This allowed us to identify holocentrid specific opsin residues corresponding to known tuning and chromophore binding pocket sites according to the protein structure of bovine rhodopsin (Palczewski et al., 2000), and to infer holocentrid pigment spectral sensitivities based on sequence differences to each primary reference sequence: Oryziaslatipes RH1 (GenBank accession no.: AB180742.1) (Matsumoto et al., 2006); Oreochromisniloticus RH2B (JF262086.1) (Parry et al., 2005); O. niloticus RH2Abeta (JF262086.1) (Parry et al., 2005); O. niloticus SWS2A (JF262088.1) (Parry et al., 2005). For these, we focused on variable amino acid residues either at known tuning sites or at retinal binding pocket sites. Site effects were then inferred either for known substitutions or for substitutions that cause a change in polarity compared to the residue found in the primary reference sequence.

The following sites and effects for the different opsins were considered. RH1: E122M (−7 nm; all species; Yokoyama and Takenaka, 2004), F261Y (+10 nm; S. spiniferum, S. rubrum and N. sammara; Chan et al., 1992), A292S (−10 nm; S. spiniferum and S. rubrum; Fasick and Robinson, 1998), A295S (−4 nm; M. jacobus; Lin et al., 1998; Janz and Farrens, 2001).

SWS2A: I49V (−2 nm; all species; Yokoyama and Tada, 2003), A164S (−2 nm; S. spiniferum, M. violacea, M. murdjan, M. berndti, M. jacobus; Yokoyama and Tada, 2003), A269T (+6 nm; all species except M. jacobus; Yokoyama and Tada, 2003), L216F (−4/−8 nm upper/lower limit; all species). L216F in combination with M205I most likely explain the 8 nm difference between O. niloticus SWS2A (456 nm λ_max; Parry et al., 2005) and Pseudochromis fuscus SWS2Aβ (448 nm λ_max; Cortesi et al., 2015). Since all holocentrid sequences contained a methionine at residue 205, and as it is unclear whether L216F alone, M205I alone, or both substitutions together contribute to the 8 nm shift, we calculated upper and lower limit λ_max values, accounting for a 4 and an 8 nm blue shift, respectively, caused by L216F.

RH2A (only Holocentrinae): F60I/M/L/V (0/−1 nm upper/lower limit; all species; Yokoyama and Jia, 2020), Y74F (0/−1 nm; all species; Yokoyama and Jia, 2020), V255I (0/−1 nm; all species except N. sammara RH2A-1, S. spiniferum RH2A-2 and RH2A-3; Yokoyama and Jia, 2020), M259F/V (0/−1 nm; all species; Yokoyama and Jia, 2020), G273A (0/−1 nm; all species; Yokoyama and Jia, 2020). Substitutions at these sites are part of a site effect complex shown to cause strong red shifts in several teleost lineages compared with their ancestors (+21 nm; Yokoyama and Jia, 2020). However, the bulk of this shift is caused by the substitutions Y96T, Q122E and C213F, whereas the other substitutions may cause much smaller individual effects, if any. The substitutions observed at these five sites in Holocentrid RH2A opsins are inversions (either of the amino acids involved or of the polarity of substituted amino acids) and were therefore tentatively hypothesized to cause effects opposite to those described by Yokoyama and Jia (2020). For lower limit calculations we thus hypothesized each site, where present in the sequence, to cause a −1 nm blue shift.

RH2B (only Myripristinae): Y96T/Q122E/C213I (+20 nm; all species; Yokoyama and Jia, 2020), I49C/S and S109G combined (+8 nm; all species; Luehrmann et al., 2019).

After dark adaptation for 1 hour, the eyes of two N. sammara and one M. berndti were enucleated and prepared following previously described methods (Barthel and Raymond, 2000). Dual-labelling FISH was performed on whole mount retinas or quadrants of retina for very large retinas following standard protocols (Raymond and Barthel, 2004; Allison et al., 2010; Dalton et al., 2014, 2015). In brief, RNA was reverse transcribed using the High Capacity RNA-to-cDNA kit (Applied Biosystems). cDNA was then used to generate the probe template by standard PCR using the MyTaq™ HSRED DNA Polymerase (Bioline) and opsin specific primers (listed in Table S1) designed to bind to the 3′ untranslated region (3′UTR) (RH2A-1 and RH2A-2 in N. sammara) or the coding sequence (SWS2A in N. sammara, SWS2A and RH2B in M. berndti). Probes were then labelled with DIG or Fluorescein (Roche DIG/Fluorescein RNA Labeling Mix, Sigma Aldrich), tagged with Alexa Fluor 594 or 488 dyes (Invitrogen), and the signal enzymatically augmented with sequential tyramide signal amplification (TSA amplification kits, Invitrogen). Finally, retinas or retinal pieces were mounted, photoreceptor side up, on coverslips in 70% glycerol in PBS.

