The cell bodies of most ganglion cells, and the layer formed by their axons, lie close to the inner limiting membrane adjacent to the vitreous, although occasional “displaced” ganglion cells are seen among amacrine cells in the inner nuclear layer [(22, 23), p. 309]. In the central regions of the retina, there are up to eight layers of ganglion cells, whereas in the far periphery, near the ora serrata, there are only sparse clumps of two or three ganglion cells with gaps between them (23).

A fundamental basis for classification—and the one mainly adopted in the present review—is the size and morphology of the cell body and dendrites, as well as the extent of the dendritic field (Figure 1). Already in 1893, using the silver staining method of Golgi, Cajal distinguished several types of ganglion cells by such features, but it was not until 1935 that the predominant ganglion cells of the central region of the primate retina, the midget cells, were described by Stephen Polyak (who had not yet anglicised his name) (24).

The functional properties of a given ganglion cell must necessarily depend on the excitatory and inhibitory inputs it receives from bipolar cells and from amacrine cells. The many different types of amacrine cells allow the construction of specialised circuits that determine the responses of their associated ganglion cells. In principle, it would be possible for two ganglion cells of identical morphology to receive different inputs and thus differ in their functional properties.

A fundamental basis for classifying ganglion cells is therefore the stratum, or strata, of the inner plexiform layer (IPL) in which their dendrites extend. Some are “monostratified,” their dendrites confined to one stratum; some are “bistratified,” having two distinct layers of dendrites (Figure 1). The level or levels at which the dendrites of a given ganglion cell stratify are very characteristic (6). One gross division is between the inner and outer layers of the inner plexiform layers. Bipolar cells of the OFF type predominantly make contact with ganglion cells in the outer part whereas bipolar cells of the ON-type predominantly synapse in the inner part.

A further anatomical classification, and one of special interest to the neuro-ophthalmologist, can be based on the projection sites of each class of ganglion cell. Besides the lateral geniculate nucleus, there are several other brain areas that receive direct projections from the retina, including the superior colliculus, the pulvinar complex, the olivary pretectal nucleus, the supraoptic nucleus of the optic tract, the paraventricular nucleus, the suprachiasmatic nucleus, and the dorsal raphe nucleus (see Figure 2) (25–27).

Classically, the projections of the ganglion cell layer could be established by anterograde tracing, e.g., by intraocular injection of a radioactive agent or cholera toxin subunit B (28), but anterograde tracing of this kind does not identify the type, or types, of ganglion cell in which the projection originates. Retrograde methods, e.g., injecting horseradish peroxidase into the central site, allow specific ganglion cells to be labelled in the retina. An example would be the study by Cowey et al. showing retrograde labelling of several types of primate ganglion cells—including parasol and midget—after injections to the pulvinar (29). One modification of the retrograde method was introduced by Rodieck and Watanabe, who used a fluorescent marker for retrograde tracing and then, in an in vitro preparation, injected rhodamine-conjugated horseradish peroxidase into individual cells that had been labelled by the retrograde marker—a procedure that gave better filling of dendrites than did simple retrograde tracing with horseradish peroxidase (30).

The Golgi method is “noted for its fickleness” (31): this is its strength, in that it exquisitely reveals isolated neurons against a pale background, but it is also its weakness, in that the random staining is unpredictable and not specific to a particular type of cell (32). What are obviously desirable, especially for quantitative purposes, are methods that label either a single class of ganglion cells or only a small number of classes. In the case of lower mammals, many molecular markers (e.g., using antibodies or gene expression) have been developed to identify individual classes (Table 2) [e.g., (46)].

There has been less work of this kind in primates, but of particularly note is the work by Peng et al. (19) and Yan et al. (20) who identified RNA expressed in individual cells from macaque and human retinas (Table 3). In macaque, they were able to group peripheral ganglion cells into 18 clusters and foveal ganglion cells into 16 clusters. Three of the peripheral clusters in macaque retina (and two in the case of human) expressed OPN4, the gene for melanopsin and thus a marker of intrinsically photosensitive RGCs (ipRGCs). Fourteen of the foveal clusters corresponded to peripheral clusters. Although a similar set of transcription factors is used in mouse and primate ganglion cells, there was little correspondence in the detailed RNA expression patterns of individual cell types. In particular, there was no clear mouse equivalent of the midget ganglion cell of the primate. The patterns of RNA expression were very similar for human and macaque retinas, but occasional differences were seen. For example, the gene RBPMS2 was expressed in human, but not macaque midget ganglion cells. In the present context, it is significant that genes known to be associated with glaucoma were found to be predominantly expressed in ganglion cells, sometimes selectively—e.g., SIX6 in midgets (19).

