Eye enlargement.

This transition, which preceded the emergence of well-developed limbs and digits [14], is chiselled into the fossil record: Within only a few million years, early lobe-finned fish transitioned from a swimming form with small eyes, to an “underwater walking” form with large eyes (Fig 2C). Our last fully aquatic ancestors probably lived in the shallows but peeked out of the water in search of food. Invertebrates had colonised the land some 50 million years prior [23], and with no vertebrate predators around, their initial lack of defences [24] would have made for an easy meal. While tetrapod eye enlargement was therefore likely an adaptation for a crocodilian-like lifestyle, it opened the floodgates for much to come.

Shattering the cone mosaic.

Many surface dwelling fish have a crystalline cone mosaic with fixed photoreceptor-stoichiometry of 2:2:1:1 for red, green, blue, and UV, and with rods slotted around UV cones [4,17,18] (Fig 2D, left). In species where subsets of cone types are missing the mosaic rearranges accordingly to close any gaps [17,25]. The functional purpose of this crystalline mosaic remains poorly understood; however, it may confer important computational advantages, perhaps not too dissimilar to some of those conferred by the crystalline organisation of crustacean eyes [26], which notably includes insects [27–29]. Conceivably, the mosaic might also be exploited to reliably match “driving” and “regulatory” cone circuits (cf. Fig 1B). In support, in some fish the crystalline lattice propagates onto subsets of horizontal and bipolar cells [17,30].

However, the crystalline solution of fish does not work in air. First, it would lead to spatial oversampling in the periphery, because unlike in water, terrestrial optics cannot simultaneously maintain perfect focus across the entire back of the eye [31]. Second, due to greatly reduced photon scatter and absorption in air, everything is visible with high contrast for as far as the eye can see. Sampling such an environment with a regular cone lattice leads to aliasing [32], which would in turn incapacitate motion circuits. This problem is long recognised in digital photography where it is solved by introducing blur; however, this solution is energy inefficient. In biology, a better option is to break the crystalline arrangement in favour of a locally randomised pattern (Fig 2D, right). This also opened up the possibility to discard the fixed cone stoichiometry in favour of distinct and locally flexible cone ratios that best serve an animal’s specific visual ecology—as is the case in all extant tetrapods [3,20].

And yet, one mystery remains: Extant lobe-finned fishes, the closest living relatives to the aquatic ancestors of tetrapods, have no crystalline mosaic [17,33,34]. The arrangement is also not found in the “preceding” cartilaginous fish or in lampreys [17,35]. The crystalline pattern may therefore be a clade-specific trait of ray-finned fishes that is not ancestral to the tetrapod eye. Alternatively, the crystalline arrangement could have evolved in the common ancestor to all teleosts but have subsequently been lost in extant lobe-finned species, of which there are only 2 small lineages: coelacanth, who live in the deep where light is limiting [36] and where the mosaic breaks also in ray-finned fish [37], and lungfish [33,34], who like early tetrapods routinely peek out of the water.

New photoreceptors.

At least 2 new sets of photoreceptor types emerged in early tetrapods: “double cones” [11] and “blue rods” [21,38,39]. Both are present in amphibians but absent in fish, suggesting that their emergence coincided with the first major presence of vertebrates on land. Correspondingly, their purpose is probably related to exploiting new opportunities that presented themselves above the water surface.

Of the 2 new photoreceptor systems, double cones probably appeared first. They exist in amphibians [39], reptiles [40,41], birds [11,22], monotremes [42–44], and marsupials [44–46], and therefore probably appeared in a common ancestor to all these lineages: the first tetrapods (Figs 2B and 1C, cf). They are again missing in eutherian mammals [3]. By contrast, blue rods are only present in amphibians [21,38] and therefore probably emerged after their lineage had already diverged from that of subsequent terrestrial species. The origin of double cones is debated, but their anatomical similarity to pairs of ancestral red and green cones in the eyes of fish [18] points at their joint duplication as a likely candidate. Pairs of teleost red and green single cones are also sometimes referred to as double cones [47]; however, this is a historical misnomer: Fish only have 1 set of these cones, while non-eutherian tetrapods have 2: the ancestral set of red and green single cones, plus an independent population of double cones.

Duplicating red/green circuits?

Double cones are made up of 2 tightly associated photoreceptors, called the principal and accessory member [11]. In line with its putative origin from the red single cone, the principal member expresses the red LWS opsin, but the same opsin is also usually [48] found in the accessory member [22]. The latter may have resulted from an opsin expression switch from RH2 to LWS. The 2 members of the double cone also differ in their oil droplets [49,50] (see below).