For visualization of labelled opsin genes, multi-channel scans for each dual-labelled opsin pair were performed using a spinning-disk confocal microscope consisting of a Nikon Ti-E (Nikon Instruments Inc.) equipped with a Diskovery spinning-disk platform (Spectral Applied Research), and Zyla 4.2 sCMOS cameras (Andor). NIS Elements (Nikon Instruments Inc.) were used to perform multi-channel imaging with a CFI Plan Apochromat VC 20x objective (NA 0.75, WD 1.00 mm), and a water immersion CFI Apo Lambda S 40× objective (NA 1.25, WD 0.18 mm) for high resolution images. All scans were exported as TIFs and further processed (merging of colour channels, adjusting of brightness) with ImageJ v.1.8.0_66 (National Institutes of Health, USA).

Eyes were fixed in 4% PFA in 0.1 mol l⁻¹ PBS (pH=7.4) for 48 h. Retinal whole mounts were then prepared according to standard protocols (Stone, 1981; Coimbra et al., 2006; Ullmann et al., 2011). The orientation of the retina was kept by referring to the position of the falciform process that ends ventrally for Holocentrinae and naso-ventrally for the Myripristinae. Each retina was bleached overnight at room temperature in a solution of 3% hydrogen peroxide in 0.1 mol l⁻¹ PBS.

For photoreceptor analysis, retinas were whole mounted (photoreceptor layer up) in 100% glycerol on a microscope slide. For ganglion cell analysis, the retinas were whole mounted, ganglion cell layer facing up, on a gelatinised slide, left to dry overnight in formalin vapour to improve fixation and cell differentiation (Coimbra et al., 2006, 2012) and stained in 0.1% Cresyl Violet (Coimbra et al., 2006). Possible shrinkage during staining was considered negligible and if present confined to the retinal margins, since the retinal whole mount remained attached to the slide throughout the staining process (Coimbra et al., 2006).

Different types of analyses were performed for high-density cell types (that is, single cones, double cones and ganglion cells) and low-density cell types (triple cones). Following the protocols described in de Busserolles et al. (2014a,b), topographic distribution of single cones, double cones, total cones and ganglion cells were assessed using the optical fractionator technique (West et al., 1991) modified by Coimbra et al. (2009, 2012). Briefly, using the parameters listed in Table S2 and a 63× oil objective (NA 1.40) mounted on a compound microscope (Zeiss Imager.Z2) equipped with a motorised stage (MAC 6000 System, Microbrightfield, USA), a digital colour camera (Microbrightfield) and a computer running StereoInvestigator software (Microbrightfield), cells were randomly and systematically counted. The counting frame and grid size were chosen carefully to maintain the highest level of sampling (∼200 sampling sites) and achieve an acceptable Schaeffer coefficient of error (CE <0.1; Glaser and Wilson, 1998).

Single cones and double cones were easily distinguished (Fig. 1) and counted separately and simultaneously using two different markers to generate data for single cones alone, double cones alone, and the two cell types combined (total cones). Owing to the low number of single cones present in the retinas of all holocentrids, the analysis for the single cones was repeated using a larger counting frame (Table S2).

In Holocentrinae, ganglion cells were arranged in a single layer in the ganglion cell layer. However, in Myripristinae, several ganglion cells were displaced and present in the inner nuclear layer. Since it was not always possible to confidently identify the ganglion cells present in the inner nuclear layer from the other cell types (amacrine and bipolar cells), only ganglion cells present in the ganglion cell layer were counted in this study. Furthermore, only ganglion cells were counted as they were easily distinguished from the other cell types present in the ganglion cell layer (amacrine and glial cells) using cytological criteria alone (Hughes, 1975; Collin and Collin, 1988).

Topographic maps were constructed in R v.2.15.0 (R Foundation for Statistical Computing, 2012) with the results exported from the Stereo Investigator Software according to Garza-Gisholt et al. (2014). The Gaussian kernel smoother from the Spatstat package (Baddeley and Turner, 2005) was used and the sigma value was adjusted to the grid size.

The distribution of the triple cones was mapped from one retina of S. rubrum using the Neurolucida software (MicroBrightField). The outline of the retinal whole mount was digitized using a 5× objective (numerical aperture, 0.13). The entire retina was then scanned in contiguous steps using a 20× objective (numerical aperture, 0.8), and each triple cone was marked. Results were exported from the Neurolucida software, and a dot map representing the location of each triple cone was constructed in R using a customized script based on Garza-Gisholt et al. (2014).

