Intrinsically photosensitive ganglion cells contribute to plasticity in retinal wave circuits

Abstract

Correlated spontaneous activity in the developing nervous system is robust to perturbations in the circuits that generate it, suggesting that mechanisms exist to ensure its maintenance. We examine this phenomenon in the developing retina, where blockade of cholinergic circuits that mediate retinal waves during the first postnatal week leads to the generation of “recovered” waves through a distinct, gap junction–mediated circuit. Unlike cholinergic waves, these recovered waves were modulated by dopaminergic and glutamatergic signaling, and required the presence of the gap junction protein connexin 36. Moreover, in contrast to cholinergic waves, recovered waves were stimulated by ambient light via activation of melanopsin-expressing intrinsically photosensitive retinal ganglion cells. The involvement of intrinsically photosensitive retinal ganglion cells in this reconfiguration of wave-generating circuits offers an avenue of retinal circuit plasticity during development that was previously unknown.

The computations performed by neural circuits are not determined by hard-wired anatomy but rather can be altered by experience or different neuromodulatory states (1–3). This plasticity is particularly important during development, when neural circuits show remarkable robustness against perturbations that disrupt the patterned, spontaneous activity required for normal development (4). For example, giant depolarizing potentials in the developing hippocampus are maintained against decreases in gamma-aminobutyric acid (GABAergic) transmission by increasing the strength of glutamatergic transmission (5). Similarly, spontaneous network activity in the developing spinal cord is maintained against alterations in GABAergic transmission by changes in both the intrinsic excitability of individual neurons and changes in synaptic strength of glutamatergic synapses (6). The developing retina also shows robustness against perturbations in circuits that generate spontaneous retinal waves (4). For example, disruption of normal cholinergic transmission during the first postnatal week leads to the generation of waves via a distinct gap junction coupled network (7–9). These observations indicate that degenerate circuit mechanisms exist in the developing retina to maintain spontaneous activity.

Here we explore the hypothesis that intrinsically photosensitive retinal ganglion cells (ipRGCs) contribute to this wave circuit plasticity. ipRGCs are a recently discovered class of photoreceptors that express the photopigment melanopsin (10) and are light sensitive in mice from birth, unlike rod and cone photoreceptors, which become photosensitive after 2 postnatal weeks of development (11). Although ipRGCs are typically involved in non-image-forming functions, such as entrainment of circadian rhythms (12), they have been shown to support intraretinal signaling via gap junction coupling and by signaling to dopaminergic amacrine cells (DACs) (13–15). Indeed, light stimulation of ipRGCs can modulate cholinergic retinal circuits during development (16).

We use multielectrode array (MEA) recordings to compare the spatial and temporal properties of firing patterns recorded in the dark versus the light from wild-type (WT) mice and knockout mice lacking normal cholinergic waves (β2KO), in addition to lacking the gap junction protein connexin 36 (β2-cx36 dKO) or the photopigment melanopsin (β2-Opn4 dKO). Our data support the hypothesis that early light responses from ipRGCs contribute to the circuit that mediates the recovery of correlated spontaneous firing patterns in the absence of cholinergic waves.

Results

Recovered Waves in Mice Lacking Cholinergic Waves Are Modulated by Light.

To investigate firing patterns across retinal ganglion cells (RGCs) in both the dark and light, we performed MEA recordings on retinas acutely isolated from WT mice and in mice lacking the β2 subunit of nicotinic acetylcholine receptors (nAChRs) (β2KO) at postnatal days 4–7 (P4–P7), when cholinergic waves normally occur (17, 18). At these ages, β2KO mice exhibit gap junction waves in place of cholinergic ones (8, 9). In addition, rod and cone photoreceptors are not yet photosensitive and do not contribute to ganglion cell light responses (19). WT recordings confirmed previous observations (7, 9, 20, 21), where RGCs fired periodic bursts of action potentials that swept across the retina as a wave (Fig. 1A). Immediately following light onset, a subset of cells fired sustained bursts of action potentials, consistent with previously described light responses of ipRGCs in early postnatal development (22, 23).

Fig. 1.

Recovered waves in mice lacking cholinergic waves are modulated by light. (A and B) MEA recordings from a WT P5 (A) and β2KO P5 (B) retina. (Top) Activity pattern corresponding to boxed region in Middle. Each dot represents an electrode site, and the radius of the dot is proportional to the single unit firing rate recorded at that site. Frames correspond to 2 s. (Middle) Raster plot of spike trains of all single units. (Bottom) Average firing rate of all units. (C) Summary data of wave frequency in dark and light for WT, β2KO, and β2-Opn4 dKO mice. Open circles correspond to individual retinas and recordings from the same retinas are connected by dotted lines. (D) Summary data of burst duration during a wave in dark and light for WT, β2KO, and β2-Opn4 dKO mice. Error bars in C and D correspond to SEM. *P < 0.05; ***P < 0.001, paired t test.