Beyond inheriting opsins and morphological characteristics, any newly duplicated cones would presumably also have inherited their ancestors’ postsynaptic circuitry. This would have opened the option to share or take over some of their original functions. In agreement, just like red and green single cones, the 2 members of the double cone feed into mutually independent horizontal and bipolar cell circuits [11]. Conceivably, the ancestral one-size-fits-all strategy of using red cones to drive most of vision could therefore now be broken into 2 independent circuits, each with their own regulatory system. This in turn would have enabled individual specialisation of the 2 circuits for different sets of visual tasks. One of these tasks, at long last, might have been the emergence of a dedicated system for colour vision as we think of it today.

Co-option of ancestral single cones for terrestrial colour vision?

This view is supported by observing phylogenetic patterns in photoreceptor oil droplets—small optical elements that are positioned immediately in front of cones’ light sensitive outer segments [49]. Oil droplets are common in vertebrates, from lobe-finned fish to marsupials, and again absent in eutherians. Moreover, some ray-finned fish have ellipsosomes, which though developmentally distinct, appear functionally reminiscent of oil droplets [51]. In general, oil droplets can serve many functions, but one of them is spectral sharpening [49,52,53]. By mixing coloured pigments into the oil, the droplets can be used as filters that restrict a cone’s spectral sensitivity. This limits spectral overlap between cones and thereby increases colour resolution. However, this sharpening comes at the cost of reduced light sensitivity, which means that it is not possible to simultaneously optimise the same photoreceptor for high signal to noise greyscale vision and for spectrally narrow colour vision. Prior to the emergence of double cones, this trade-off would have precluded ancestral red cones from using oil droplets (or ellipsosomes) in this way, and correspondingly, all teleosts [49] except lungfish [33,34] use them for non-spectral tasks. Conversely, oil droplets are often heavily pigmented in red cones of amniotes, while double cone droplets remain at most weakly pigmented, and sometimes absent [49,52,53]. Together, this hints that double cones, led by the principal member and regulated by the accessory member, took over some of the original roles of the long wavelength system in supporting high acuity spatiotemporal tasks [54]. This could have freed up the use of single cones for spectral vision.

In agreement, most diurnal birds retain the full complement of ancestral single cones, and these cones’ interplay with oil droplets leads to the probably most highly resolved colour vision system among vertebrates [22,49,52,53] (Fig 3A and 3B). Double cones are theoretically not needed for this task [54]. Among single cones, any surviving opponency of ancestral green and blue circuits [55] would no longer be overly useful for estimating distance in air [1,10,56], and similarly, their requirements for regulating motion circuits likely changed [57]. This would have opened the possibility to co-opt some of their existing, spectrally nuanced retinal and central circuits for new functions, such as colour vision. Likewise, some of the ancestral functions of UV cones, such as prey capture, are less applicable in air, and probably opened further options for spectral specialisations. A decreased dependence on single cones for supporting achromatic vision, alongside the no longer fixed stoichiometry as necessitated in fish, would have allowed retinal circuits to reduce their relative abundances, including of their postsynaptic circuits. For example, in chicken only around 20% of cones are ancestral blue- or UV cones [19], and these feed into fewer than 10% of bipolar cells [11], down from more than half in zebrafish [58]. This reduction in short wavelength contributions to retinal circuits is pervasive across terrestrial species and taken to the extreme in our own eye: Fewer than 0.1% of photoreceptors in the human retina are ancestral UV cones, and only a single type of bipolar cell specifically represents their signals [59,60].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Retinal circuits of nonmammalian tetrapods.

(a) Conceptual retinal wiring schema of diurnal birds and reptiles, illustrating how double cones might fit into the circuit blueprint inherited from fish (cf. Fig 1B). While double cones might partially take over greyscale functions of the ancestral red system, existing and new interactions between single cone circuits might have increasingly been coopted for colour vision. (b) Spectral sensitivity functions of chicken cones with oil droplets included (based on ref [125]). (c) Mean spike rates of 2 functionally identified ganglion cell clusters from poultry chick retina in response to a time-accelerating full field chirp stimulus as indicated (mod. from ref [65]). (d) As (a), but for frogs, who lost ancestral green cones (RH2) but evolved “blue rods.” Potentially, blue rods serve functions similar to those served by ancestral blue cones of fish, however at lower light levels.

https://doi.org/10.1371/journal.pbio.3002422.g003

Double cones took over (some of) achromatic vision.