Holocentrids have a typical vertebrate retina organised in several layers: photoreceptor, outer nuclear, inner nuclear and ganglion cell layer (Fig. 2A,C). However, compared to most vertebrates that possess a single photoreceptor layer containing both rod and cone cells, holocentrids possess a multibank retina composed of one layer of cones and several layers of rods. In the three species investigated (from three different genera and the two subfamilies) this multibank organisation was found across the entire retina (Fig. S1). However, the number of rod banks varied in different areas of the retina and between species (Fig. 2, Fig. S1). Up to six and seven banks could be identified in N. sammara and S. diadema (Holocentrinae), respectively, while in M. murdjan (Myripristinae) up to 17 banks were observed. The highest number of banks in the Holocentrinae was found in the temporal and central areas whereas in M. murdjan the highest number of banks was found in the ventral area (Fig. S1). Indicative of a higher rod photoreceptor density, the outer nuclear layer (ONL) in M. murdjan also showed an increased thickness compared to the ONL in Holocentrinae (Fig. 2). In all three species, the rod outer segment length appeared to be uniform across all banks but varied between species. Overall the higher the numbers of banks, the shorter were the outer segments (∼15 µm in M. murdjan, ∼21 µm in S. diadema and ∼31 µm in N. sammara). However, at least for S. diadema, the width of the outer segment seemed to increase from layer to layer, with the first layer (B1, most scleral layer) having the thinnest rods (Fig. 2B). Consequently, in S. diadema, the first layer of rods (B1) had the highest density of cells and the last layer (B7) had the lowest. It is notable that for several sections from all three species, the rod nuclei in the ONL were arranged in vertical lines, as illustrated for M. murdjan in Fig. 2C.

Although rod-dominated, three types of cones (single, double and triple) were also found in the retinas of holocentrids (Fig. 3A). Double cones were the most frequent type, followed by single cones, while triple cones were rarely found. In Myripristis spp., double and single cones were not organised in a regular array or mosaic, but instead were arranged randomly throughout the retina (Fig. 1). In Holocentrinae, cone arrangements varied in different parts of the retina. In general, double cones were organised in regular rows, except in the temporal retina where the arrangement was more squared (i.e. double cones were positioned at an angle) and in the central retina where there was no apparent organisation. Single cones appeared evenly spread out throughout the retina although no obvious general pattern was observed, except in the nasal retina of N. sammara where they were arranged in a square pattern. When present, single cones in Holocentrinae were always placed in the middle of four double cones, as seen in the classic teleost square mosaic (Fig. 1C,D).

The holocentrid retina also showed clearly discernible photoreceptor and retinal pigment epithelium (RPE) retinomotor movements. In the light-adapted state (Fig. 2D,E), the cones and the melanin pigment granules within the RPE were positioned in the most vitreal part of the photoreceptor layer at the level of the first rod bank (Fig. 2D,E). Conversely, in the dark-adapted state, cones were positioned at the level of the third rod bank and the melanin pigment granules migrated all the way to the top of the last rod layer (Fig. 2B).

Transcriptomes from eight holocentrid species (four species per subfamily) showed that they predominantly express one rod opsin and either two (Myripristinae) or three cone opsins (Holocentrinae) within their retinas. Phylogenetic reconstruction identified these opsins to be RH1 (rod opsin; dim-light vision), SWS2A (blue-sensitive), and either RH2B-1 or two RH2A paralogs (RH2A-1 and RH2A-2; blue-green-sensitive) for Myripristinae and Holocentrinae, respectively. We were also able to extract a partial RH2B-2 sequence from the M. murdjan transcriptome and found evidence for a third RH2A-3 copy in S. spiniferum (Fig. 4A).

Quantitative opsin gene expression was highly similar within subfamilies with detailed values for each individual and species listed in Table S3. Opsin gene expression was strongly rod dominated; RH1 expression made up 99.45±0.08% (mean±s.e.m.) of the total opsin gene expression in Myripristinae and 94.92±0.97% in Holocentrinae. Within cone opsins, SWS2A (Myripristinae: 6.83±0.92%, Holocentrinae: 5.12±0.16%) showed much lower expression compared to RH2 genes (Myripristinae RH2B: 93.17±0.92%, Holocentrinae RH2A-1: 45±3.90% and RH2A-2: 49.70±3.91%). Finally, while RH2B-2 in M. murdjan was expressed at levels that were too low to reconstruct its full coding sequence, RH2A-3 in S. spiniferum made up ∼1.1% of its cone opsin expression (Fig. 4B; Table S3).