In functional experiments, a physiological measure of a cell's response is recorded when the retina is stimulated with a specific stimulus. Psychophysical work has often guided the choice of stimulus. In the second half of the twentieth century, psychophysicists endeavoured to isolate “mechanisms” or “channels” within the visual system. These constructs were hypothetical, but the hope—not without foundation—was that they corresponded to independent neural channels. The isolation of a given channel was achieved by construction of a stimulus to which the channel might be maximally sensitive and by the use of selective adaptation stimuli to reduce the sensitivity of other channels [e.g., (52–54)]. The techniques that were honed by psychophysicists were often adopted by electrophysiologists and applied to individual ganglion cells. The same techniques often also inspired new clinical testing methods, such as frequency-doubling perimetry, designed to isolate channels with non-linear responses (55). Psychophysics has also inspired the instruments used to deliver the carefully crafted stimuli needed in electrophysiological work—Maxwellian-view optics in the 1960's, computer-controlled CRT displays in the 1980's, and digital light processors in this century.

The celebrated study of Kuffler initiated the extracellular recording of action potentials from individual ganglion cells in the mammalian retina by means of fine-tipped microelectrodes (56). Kuffler demonstrated the antagonistic centre-surround arrangement that characterises the receptive fields of many ganglion cells: stimulation of the centre of the receptive field evokes an ON response in some cells and an OFF response in others, whereas stimulation of a surrounding region evokes the opposite response (Figure 3).

A pioneer in the study of primate RGCs was Peter Gouras, who made microelectrode recordings in vivo from Macaca mulatta and made a basic distinction between cells with transient (“phasic”) responses and those with sustained (“tonic”) responses (57). Influenced by the psychophysical work of Stiles (58), Gouras presented small monochromatic flashes of varying wavelength on monochromatic adapting fields and showed that sustained cells typically drew inputs from one class of cone in the centre of their receptive field and inhibitory inputs from other types of cone in the surround. Given the specificity of the centre input and the predominance of sustained cells in the central field, he identified them with the midget cells of Polyak. The phasic cells were more common in the periphery and appeared to draw inputs of the same sign from long- and middle-wave cones, with little input from short-wave cones.

In the last two decades, it has become possible to record concurrently from several hundred ganglion cells in an eye-cup preparation. A segment of peripheral retina, with pigment epithelium intact, can be placed with ganglion cell layer downwards on a planar array of, say, 512 extracellular micro-electrodes. In the work of Field et al. for example, the macaque retina was stimulated with a lattice of square pixels that varied randomly and independently in chromaticity (59). The responses of individual ganglion cells, identified later off-line, were correlated with the random pattern of stimulation to determine their preferred stimuli.

The introduction of adaptive-optics scanning-light ophthalmoscopy (AOSLO) combined with calcium imaging has been used to monitor the responses of individual ganglion cells in the eye of a living primate [e.g., (60)]. Action potentials cause rapid changes in intracellular free calcium and these can be revealed with a fluorescent protein calcium sensor. McGregor et al. used the sensor GCaMP6s for that purpose. While the retina was stimulated with orange (590 nm) drifting gratings, a 488-nm laser was focussed on the ganglion cell layer to excite the calcium sensor, and the fluorescence was detected in a band at 517–520 nm. A limitation of currently available calcium sensors, such as GCaMP6s, is that they have relatively large time constants (0.6 s) and so cannot follow high frequencies of modulation.