Recordings of β2KO mice also exhibited retinal waves with distinct propagation properties from WT mice. In particular, waves in β2KO mice were less frequent and many RGCs fired action potentials that were not associated with retinal waves and therefore were not significantly correlated with one another, in agreement with previous studies (8, 9, 21, 24) (Fig. 1B). Surprisingly, we also observed that light stimulation led to an almost twofold increase in the frequency of waves in β2KO retinas, in contrast to WT retinas, for which light stimulation had no effect on wave frequency (Fig. 1C, ***P < 0.001). This light-induced increase was observed over a range of starting frequencies, suggesting that the light-induced effect is independent of initial wave frequency. Light stimulation also increased the burst duration during a wave in both WT and β2KO retinas (Fig. 1D), as previously reported for WT retinas (16).

Because ipRGCs are the only functional photoreceptors at these ages and because the light-evoked increase in burst duration of cholinergic WT waves is eliminated in melanopsin knockout mice (Opn4 KO) (16), we presumed that the observed light-evoked effects on waves in β2KO retinas were mediated by ipRGCs. To test this directly, we performed MEA recordings on a β2-melanopsin double KO mouse (β2-Opn4 dKO, kindly provided by David Copenhagen, University of California, San Francisco). These mice exhibited retinal waves but did not show an increase in wave frequency or burst duration during a wave in the light (Fig. 1 C and D). These observations confirm that the photic effects on waves in β2KO mice were mediated by ipRGCs.

Waves in β2KO Mice Require cx36.

Previous studies have shown that the recovered waves in β2KO mice are blocked by gap junction antagonists but not fast neurotransmitter antagonists (8). We assessed whether the presence of the neuronal gap junction protein cx36 was necessary for waves in β2KO mice by performing MEA recordings on a β2-cx36 dKO mouse (24). Cx36 is the most abundant retinal connexin and couples most ganglion cells to other ganglion or amacrine cells (25). Although cx36KO mice have an increase in asynchronous firing in between retinal waves, they exhibit cholinergic retinal waves with propagation patterns that are indistinguishable from WT during the first postnatal week (24, 26). We found that β2-cx36 dKO mice did not exhibit recovered retinal waves in either the dark or light (Fig. 2A). Rather, many RGCs fired asynchronous action potentials. We characterized the correlation properties of spiking neurons by computing correlation indices as a function of intercellular distance for all cell pairs. This gives a measure of the likelihood relative to chance that two cells fire together within a given time window, where retinal waves are characterized by a correlation index that is high for nearest neighbors and that falls off with increasing intercellular distance (17). In contrast, correlation index curves of β2-cx36 dKO mice were flat, confirming the absence of retinal waves (Fig. 2B). In some retinas, we observed synchronous bursting among subsets of cells in the light, however this activity did not propagate in a wave-like manner (Fig. S1A, Top). These observations confirm that recovered waves in β2KO mice are gap junction mediated and show that cx36 is required for their propagation.

Fig. 2.

β2KO waves require cx36. (A) MEA recording of spontaneous activity in β2-cx36 dKO mouse (as in Fig. 1A). (B) Correlation index versus interelectrode distance for pairs of spike trains in the light for β2-cx36 dKO mice. Data points correspond to averages of median values from individual retinas and error bars correspond to SEM.

ipRGCs Do Not Function as “Hub” Neurons for Recovered Waves.

One way in which ipRGCs may increase β2KO wave frequency in the light is by functioning as hub neurons, or highly connected nodes, that link together many cells and thereby impart synchrony within the network. To test this possibility, we first identified neurons whose firing was highly correlated with a large number of other cells and then determined whether these highly connected cells corresponded to ipRGCs.

We first computed correlation indices as a function of intercellular distance for all cell pairs. WT correlation indices were high for nearest neighbors and fell off with increasing intercellular distance, characteristic of retinal waves (Fig. 3A). β2KO correlation index curves were flatter than those for WT (Fig. 3A), consistent with larger waves and a faster propagation speed, as described in previous studies (8, 9). Correlation index curves showed no difference in the dark and light for both WT and β2KO mice. We next used these correlation indices to construct connectivity maps, which show connections between neurons with the highest correlation indices (26, 27). In particular, this strategy has been used in the developing hippocampus to identify highly connected hub neurons, which repeatedly initiate activity (28). We defined two cells as being connected if their correlation index fell in the top 5% of all correlation index values and extracted cells that were connected to at least 15% of other cells (Fig. 3 B and C; “highly connected cells” correspond to red units and their connections are shown by blue lines).