FISH was performed on the retinas of two holocentrid species, N. sammara (Holocentrinae) and M. berndti (Myripristinae). In both species, SWS2A expression was limited to single cones while RH2 opsins (RH2B-1 for M. berndti; RH2A-1 and RH2A-2 for N. sammara) were expressed in the double cones only (Fig. 5). Moreover, in N. sammara, the two RH2A paralogs were co-expressed in both members of the double cones (Fig. 5). These expression patterns were consistent throughout the retina of each respective species.

Since spectral sensitivity information for Holocentridae was sparse in the literature, especially for cones, and since transcriptomic data revealed the presence of several RH2 genes in Holocentrinae, we estimated the maximum absorbance (λ_max) of each opsin protein based on their amino acid sequence for the eight species from which retinal transcriptomes were available (Table 1). Predicted values were then compared with measured spectral sensitivities from the literature (Munz and McFarland, 1973; McFarland, 1991; Toller, 1996; Losey et al., 2003). Estimated λ_max of holocentrid RH1 pigments ranged from 491 nm in M. jacobus to 505 nm in N. sammara. RH1 estimations in all other species fell within this range with a predicted λ_max value of 495 nm. With the exception of M. violacea (λ_max MSP: 499 nm), these predictions fit well with previous MSP measurements of holocentrid rods (Munz and McFarland, 1973; McFarland, 1991; Toller, 1996; Losey et al., 2003).

Estimated spectral sensitivities for the SWS2A pigments ranged from 442/448 nm (lower/upper limit) in M. jacobus to 450/456 nm in N. sammara, S. diadema and S. rubrum. For S. spiniferum, M. violacea, M. murdjan and M. berndti, SWS2A λ_max were estimated at 448/454 nm. Compared with the available single cone MSP measurements (Losey et al., 2003), these predictions are all slightly long-wavelength shifted (λ_max MSP: M. berndti, 443/453 nm; N. sammara, 446 nm) (Table 1).

For all Myripristinae, RH2B-1 was estimated to be maximally sensitive at 500 nm. In M. berndti, this estimate was similar to the lower limit of the double cone spectral sensitivities measured by MSP (mean±s.d.: 506±2.4 and 514±1.4 nm) (Losey et al., 2003). Moreover, while MSP measurements suggested the presence of two spectrally different double cone types, the retinal transcriptome of M. berndti only contained a single RH2 opsin gene. In Holocentrinae, there was no or only very little difference (±1 nm) in the λ_max estimates of the different RH2A paralogs (λ_max=513–514/518 nm). These predictions were comparable to the N. sammara double cone λ_max measured by MSP (λ_max=512±3.0 nm; Losey et al., 2003) (Table 1).

The topographic distribution of ganglion cells and cone photoreceptors (double, single and total cones) was investigated in four Myripristinae and five Holocentrinae species from three different genera: Myripritis, Neoniphon and Sargocentron. Overall, while individuals and species within the same subfamily showed similar distributions of ganglion and cone photoreceptor cells (Figs S2 and S3) retinal topographies did differ between subfamilies. The retinal topography of each cell type for one representative species per genus is shown in Fig. 6 and detailed retinal topographies for the remaining species are provided in Fig. S3.

The ganglion cell distribution revealed subfamily specific specialisations (Fig. 6). Holocentrinae had a well-defined area temporalis that weakly extended along the horizontal meridian but did not reach the nasal area of the retina. Myripristinae, on the other hand, had a very large area centralis with a peak density in the ventral–temporal part of the retina. In addition, M. berndti and M. murdjan also had a horizontal streak (Fig. S3). The subfamilies also differed in their ganglion cell numbers and densities (Table 2). Holocentrinae had a higher total number of ganglion cells compared with Myripristinae, with an average of 830,000 cells and 580,000 cells, respectively. Ganglion cell peak densities were also much higher in Holocentrinae compared with Myripristinae, with densities ranging from 9510 to 23,786 cells mm⁻² and 2150 to 5990 cells mm⁻², respectively. Consequently, although Myripristinae usually had bigger lenses (Table S2), Holocentrinae had higher visual acuity estimates. Most Holocentrinae had an estimated spatial resolving power (SRP) of around 7 cycles per degree with the highest acuity recorded for N. sammara at 11 cycles per degree while all Myripristinae had an estimated SRP of around 4 cycles per degree (Table 2).