While single cones potentially specialised for colour vision, double cones may have specialised for high acuity spatiotemporal vision [54,61]. In chicken, they co-wire with rods [11], the ancestral remit of red single cones [58,62]. Double cones are usually also the largest and most numerous cone types, which maximises their potential for temporal and spatial resolution, respectively [63]. In support, birds generally have some of the fastest eyes of any vertebrate [64], and the spectral tuning of birds’ fastest retinal output channels (Fig 3C) is consistent with a primary drive from a red opsin expressing cone [65]. However, the extent to which these fast channels are driven by red single and/or double cones remains unknown. It is also unclear if the principal and accessory members of the double cone work jointly (as suggested by their intimate coupling), in mutual opposition (as hinted by their putative ancestry), or both. Despite their pervasive presence across vertebrates, from frogs [39] to kangaroos [44], systematic data on the function of double cones and their downstream circuits remains close to nonexistent.

Moreover, despite the newfound potential for differentially specialising red circuits, some important aspects of ancestral red single cones for driving achromatic vision must have been retained: First, diurnal raptors have the highest spatial resolution of any vertebrate, but in most species, this is achieved without the use of double cones which are systematically absent from their foveae [66]. Second, eutherian mammals lost the double cone, but they can see perfectly fine. If double cones would have completely taken over general-purpose greyscale vision in the ancestors of eutherian mammals, our own eyes today would probably be covered in double cones, not red single cones.

Extending red regulation into the night.

While regulation of red and UV functions by mid-wavelength circuits may work well during the day (Fig 1B), the strategy has 1 major flaw: Any regulation would cease as soon as rods take over, because most species have no “second rod” to regulate the first one. In fish, this absence for the potential to regulate rods may in part reflect a compromise driven by the reduced light levels in the water compared to in air. As photons become sparse, visual systems increasingly struggle with signal to noise more than they do with the issues that putatively drove the evolution of regulatory systems in the first place [67,68]. However, with the appearance of early amphibians, this pressure may have lifted: Amphibians must live near or above the water surface where there is more light, including at night. They also have proportionally larger eyes compared to most surface-dwelling fish that share parts of their habitat [69], and their rods are some of the largest among vertebrates [21]. All these factors contribute to improving night vision, and this in turn might have opened a path towards rod-regulation (Fig 3D), much like vertebrates’ earliest fully aquatic ancestors might have done some 300 million years prior for the cone system. Moreover, unlike most other tetrapods, amphibians routinely return to the water where the previously discussed spectral features [1] still apply.

The origin of amphibian “blue-rods” remains debated [21,70]. Its rod-like features (e.g., [71]) hint that it might reflect a duplication of the ancestral “green” rod, while its expression of the SWS2 opsin instead points to its duplication from blue cones. In general, cones can adopt rod like features over time—even the ancestral “green” rod (RH1) may represent an ancient duplication of green cones (RH2) [72]. Similarly, some snake [73] and gecko [74] photoreceptors are thought to have switched back and forth between rod- and cone-like features as their lineages iteratively adopted diurnal and nocturnal lifestyles. Paralleling the possible duplication of the red-green system to yield double cones, a blue cone origin of “blue rods” would also be in line with their possible regulation of green rod circuits (Fig 3D). However, as with double cones, the purpose and functional substrate of “blue-rod” circuits remain sparsely explored. Nevertheless, anatomical work hints that blue rods do make outer retinal connection that differ from those of ancestral rods [21,39], and the 2 rod systems can be differentially read out at the level of behaviour [70].

Together, the 2 new major photoreceptor systems that emerged well-timed with the first long-term emergence of vertebrate life on land might represent a near-complete duplication of the full ancestral cone complement. That is, except for UV. Plausibly, there was little selection pressure for duplicating this system because the background subtraction effect offered by aquatic UV-vision no longer applies in air. Accordingly, UV cones would have become less useful for supporting basic figure ground segmentation tasks including prey capture. Their utility in threat detection, however, would remain [75,76].

Revisiting the nocturnal bottleneck.

From early jawless species via fish and amphibians to present-day reptiles and birds, vertebrate retinas have increased considerably in their capabilities over their more than 500-million-year histories (Fig 2B). And yet, much of this ancestral accumulation of complexity is lost in our own lineage. Unlike even our still relatively close marsupial “cousins,” retinas of eutherian mammals have shed most of the early inventions. Gone are 2 of the original 4 ancestral single cones, alongside both double cones, and oil droplets are not to be found anywhere from mice to men. Even ancestral UV cones, though present, appear to play an at best peripheral role in shaping the visual world of eutherians [16].