A critical task has been to relate one method of classification to a second. An early success was achieved by intracellular recording with a micropipette electrode filled with a dye such as Procion yellow (61): after a basic characterisation of the cell's response, the passage of hyperpolarizing current could be used to inject the stain into the cell, for later histological examination. Nelson et al. used such a method in eye-cup preparations from cat to show the fundamental mammalian separation of the inner plexiform layer into ON and OFF sublaminae (62). Dacey and Lee used a refinement of this technique in which primate ganglion cells of specific morphology were targeted in a flat-mount preparation under visual inspection, recordings were made with an intracellular micropipette, and the cell was stained by intracellular injection of the fluorescent dye pyranine during recording (63). It was this work that first identified the small bistratified ganglion cell as carrying the excitatory signal of short-wave primate cones.

In 2003, Dacey et al. introduced a powerful technique that has been central to our modern understanding of the range of ganglion cell types (64). The method allows the morphology of the cell to be related not only to stratification level in the inner plexiform layer and to downstream projection sites, but also to the functional characteristics of the cell. Rhodamine dextran was injected in vivo into a central site (e.g., the lateral geniculate nucleus or the superior colliculus) and the dye travelled retrogradely to the retina over the course of 4 to 7 days. In a subsequent in vitro preparation of retina, including the retinal pigment epithelium (RPE) and choroid, the rhodamine dextran was seen to be sequestered within the cytoplasm of ganglion cell bodies, but if, under visual inspection, an individual cell was briefly exposed to light, then the tracer was liberated and spread throughout all the processes of the cell. The tracer did not appear to impair neuronal function, and the responses of the cell could be examined with an extracellular microelectrode, while visual stimuli were delivered via the microscope used for selecting cells and placing the electrode (64).

Peng et al. have linked their single-cell RNA analyses to morphology by combining fluorescent in situ hybridisation (“FISH”) with sparse viral labelling (19). By this means, they were able to confirm—for ON- and OFF-midgets and ON- and OFF-parasols—the tentative identifications that had emerged from their analyses of RNA patterns.

This major cell type accounts for 70% of RGCs. The midget RGCs have a small-sized body with small fields of dendrites (5–10 μm in diameter in the central retina and up to 225 μm in the periphery), which correspond to smaller receptive fields than those of other RGCs. These cells are located mainly in the central retina and project to the parvocellular pathway (66, 67). Midget RGCs have a one-to-one connectivity with midget bipolar cells, which draw their input from a single cone (68). There are two types of midget RGCs: the outer stratified OFF-midget cells show smaller dendritic fields and higher cell densities than the inner ON-midget cells.

The parvocellular pathway is dominated by midget RGCs. Functional assessments of these cells demonstrate that their luminance contrast sensitivity is lower than that of parasol RGCs and most show clear chromatic opponency (69). In general, midget cells have red–green opponency, parasol RGCs are achromatic, and bistratified ganglion cells connect with S-cone ON and L-M cone OFF pathways. However, recent studies suggest that some OFF-midget cells receive signals from short wavelength (blue) sensitive cones (14, 70). Electron microscopy reconstructions of retinal circuits suggest the possibility that a small proportion of midget ganglion cells might have blue–OFF, yellow–ON receptive fields. In addition to colour discrimination, midget RGCs also subserve pattern, texture and stereoscopic depth perception (71).

Parasol RGCs project to the magnocellular layer of the LGN. As with midget cells, there are two types of parasol cells in primates: ON-parasol cells respond with an increase in firing rate at the onset of light in the centre of their receptive field whereas OFF-parasol cells react to off stimuli (63). In synaptic connexions between ON-centre parasol cells and other cells, ~20% of the input is from bipolar cells and the remainder of the signal is introduced from amacrine cells (72).

Parasol RGCs have larger receptive fields and cell bodies, have higher sensitivity to luminance contrast, and present little or no chromatic antagonism, in contrast to midget RGCs (73). Parasol RGCs play a role in motion perception, flicker perception and depth processing based on motion parallax (71). They largely comprise the magnocellular pathway.

This cell type accounts for ~5–8% of RGCs (8). The small bistratified RGCs (sbRGCs) project to the koniocellular layers of the LGN. Small bistratified cells have branches in both layers (inner ON- IPL and outer OFF-IPL): inner ON-IPL branches receive excitatory input from S-ON bipolar cells initiated by S-cones, while opposed (L+M)-OFF light responses arrive through outer OFF-IPL branches (63). This arrangement is thought to give good colour vision with low spatial resolution.