Similarly to the ganglion cell distribution, total cone topography differed between subfamilies (Fig. 6). Myripristinae had a strong horizontal streak, slightly oblique in orientation, with a peak cell density in the central to temporal area. Holocentrinae, on the other hand, had two areas located temporally and nasally, as well as a weak horizontal streak. Moreover, their peak cell density was found in the temporal area (Fig. S3) with the exception of S. rubrum where the peak cell density was found nasally. M. pralinia had an intermediate specialisation (one area and a weak horizontal streak, Fig. S3) between the one found for the remaining Myripristinae (Fig. S3) and the Holocentrinae (Fig. S3). Interestingly, for all Myripristinae, peak cell densities and topography patterns of total cone photoreceptors did not match the ones found for ganglion cells (Fig. 3, Fig. S3). In Holocentrinae, peak cell densities of total photoreceptors and ganglion cells matched pretty well even though the topography patterns were slightly different between cell types; the ganglion cell pattern was mainly defined by an area temporalis while the total photoreceptor pattern was characterised by two areas and a weak horizontal streak. Similarly to the ganglion cell numbers, Holocentrinae had more cones than Myripristinae with numbers ranging from 586,815 to 984,127 cells and 308,075 to 637,051 cells, respectively (Table 3).

Double cones were the main cone photoreceptor type found in the holocentrid retina, accounting for ∼87% of all cones (Table 3). As a result, the double cone topography matched the total cone topography for all species (Fig. 6, Fig. S3).

Single cones only accounted for ∼13% of the total cone population in the holocentrid retina (Table 3). The difference in single cone topography between the two subfamilies was the most pronounced of all neural cell types (Fig. 6, Fig. S3). In Myripristinae, the strong streak seen in the double cone topographies nearly disappeared and was replaced by a small area temporalis. In Holocentrinae, the single cone pattern was quite similar to the double cone patterns, with the two areas (temporal and nasal) and a weak streak. However, single cones were also more numerous in the ventral part, and interspecific variability was greater compared with the double cones in this subfamily (Fig. S3). With the exception of N. sammara, for which the single cone peak density was in the nasal area while the double cone peak density was in the temporal part, all Holocentrinae species had their single and double cone peak densities in the temporal part of the retina.

The mean double to single cone ratio (DC/SC ratio) across the retina varied between species from 4:1 in S. rubrum to 9:1 in M. berndti, with most species having a ratio of 7:1 (Table 3). The topography of DC/SC ratio was similar for all species with a higher DC/SC ratio in the dorsal part of the retina (Fig. 6, Fig. S3).

Triple cone density and distribution was assessed for one individual of S. rubrum, the species for which the most triple cones were observed across the retina. Triple cones for this individual only represented 0.5% of the total cone population. While triple cones were present throughout most of the retina, they were more concentrated along the horizontal meridian, in a similar pattern to the total cone distribution (Fig. 3).

The most striking feature of the holocentrid visual system is its multibank retina. While anecdotally mentioned by McFarland (1991), the multibank aspect of the holocentrid retina was not properly assessed or described. Multibank retinas are usually found in teleost species that live in dim-light conditions, mainly deep-sea fishes (Wagner et al., 1998) and a few nocturnal shallow or freshwater species (Gordon et al., 1978; Shapley et al., 1980; Hess et al., 1998; Omura and Yoshimura, 1999; Bozzano, 2003; Meyer-Rochow and Coddington, 2003; Omura et al., 2003; Taylor and Grace, 2005; Taylor et al., 2015). Within the shallow-water representatives, two are reef associated: the moray eel, Muraena helena, and the conger eel, Ariosoma balearicum (Hess et al., 1998). Therefore, the presence of a multibank retina in Holocentridae, a nocturnal coral reef fish family with a strong link to the deep sea (Yokoyama et al., 2008; Greenfield et al., 2017), while certainly unusual, may not be surprising.

Similarly to the majority of teleosts with multibank retinas, rods in holocentrids are organised in well-defined banks. In other species the number of banks may vary in different parts of the retina (Locket, 1985) and/or during ontogeny with banks added as fishes grow (Fröhlich and Wagner, 1998; Omura et al., 2003). While the number of banks seems to vary across the holocentrid retina, it is currently not known if banks are added ontogenetically, and at what stage/age the multibank starts to develop in the first place. Shand (1994b), studied the gross retinal structure of holocentrid larvae at settlement stage (i.e. after metamorphosis) and did not report a multibank retina. This suggests that the extra banks are added later, although a more in-depth study of the development of the holocentrid retina is needed to confirm this.

The number of banks found in the holocentrid retina (6–17) is high compared with other species. All shallow and freshwater species and most deep-sea fishes with multibank retinas only have 2–6 banks. To date, only five deep-sea species are known to possess more than six banks, with a record of 28 banks found in the deep-sea bigeye smooth-head, Bajacalifornia megalops (Locket, 1985; Denton and Locket, 1989; Fröhlich and Wagner, 1998; Landgren et al., 2014). However, in B. megalops, this extremely high number of banks is constrained to the fovea while the rest of the retina has 2–3 banks (Locket, 1985). Therefore, holocentrids, and especially species from the genus Myripristis, are part of a small group of fishes with an exceptionally high number of rod banks. Since the three species analysed in this study are shallow-water representatives, a comparison with the retinal structure of deep-sea holocentrids would be of particular interest.