Fig. 3.

ipRGCs do not function as hub neurons for retinal waves. (A) Correlation index versus interelectrode distance for pairs of spike trains for WT (black) and β2KO (blue) mice, in the light. Data points correspond to averages of median values from individual retinas and error bars correspond to SEM. (B and C) Connectivity maps of WT P5 (B) and β2KO P5 (C) in the light. Circles correspond to location of single units with diameter scaled by the magnitude of the normalized correlation index. Red circles correspond to units that were connected to at least 15% of other units, and blue lines show their connections. Turquoise circles correspond to units identified as ipRGCs. (D) Cumulative probability distributions of the distances of connections from highly connected cells for WT (black) and β2KO (blue) retinas in the dark and light. (E) Percent of ipRGCs that are highly connected hub neurons for WT (gray) and β2KO (blue) activity, in dark and light. Box plots range from lower to upper quartiles (25% and 75%) with median values indicated by central black line; whiskers (dotted lines) range from 5% to 95%. Faded box plots correspond to percentage of ipRGCs that are highly connected that we would expect from chance. **P < 0.01, Wilcoxon rank sum test.

For WT retinas, highly connected cells fell in a cluster of adjacent units with a median connection length of 145 μm (Fig. 3 B and D). Highly connected cells in β2KO mice were more dispersed than in WT, with a median connection length of 260 μm (Fig. 3 C and D). Median connection lengths in the dark were similar to those in the light for both WT and β2KO mice (Fig. 3D), indicating that light does not alter the underlying spatial architecture of waves.

We determined the locations of ipRGCs on the MEA by identifying units that showed at least 45 additional spikes in the 60-s window following light onset compared with the 60 s preceding light onset (0.75 Hz increase; Fig. S2). This threshold was determined using targeted cell-attached recordings of ipRGCs in Opn4–EGFP mice, in which GFP is expressed under the melanopsin promoter (29). We next determined whether identified ipRGCs (turquoise units in Fig. 3 B and C) function as hub neurons for retinal waves by computing the percentage of ipRGCs that were highly connected and comparing this to the percentage we would expect by chance. In the dark, the percentage of ipRGCs that were highly connected did not differ from chance for both WT and β2KO mice (Fig. 3E). Surprisingly, in the light the percentage of ipRGCs that were highly connected was significantly less than chance (Fig. 3E, **P < 0.01). These observations show that ipRGCs do not function as hub neurons, as activation of ipRGCs in the light does not directly activate many other RGCs. Rather, they suggest that the action of light on wave frequencies in β2KO mice is an indirect effect of ipRGCs on the network. Below we explore whether this modulation occurs via dopaminergic signaling.

β2KO Waves Are Modulated by Dopaminergic and Glutamatergic Signaling.

ipRGCs are thought to influence retinal networks via glutamatergic stimulation of dopamine release from DACs [14, 15), but see ref. 30]. Dopamine is produced in mice as early as P4 by DACs (31) and is a major player in regulation of gap junction coupling via cAMP-dependent posttranslational phosphorylation (25, 32). In general, activation of D1-like receptors decreases gap junction coupling while activation of D2-like receptors increases coupling, although the magnitude of the effects are highly cell-type specific (33–35). Thus, one way that ipRGCs could exert an indirect effect on network synchrony is via a dopaminergic pathway.

We first tested whether dopamine signaling regulates wave frequencies in WT and β2KO mice by performing MEA recordings in dopamine receptor antagonists. We found that neither a D1 receptor antagonist (SCH23390, 10 μM) nor a D2 receptor antagonist (raclopride, 8 μM) had an effect on wave frequency in WT retinas (Fig. 4A). However, raclopride strongly reduced wave frequency in β2KO mice in the light, whereas SCH23390 increased their frequency (Fig. 4B). Neither antagonist significantly influenced mean baseline firing rates relative to the mean firing rate in the light (Fig. 4C), suggesting that the observed effects are likely due to modulation of gap junction coupling required for waves and not modulation of a cell’s overall firing properties. Furthermore, application of SCH23390 to β2-cx36 dKO mice did not induce the generation of waves, nor did it change mean baseline firing rates relative to the mean firing rate in the light (Fig. S1B), indicating that cx36 is a possible target of D1 receptor activation. These observations show that dopamine signaling strongly modulates the frequency of recovered waves, where waves are stimulated by D2 receptor activation but suppressed by D1 receptor activation.

Fig. 4.