The inputs of large bistratified RGCs have not been elucidated. Large bistratified cells may receive not only S-cone ON-pathway input, but also (L+M) cone OFF-opponency (inhibitory) signals. However, the precise role of this cell type is not yet clear (74).

Smooth monostratified RGCs (smRGCs) have irregular receptive fields with multiple distinct hotspots of light sensitivity. These cells again can be divided into ON- and OFF-cells (75). They might contribute to signalling spatial information via a non-linear mechanism, whereby output is not linearly related to input (76).

The recursive RGCs have moderately densely branched dendritic trees in which many secondary branches tend to curve back towards the soma or close loops of apposing and recursive dendrites. In addition, many dendrites overlap those of neighbouring cells (77). These features resemble those of the directionally selective, motion-sensitive RGCs (dsRGCs) of the rabbit, in which seven types of dsRGCs have been described, namely, ON-types specific to three different directions and ON-OFF types, specific to four different directions. To date, only one population of a bistratified ON-OFF type has been described in the macaque retina (78–80).

There are three types of thorny RGCs in the primate retina that account for ~1% of ganglion cells (77). Broad thorny RGCs are given various names such as thorny diffuse, T-group cells, S3 narrow thorny, and hedge cells (30, 49, 81). The dendrites of broad thorny RGCs span a whole layer of the inner plexiform layer. It is presumed that the cells may contribute to ON/OFF-centre light responses that are strongly suppressed by stimulation of the receptive field surround, such as local edge detector cells in rabbits (82). Additionally, two narrowly stratified cells, including outer and inner, are found in primates and their connectivity has not been clarified yet.

These cell types are monostratified cells that receive input from amacrine and bipolar cells. The transcription factor Satb2 is expressed in large sparse RGCs in macaque and human retina (83).

The ipRGCs constitute 1% of the total RGC population in humans (84). These cells have large, sparse dendritic fields. They are intrinsically photosensitive because of the expression of the melanopsin photopigment and capable of phototransduction independently of rods and cones (84). In mouse retinas, six subtypes (M1, M2, M3, M4, M5, and M6) of ipRGCs have been identified with distinctive anatomical and functional properties (Figure 4) (85). M1 and M2 ipRGCs account for the majority (74–90%) of ipRGCs. The main function of ipRGCs, in particular of the Brn3b-M1subtype, is to contribute to circadian photoentrainment through the projections to the suprachiasmatic nucleus (SCN) of the hypothalamus (86), but they are also relevant for other non-image forming functions of the eye, including the regulation of the pupillary light reflex through the projections to the OPN. M1 and M2 ipRGCs project to both the SCN and the olivary pretectal nucleus (OPN); however, M1 ipRGCs innervate the outer shell region of the OPN, where projection neurons that innervate the pre-autonomic Edinger-Westphal nucleus reside, whilst M2 ipRGCs innervate the OPN central core (87). The input dendrites of M2 and M4 ipRGCs are in the inner retina (ON-pathway), whereas those of M1 ipRGCs are located near the inner nuclear layer (OFF-pathway). The dendrites of M4 and M5 ipRGCs are located in the inner lamina of the inner plexiform layer. M4 ipRGCs have a larger cell body compared with M5 ipRGCs that have small, highly branched dendrites arrayed uniformly around the soma.

In humans, three ipRGC subtypes (M1, M2, and M4) have been defined (88). M1 ipRGCs have outer stratifying dendrites with a few smooth spines in the outer IPL, whereas M2 ipRGCs stratify in the inner part of the inner plexiform layer (IPL) close to the ganglion cell layer. M1 ipRGCs have been divided into two subtypes, gigantic M1 RGCs (GM1 cells) with round or oval large soma and displaced M1 RGCs (DM1 cells). Both ipRGC types receive inputs from DB6 bipolar cells and project to the dorsal LGN (89). M2 ipRGCs have larger soma and more branched dendrites than M1 ipRGCs. M1 ipRGCs are reported to receive an inhibitory input from short-wave cones via an S-cone amacrine cell (90), whereas M2 ipRGCs receive input from S-On bipolar cells and contribute to the blue cone pathway (91). M1 ipRGCs project to the SCN to synchronise the circadian clock and M2 ipRGCs project to the OPN in the thalamus to control pupillary response.