The series of loss events detailed in the above have long been linked to Walls’ “nocturnal bottleneck theory” [77]. With the appearance of diurnal predatory dinosaurs some 240 million years ago, early ancestors of today’s eutherian mammals were forced into a nocturnal niche. At night, less light is available for vision, and this would then have gradually led to the loss of all 4 photoreceptor systems (both double cones as well as green and blue single cones), and oil droplets along with it. However, it might be time to revisit this idea. The only major mammalian lineage to have systematically lost these photoreceptors are eutherians, which diverged from marsupials only some 170 million years ago [78]. However, marsupials retain both double cones and oil droplets [45,46,79], indicating that these features must have been retained in the last eutherian ancestor for at least 70 million years of previous nocturnalisation. Similarly, monotremes, diverging some 220 mya, retain the ancestral blue cone [42,43]. Accordingly, blue cones were retained for at least 20 million years of nocturnalisation. Green cones present a further conundrum. The absence of the green cone’s RH2 gene in the genome of all mammals, including monotremes and marsupials, suggests that they were lost early on [80]. And yet, some marsupials including dunnarts have a “third” unaccounted for mid-wavelength sensitive cone [45,46,81]. The identity and origin of this “extra” cone remains debated among 2 perhaps equally plausible options: Either, this is the ancestral green cone, but now expressing a non-RH2 opsin, or it is the result of yet another cone-duplication. In either case, already the retention of blue cones for at least 20 million years of sustained nocturnalisation, and of double cones plus oil droplets for an additional 50, poses a serious challenge for the nocturnal bottleneck as the main driving force for cone loss in the ancestors of eutherian mammals. Instead, it seems reasonable to suggest that other factors contributed, at least in part. One of these factors might be related to their “other half” of vision: the brain itself (cf. Fig 1C)!

Eutherian mammals may use a different retinal “code”.

Mammals have some of the largest brains relative to body size of all vertebrates, with eutherians somewhat outcompeting marsupials and monotremes [82,83]. The concomitant expansion in brain complexity might well have allowed emerging central circuits to take over some of the more complex computations that were previously supported by the retina. Over evolutionary time, and particularly upon their emergence on land, vertebrates’ increasingly complex visual requirements perhaps inevitably outran the increasing capacity of retinal circuits to deliver an immediately actionable signal for brain circuits to use. Compared to our earliest sighted ancestors, some degree of centralisation of key visual tasks has probably occurred in all extant vertebrates. Mammals, however, seem to have embarked upon this path more than any other lineage. In eutherians, vision has pivoted to a point where it is driven almost exclusively by ancestral red cones and rods, the original “general purpose system” of the vertebrate eye. From a retinal perspective, this strategy is relatively simple, general purpose, and probably efficient. After all, vertebrate retinal complexity—with a pinnacle probably in birds—is perhaps not the only option for building a powerful and versatile visual system.

Based on our still very limited understanding of retinal coding outside of mammals, the “visual way” of eutherians appears to be diametrically opposite to that of most fish and their diverse non-eutherian descendants including birds [65]. Eutherians tend to represent On and Off or fast and slow signals via parallel streams, with only a minority of neurons occupying the coding space in between these extremes [65,84–87]. Conversely, in chicken at least [61], the output from the avian retina is dominated by a highly correlated “compound code” that routinely mixes On and Off as well as fast and slow signals within individual ganglion cell axons [65]. Overall, the eutherian retina strategy therefore appears to be one of decorrelation [88,89], likely to achieve coding efficiency by keeping axonal firing rates low [90–94]. In birds, firing rates are probably higher on average, which hints their retinal code is not very energy efficient [65]. It is, however, space efficient, in the sense that multiple messages can be “multiplexed” via common axons. Conceivably, this avian strategy is a consequence of their anatomically much more dense retinal input-output organisation and enabled by a 3-fold lower energy consumption per neuron compared to those of mammals [95].