Although common in deep-sea fishes, the function of multibank retinas is still poorly understood. Amongst the several theories put forward, two non-mutually exclusive hypotheses are common: (1) multibank retinas enhance the sensitivity of the eye by increasing the density and length of the rods (Wagner et al., 1998; Warrant et al., 2003); (2) they enable rod-based colour discrimination by changing the light chromatically as it passes through the different banks (Denton and Locket, 1989). Support for either hypothesis is lacking, mostly because of the difficulty to access live deep-sea specimens to conduct physiological and/or behavioural experiments. Therefore, holocentrids may offer an ideal model to address these questions as they can be readily accessed and trained in captivity. Their large eyes (Schmitz and Wainwright, 2011), short focal length (McFarland, 1991), rod-dominated retina (this study), extremely high RH1 expression compared to other coral reef fishes [i.e. >95% in holocentrids (this study) versus 50–70% in damselfishes (Stieb et al., 2016, 2019) and diurnal apogonids (Luehrmann et al., 2019)], and their rod spectral sensitivity tuned to their respective depth range (Toller, 1996), indicate a visual system well-adapted to dim-light vision and a need for increased sensitivity. Therefore, multibank retinas in holocentrids are likely to contribute toward enhancing the overall light sensitivity of their eyes. However, a use in colour vision during crepuscular hours or at night is also a possibility (Denton and Locket, 1989).

Colour vision relies on the opponent processing of light caught by a minimum of two differently tuned photoreceptors (Kelber et al., 2003). Since rods and cones usually function at different light levels and most vertebrates possess a single rod type, colour vision is often thought to be exclusively cone based and limited to bright-light conditions. However, dim-light colour vision based on modified cones or on a combination of cones and rods does exist and has been demonstrated in geckos (Roth and Kelber, 2004) and amphibians (Yovanovich et al., 2017), respectively. Moreover, several deep-sea fishes that possess multiple rod types (de Busserolles et al., 2017; Musilova et al., 2019) and/or a multibank retina (Denton and Locket, 1989) are likely candidates for purely rod-based colour discrimination. In the case of holocentrids, nocturnal colour vision may be achieved by their multibank retina, even with a single rod type, on the assumption that each bank acts as a spectral filter (Denton and Locket, 1989). Under this scenario, each bank is assumed to have a different spectral sensitivity and colour vision is made possible by comparing the outputs of the different banks (Denton and Locket, 1989). While this is possible in theory, only behavioural experiments combined with neurophysiological measurements will be able to attest whether holocentrids use their multibank retinas for dim-light colour discrimination.

In addition to their multibank retina, Holocentridae do possess several cone types and as such also have the potential for ‘classic’ colour vision during the daytime. Results from this study indicate that Holocentridae possess at least two spectrally distinct cone types and therefore are likely dichromats. All species investigated here possess single cones that express the SWS2A gene sensitive to the blue range of the spectrum, and double cones that express one or two RH2 genes, sensitive to the green range of the spectrum. While cone spectral sensitivity estimations made in this study were comparable to MSP measurements performed by Losey et al. (2003), some differences were observed, notably in M. berndti in which two different spectral sensitivities were found in double cones while only a single RH2B gene was expressed in the retinas of the fish from our study. The discrepancy between opsin gene expression and MSP data may be due to the different origin of the individuals studied; opsin gene expression was measured in fishes from the Great Barrier Reef while MSP was performed in fishes from Hawaii (Losey et al., 2003). Moreover, opsin gene expression, and by extension spectral sensitivities, provide a snapshot of the visual system of the animal at the time of sampling, a state that may change in teleost fishes over the course of the day (Johnson et al., 2013), between seasons (Shimmura et al., 2017) and at different habitat depths (Stieb et al., 2016). Therefore, it is possible that M. berndti individuals collected in February in Australia possess a different set of visual pigments to individuals collected in Hawaii in May–June. Notably, holocentrids do possess up to eight RH2 copies in their genomes (Musilova et al., 2019), but only a maximum of three were expressed at the same time in our dataset. Additional copies could therefore be used under different light settings or at different ontogenetic stages. Further in situ and experimental studies combining RNA-seq and MSP will be needed to further explore this. Finally, it is important to keep in mind that although a number of mutagenesis studies have demonstrated the effects of specific tuning sites in fishes (reviewed in Takahashi and Ebrey, 2003; Yokoyama, 2008; Yokoyama and Jia, 2020), the combined effects and/or interactions of multiple tuning sites with one another remain mostly unclear and may also explain the small differences in spectral sensitivities observed between measured and predicted data.