The ipRGCs are relatively preserved in the mitochondrial optic neuropathies, such as Leber hereditary optic neuropathy (LHON) and autosomal dominant optic atrophy (ADOA). However, ipRGCs are affected in other optic neuropathies, such as glaucoma, and late-onset neurodegenerative disorders, such as Alzheimer disease and Parkinson disease (92–95).

There are a small number of unclassified RGCs in primates that do not fit with any of the previously described RGCs (8). Further work is needed to elucidate the characteristics of this miscellaneous group of RGCs.

Although RGCs have been extensively studied in primates, the clinical assessment of RGCs has proven more challenging as they cannot be evaluated directly.

A number of psychophysical measurements can be used to investigate changes in RGC function.

There is a long list of aetiological factors that can result in RGC injury and optic nerve degeneration. Compared with inherited optic neuropathies, fewer studies have focused specifically on RGC pathophysiology in acquired optic neuropathies. More work is, therefore, needed to elucidate subtype selectivity, if any, of RGC loss in ischemic, compressive, inflammatory, autoimmune and paraneoplastic optic neuropathies. However, we do know that most toxic optic neuropathies have an underlying mitochondrial aetiology (180). There is a growing body of evidence that mitochondrial dysfunction plays a prominent pathophysiological role in glaucoma, demyelinating optic neuritis and toxic optic neuropathies (181, 182). This aetiological link is relevant and comparing the pattern of RGC loss between these acquired optic neuropathies and classical monogenic optic neuropathies could reveal common pathways amenable to therapeutic intervention.

(1) Glaucoma

Glaucoma is a leading cause of irreversible blindness affecting 3–5% of the population over the age of 70 years (183). Extrafoveal RGCs usually deteriorate in the early stages resulting in arcuate scotomas in the visual field. Traditional anatomical studies reported greater loss of axons of large diameter, corresponding to the magnocellular pathway (parasol cells) (184), and the magnocellular LGN layers were more affected compared with the parvocellular LGN layers (185). However, there are rarer types of retinal ganglion cells with large axons and further investigations are needed to evaluate the changes of these cells in glaucoma. The relative vulnerability of large axons in glaucoma may simply reflect the anatomical location of the affected ganglion cells. Glaucoma patients have poor response to high temporal frequency light stimuli that correspond to the magnocellular pathway. In a primate study using immunohistochemistry, a decrease in large RGCs was observed after elevating IOP (186). This specific vulnerability was ascribed to calcium-permeable receptors, the relative proximity of RGCs and their dendrites to blood supply in the IPL layer, and the differing metabolic requirements of these particular large cell types (187). However, other studies suggested no predilection for a specific pathway (188, 189). Compared with inherited optic neuropathies, the ipRGCs are vulnerable in both patients with confirmed glaucoma and glaucoma suspects (190, 191). In contrast, ocular hypertension does not seem to result in significant loss of ipRGCs (192).

(2) Demyelinating Optic Neuritis

Inflammatory demyelination resulting in optic neuritis is a major manifestation of multiple sclerosis. Inflammation of the retinal vascular endothelium can precede demyelination and perivascular cuffing and oedema of the optic nerve sheath leads to breakdown of myelin (193). Idiopathic demyelinating optic neuritis leads to visual loss with minimal axonal loss.

Optic neuritis is associated with alteration of both the parvocellular and magnocellular pathways (194). Viret et al. suggested that the more heavily myelinated magnocellular axons are more vulnerable in patients with optic neuritis because low spatial frequencies, which are transmitted by the magnocellular pathway, are affected predominantly 1 month after the acute phase of the optic neuritis episode (195). Despite the recovery of visual acuity, the magnocellular pathway did not fully normalise (196). In contrast, a significant loss at high spatial frequencies has been reported in the affected eye and the parvocellular pathway was more impaired in patients with resolved optic neuritis who had 20/20 visual acuity after recovery (197). Fallowfield and Krauskopf suggested that chromatic discrimination is more severely impaired than luminance discrimination in the demyelinating diseases (198). This discrepancy might be due to differences in the timing and severity of optic neuritis. Consequently, it is still unclear which pathway is more vulnerable in the context of demyelinating optic neuritis (196). Both red-green and tritan defects have been reported in optic neuritis (199). Characteristics of colour deficiency may change over time as assessed with the FM 100-hue test, with blue-yellow defects being more common in the acute stage and red-green changes being predominant in the chronic stage (200). It is possible, of course, that the variability of symptoms in optic neuritis reflects immunologically distinct conditions that differentially affect different types of RGCs.