As for eutherian mammals, nowhere is their “spring-cleaned” retinal output strategy more obvious than in the case of humans and our closest nonhuman primate relatives. More than 80% of the retinal fibres leaving our eyes, leading to a probably even greater fraction of our conscious visual experience, come from just 2 pairs of projection systems: midgets and parasols [60,96,97]. Midgets are numerous and tiny and provide exquisite spatial resolution, while parasols are larger which allows them to pool cone signals, ultimately serving fast temporal resolution. The “code,” delivered to the cortex via the thalamus, is exceptionally general and flexible [91,98–101]. Even colour-signals are delivered via this route, and not using ancestrally distinct sets of cones. In humans, green and red signals are provided by the same ancestral red cone, with the tweak that some red cones randomly express a green-shifted version of the red opsin, while others express the “original” red version [59,102] (Fig 4A, left). During infancy, we learn the statistics of the consequently red- and green-biased signals as they come into the brain [103,104], and the resultant spectral performance in red-green discrimination appears to be second to none [98,105,106]. A fifth “major” type of retinal projection neuron, the small bistratified ganglion cell [59,107,108], additionally delivers spectrally opponent signals derived from contrasting ancestral red and UV cones, thereby essentially completing our trichromatic experience (Fig 4A, right). Beyond those 5 [97], other retinal output channels do exist, including some that resemble versions of ancestral circuits that once were probably fundamental to vertebrate’s success on this planet [59,109,110]. However, in the human eye, their abundance and diversity is low even when compared with mice [84,85,109], and their roles—for the most part at least—remain uncertain.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Retinal circuits of eutherian mammals.

(a) Conceptual circuit model for human vision, based on modifications to the ancestral blueprint of fish (cf. Fig 1B). In humans, most cones are ancestrally red, but they randomly express a “red” or a “green-shifted” variant of the LWS opsin (ancestral red cones expressing the “green-shifted” LWS variant are often called green cones. They are not to be confused with ancestral green cones which express RH2, and which are lost in all mammals). A small number of ancestral UV-cones are also retained, now expressing a “blue-shifted” variant of SWS1, hence called blue cone (not to be confused with ancestral blue cones). (b) Retinal distribution of ancestral UV cones in mouse (mod from ref [76]). (c) Example linear (top/middle) and dark-biased (bottom) contrast responses of mouse cones (mod from ref [75]). Putatively, the dark biased responses can be attributed to ancestral UV cones, while the linear responses could come from ancestral red cones. (d) As (a), but for mice.

https://doi.org/10.1371/journal.pbio.3002422.g004

The remains of the ancestral UV system.

Despite our overwhelming reliance on red cones and rods, some noncolour coding ancestral UV circuits may have survived. For example, visual snow syndrome can be exacerbated by selective UV cone activation [111]. This rare human condition has no obvious relation to colour vision—instead, it describes the tendency to experience perceptual noise—that is, to see a “snowy” image. The underlying causes remain poorly understood, but its potentiation following overexcitation of the ancestral UV system hints that it might be linked to primordial UV circuits that originally served other purposes [112].

Beyond the perhaps extreme case of humans, the much-studied retinas of rodents give further insights (Fig 4B–4D). Mice, for example, retain a patch of ancestral UV cones in their ventro-nasal retina, looking straight up at the sky [76] (Fig 4B). The function of this patch remains unclear, but one possibility is in aerial predator detection. Mice are in fact remarkable when it comes to UV vision [75,113–116] (Fig 4C and 4D). Dorsally, their photoreceptor complement resembles the “eutherian-standard” of “green-shifted” ancestral red cones plus rods, sparsely interspersed with UV cones. Ventrally, however, ancestral red cones are UV-sensitive because they co-express the UV-opsin [115]. Almost poetically opposite to the 2 types of red cones of birds, mice therefore have 2 types of UV-sensitive cones: the ancestral UV patch [76], embedded within a larger patch of UV co-expressing red cones [75,116]. This intermixed dual population of UV-sensitive cones has long left it unclear what ancestral UV cones contribute to this retina. After all, it makes little sense to simply “see UV twice”—unless, that is, if the 2 UV systems were used in distinct ways. The circuitry for doing so has presumably been in place for hundreds of millions of years. In support, even though all ventral mouse cones are UV sensitive, functionally they appear to fall into 2 distinct subpopulations [75]: One low gain and approximately linear—ideal for supporting general purpose vision—and another high gain with a strong dark bias—ideal for spotting the silhouettes of overhead birds against the sky [113,117] (Fig 4C). However, it remains unclear whether and how these 2 functional signatures are in fact linked to co-expressing versus ancestral UV cones, respectively (Fig 4D).

Beyond mice, co-expression of short wavelength opsins in ancestral red cones that “look” at the sky is remarkably common among terrestrial tetrapods [16,75,118]. If ancestral red cones are indeed the primordial general-purpose channel (Fig 1B), this awkward arrangement begins to make sense: Unlike in water, on land there is plenty of downwelling short wavelength light available, including at night [113,119,120]. In many cases, a UV image can in fact be substantially sharper compared to an equivalent red image [113,117,121], especially for detecting the edges of overhead vegetation [113,119]. Rather than using the ancestral UV cone complement to exploit this new opportunity, it was probably easier and therefore evolutionarily favourable to locally update the opsin driving the ancestral red circuit.