Unlike in other coral reef fish families that show high interspecific variability in cone opsin expression (Phillips et al., 2015; Stieb et al., 2016; Luehrmann et al., 2020), variability was very low in holocentrids from the same subfamily and did not seem to correlate with habitat partitioning. Between the subfamilies, Holocentrinae had a higher proportion of cone opsin expression compared with Myripristinae and expressed two RH2A paralogs compared with the single RH2B copy expressed in Myripristinae. Since the two RH2A genes in Holocentrinae were found to be co-expressed in both members of the double cones and are also predicted to have similar spectral sensitivities, the advantage of expressing two copies over a single RH2 gene is intriguing. Certainly, the higher level of cone opsin expression in Holocentrinae does suggest that they rely on their photopic visual system more than the Myripristinae. However, whether holocentrids can discriminate colour during the day remains to be investigated. The large size of their cone photoreceptors compared with diurnal species (Munz and McFarland, 1973) may also increase sensitivity to lower light conditions and allow for cone or a mixture of cone and rod-based colour vision in dim-light conditions (Hess et al., 1998). Consequently, their large cone photoreceptors and multibank retinas may enable holocentrids to perceive colour in a wide range of light intensities.

Retinal ganglion cell topography is a powerful tool in visual ecology, highlighting areas of high cell density and therefore high acuity in a specific part of the visual field. In teleost fishes, several studies have shown a strong link between the retinal topography pattern and the habitat and/or behavioural ecology of an animal (Collin and Pettigrew, 1988a,b, 1989; Shand et al., 2000; Collin, 2008; de Busserolles et al., 2014b). In holocentrids, ganglion cell topography showed very little intraspecific and interspecific variability within subfamilies. However, density and topography patterns differed between the two subfamilies. While both subfamilies share similar habitats (Gladfelter and Johnson, 1983), they differ in feeding ecologies, with Holocentrinae mainly feeding on benthic crustaceans and species from the genus Myripristis (Myripristinae) feeding on large zooplankton in the water column (Gladfelter and Johnson, 1983; Greenfield, 2002). Accordingly, an area temporalis that extends into a weak horizontal streak may allow Holocentrinae to scan and detect crustaceans situated in front of them, on or close to the sea floor. In Myripristinae, a large area temporo-ventralis that provides higher acuity in front and above of them, may instead facilitate the detection of zooplankton when seen against the background illumination. A similar ganglion cell topography and relationship with feeding mode was also observed in the nocturnal apogonids (Luehrmann et al., 2020).

In addition to the topography pattern, spatial resolving power (SRP) also differed between the two subfamilies with Holocentrinae having higher acuities than Myripristinae. However, this result has to be interpreted carefully since Myripristis spp. had a large population of displaced ganglion cells that was not included in the analysis, potentially resulting in an underestimation of their SRP. Displaced ganglion cells have been described in many vertebrates and shown in several cases to be part of the accessory optic system (Simpson, 1984). As such, all or some of these displaced ganglion cells are likely to have a different function to the ones found in the ganglion cell layer and may not contribute to visual acuity. Future labelling and tract tracing experiments will be needed to elucidate the function of the displaced ganglion cell population in Myripristinae. Moreover, while displaced ganglion cells were not obvious in Holocentrinae, their presence/absence was not studied here and will need to be assessed further. Regardless, holocentrid visual acuity was relatively low compared with that of diurnal reef fishes with similar eye sizes (e.g. the SRP of M. murdjan and S. rubrum is 4.38 and 7.35 cpd, respectively, versus an SRP in Lethrinus chrysostamus of 22 cpd; Collin and Pettigrew, 1989). However, Holocentrids did have a similar SRP to the nocturnal apogonids (∼7 cpd; Luehrmann et al., 2020). Since apogonids are much smaller fish with much smaller eyes, this suggests a higher eye investment in visual acuity for apogonids compared to holocentrids. Conversely, the large eyes in holocentrids (Schmitz and Wainwright, 2011) coupled with a short focal length (McFarland, 1991), is likely an adaptation to increase the overall sensitivity of the eye rather than its acuity.

Photoreceptor cells constitute the first level of visual processing and as such their density and distribution provide important information about the visual demands of a species. Even though the holocentrid retina is rod-dominated, only the density and distribution of cone photoreceptors could be assessed in this study owing to the presence of the multibank retina.

In diurnal teleost fishes that have a cone-dominated retina, cones are generally arranged in a regular pattern or mosaic. Conversely, nocturnal and bottom-dwelling fishes that have a rod-dominated retina tend to have disintegrated cone mosaics (Engström, 1963). Accordingly, holocentrids, and especially Myripristinae, were found to have a mostly disintegrated cone mosaic that fits with their nocturnal activity pattern. Holocentrinae, however, did have a more organised cone arrangement, especially in the temporal part, the area with the highest visual acuity. This, in addition to their higher cone densities, cone opsin expression and visual acuity suggest that the Holocentrinae visual system is better adapted for photopic conditions than the visual system of the Myripristinae.