(3) Toxic Optic Neuropathy

Various substances such as ethambutol, isoniazid, linezolid, chloramphenicol and methanol can cause optic nerve dysfunction, probably through acquired mitochondrial dysfunction (180). As in inherited optic neuropathies, the papillomacular bundle is selectively vulnerable and this typical feature can be confirmed by optical coherence tomography, which shows a profound decrease in temporal RNFL thickness. The parvocellular pathway within the papillomacular bundle is affected extensively likely secondary to a number of factors, including smaller and more thinly myelinated nerve fibres and a faster firing response with higher average rates of action potentials (201). However, there is a lack of evidence on whether this is simply because the parvocellular neurons predominate in the papillomacular bundle, or whether the midget cells are the primary target of the triggering toxic substances.

The physiological features of the major RGC subtypes (mRGCs, pRGCs, and sbRGCs) are well-known, but the role and characteristics of other RGCs require further study. An in-depth characterisation of the chronological structural and functional changes occurring within the RGC layer in optic nerve disorders, including inherited and acquired optic neuropathies, are important to inform the future design of clinical trials. Understanding which RGC subtypes are selectively affected will help optimise outcome measures in natural history studies and trials of experimental therapies. As mentioned earlier, the FDT test is used for the early detection of glaucoma because the magnocellular pathway is more vulnerable (127). Given that a common variant in the SIX6 gene (rs33912345) is strongly associated with primary open-angle glaucoma (POAG) and the fact that this gene is highly expressed in midget RGCs, tests that evaluate this particular cell type could prove to be a useful sensitive biomarker of disease progression (19, 202).

The remarkable advances in gene delivery and editing technology have led to an increasing number of clinical trials for optic neuropathies, in particular gene replacement therapy for monogenic inherited optic neuropathies (203). Gene therapy using adeno-associated viral vectors is currently favoured and there is now cumulative evidence of its long-term safety and efficacy in delivering gene constructs to retinal cells (204, 205). Promising results have been obtained with allotopic expression of the MT-DN4 gene in patients with LHON treated within 1 year of disease onset (206, 207). To enhance success of gene therapy, optimised tissue-specific promoters, which control expression of the therapeutic gene, are needed, and these could potentially. be optimised for the relevant RGC subtype (208). Genomic editing, such as the CRISPR-Cas system, and stem cell therapy is an exciting development that has the potential to revolutionise the treatment of ophthalmological disorders given the eye's relative ease of anatomical access and its relative immune privilege (209, 210). The intriguing preservation of ipRGCs in mitochondrial optic neuropathies needs to be investigated further as the factors that confer this resilience would be obvious therapeutic targets (211).

There is increasing interest in employing RGCs to restore visual function in the retinal dystrophies marked by widespread loss of rods and cones (212). Optogenetic therapies are being developed to confer light sensitivity to inner retinal neurons, which are spared in these forms of outer retinal degeneration. Another approach is the use of electronic implants to stimulate these inner retinal neurons so that visual signals can be transmitted to the brain. A better understanding of inner retinal connectivity, specifically that of RGCs, is essential to optimise these innovative sight restoring strategies. Elucidating the selective vulnerability of RGCs compared with other retinal and neuronal cell types in inherited optic neuropathies is key to developing targeted treatments for this group of disorders. The availability of high-throughput transcriptomic techniques that can be conducted at the single cell level is an exciting development, providing us with powerful tools to identify pathways that can be modulated for generalizable, mutation-independent neuroprotective strategies (213). Although appealing, regenerative medicine will require not only the replacement of the missing RGCs, but also the establishment of the sophisticated circuitry that allows the integration of signals from various pathways to achieve a reasonable degree of visual perception (214).