Similarly to the ganglion cells, photoreceptor topography may be used to identify areas of the visual field that are ecologically meaningful for a species. While photoreceptor and ganglion cell topography usually match and have peak cell densities in the same region of the retina, variations do exist, and may indicate different visual demands in different parts of the visual field of the animal or at different times of the day (Stieb et al., 2019; Tettamanti et al., 2019). In Holocentrids, total cone and ganglion cell topographies differed, especially in Myripristinae. Since holocentrids are nocturnal fish and have a rod-dominated retina, it is likely that the topography and peak density of their rod photoreceptors matches that of their ganglion cells, as seen in some deep-sea fishes (de Busserolles and Marshall, 2017). This is supported by the highest number of banks being located in the area of the highest ganglion cell density in all three species. Unfortunately, limited information is available about holocentrid daytime activities. During that time, they appear to hover in or above their refuges and may partake in some social interactions such as courtship, aggression and predator avoidance (Winn et al., 1964; Carlson and Bass, 2000). Accordingly, a horizontal streak may allow them to scan a wide area of their visual field to look for possible intruders or conspecifics while staying within the safety of their refuges, as suggested for the highly territorial anemonefishes (Stieb et al., 2019). Moreover, a high density of cells in the nasal area, as seen in Holocentrinae, may also help in detecting predators coming from behind (Collin and Pettigrew, 1988a).

For all holocentrids, but especially in the Myripristinae, single and double cone topography differed, suggesting that the different types of cones may be used in different visual tasks. While it has been demonstrated that both single and double cones are used for colour discrimination in the coral reef Picasso triggerfish, Rhinecanthus aculeatus (Pignatelli et al., 2010), in many other species this is not clear (Marshall et al., 2019). Since the spectral sensitivity of double cones often matches the spectral distribution of the ambient/background light they may be used in luminance detection tasks (McFarland, 1991; Marshall et al., 2019; Carleton et al., 2020). In holocentrids, this idea is further supported by the ratio of double to single cones which was consistently higher in the dorsal retina. If double cones are indeed used in achromatic tasks, having a higher proportion in the dorsal retina, the area that samples light in the field of view below the fish where background illumination is lower, might assist in increasing sensitivity. However, behavioural tests will be needed to confirm this.

Holocentridae have a visual system that is well-adapted for their nocturnal lifestyle with large eyes, short focal length, rod-dominated retina, multibank retina, extremely high rod opsin expression, rods tuned to their preferred light conditions, few cone opsins that are expressed at low levels, few cone photoreceptors and relatively low visual acuity. Moreover, the fact that the holocentrid ganglion cell topography correlates with their feeding mode, a task which in this family is exclusively conducted at night, further supports their heavy reliance on their scotopic visual system. The presence of at least two spectrally different cone types with their own topography patterns also indicates the use of their cone-based visual system during the day and the potential for dichromacy. Moreover, while interspecific variability was very low within the family, differences in visual adaptations could be seen between the two subfamilies at all levels with Holocentrinae having a slightly more developed photopic visual system compared with Myripristinae. Finally, what really sets the holocentrid family apart from other coral reef fishes is their well-developed multibank retina, an adaptation mostly found in deep-sea fishes, and their potential for colour vision in a wide range of light settings, especially under scotopic conditions. Future ontogenetic and behavioural analyses should therefore be conducted in order to understand the origin and function of the multibank retina, as well as to assess whether this family is able to discriminate colours and under which light intensities. Additionally, investigation of other teleosts with intermediate depth ranges, such as mesophotic species, are likely to reveal interesting adaptations for dim-light vision.

We would like to acknowledge the Traditional Owners of Lizard Island, The Dingaal Aboriginal people. We thank Cairns Marine for supplying fish and the staff at the Lizard Island Research Station, Lorenz Sueess and Eva McClure for support during field work. We also thank Janette Edson from the Queensland Brain Institute (QBI) Genomics Facility for library preparation and transcriptome sequencing, Rumelo Amor from the QBI Advanced Microscopy Facility for technical support, Zuzana Musilová from Charles University (Czech Republic) for help with FISH experiments, Helen Cooper and Michael Langford (QBI) for providing lab facilities to conduct FISH experiments and the two anonymous reviewers and editor. The stereology were performed at the QBI Advanced Microscopy Facility using Stereo Investigator, which was supported by an ARC LIEF grant (LE100100074).