β2KO waves are modulated by dopaminergic and glutamatergic signaling. (A and B) Wave frequency in dark, light, and in presence of dopamine receptor antagonists in light for WT (A) and β2KO (B) mice. (C) Mean firing rate in dark, light, and in presence of dopamine receptor antagonists in light for β2KO mouse. (D) Wave frequency in dark, light, and in presence of glutamate receptor antagonists in light for β2KO mice. (E) Mean firing rate in dark, light, and in presence of glutamate receptor antagonists in light for β2KO mice. (F) Wave frequency in dark, light, and in presence of glutamate receptor antagonists in light for β2-Opn4 dKO mice. For all plots, open circles correspond to individual retinas, and recordings from the same retinas are connected by dotted lines. Error bars correspond to SEM. D1R antagonist, SCH23390 (SCH), 10 μM; D2R antagonist, raclopride (rac), 8 μM. AMPAR and NMDAR antagonists, DNQX and d-AP5, 20 and 50 μM. *P < 0.05; **P < 0.01, repeated measures ANOVA with Holm-Sidak posthoc test.

Recent experiments indicate that ipRGCs form excitatory glutamatergic synapses onto DACs (14, 15). Thus, if the observed light-evoked increase in wave frequency in β2KO mice were mediated by ipRGC feedback onto DACs, we would expect that blocking glutamate receptors would block the effect. To test this hypothesis, we performed MEA recordings of β2KO mice in glutamate receptor antagonists in the light [d-AP5, 50 μM and 6,7-dinitroquinoxaline-2,3-dione (DNQX), 20 μM]. In four out of five retinas tested, wave frequencies returned to dark levels in the presence of glutamate receptor antagonists with no effect on baseline firing rates (Fig. 4 D and E), suggesting that the antagonists influence wave properties but not a cell’s overall firing properties. To test whether these glutamatergic effects were dependent on light-evoked firing of ipRGCs, we repeated the measurements in β2-Opn4 dKO mice. We found that glutamate receptor antagonists had no effect on wave frequencies in these mice (Fig. 3F), indicating that the glutamatergic reduction of wave frequency in β2KO mice in the light was melanopsin dependent. Together, these observations are consistent with the model that ipRGCs influence retinal networks via glutamatergic signaling, and likely via a dopaminergic pathway.

Light-Sensitive, Noncholinergic Wave Circuit Is Present but Latent in WT Mice.

The differential modulation by light and dopamine receptor antagonists between cholinergic and recovered gap junction waves indicates that they are mediated by distinct circuits. Does a noncholinergic wave circuit exist as a “latent” circuit in WT retinas, or does it require an extended perturbation provided by the β2KO mouse? To address this question, we tested whether acute block of cholinergic waves in WT mice unmasked a light- and dopamine-sensitive wave-generating circuit. We monitored light-evoked activity after a 20-min nAChR blockade using di-hydroß-erythroidine (DHβE, 8μM). Approximately 2 min after light onset, we observed rhythmic bursting with a periodicity of a few seconds (Fig. 5A), similar to that observed in another study after prolonged (10 h) nAChR blockade (7), but we did not detect propagating waves.

Fig. 5.

Light-sensitive, noncholinergic wave circuit is present but latent in WT mice. (A and B) MEA recordings from a WT P6 retina after 20 min of DHβE (Dβ) application in dark (A) and after 20 min of DHβE+SCH (Dβ+S) application in dark (B) (as in Fig. 1A). (C–E) Frequencies (C), burst duration (D), and firing rates (E) of waves in WT (gray), β2KO (blue), and WT in DHβE+SCH (red). Open circles correspond to individual retinas. (F and G) Correlation indices versus interelectrode distance for WT in DHβE alone (F) and DHβE+SCH (G) in the light. (H) Data from Figs. 3A and 5G for WT (gray), β2KO (blue), and WT in DHβE+SCH (red). (I) Connectivity map for WT in DHβE+SCH in the light (as in Fig. 3B). (J) Cumulative probability distributions of the lengths of connections from highly connected cells (WT and β2KO curves correspond to those from Fig. 3D). (K) Percent of ipRGCs that are highly connected hub neurons for WT in DHβE+SCH in the light compared to chance (box plots as in Fig. 3E). C–E, *P < 0.05, **P < 0.01, one-way ANOVA; C, ***P < 0.001, paired t test; K, **P < 0.01 Wilcoxon rank sum test. nAChR antagonist, DHβE (Dβ), 8 μM; D1R antagonist, SCH23390 (SCH or S), 10 μM.

We next tested whether recovery of waves was being suppressed by D1 receptor signaling by applying the D1 receptor antagonist SCH23390 for 20 min in the dark to DHβE-treated retinas, and monitoring activity in the dark and light. Light stimulation led to the generation of propagating, correlated wave-like events, with similar spatial-temporal properties to waves in β2KO mice (Fig. 5 B–K). Specifically, connectivity maps of recovered waves matched those observed in β2KO mice, where highly connected cells had a dispersed, gap junction signature rather than a clustered, cholinergic one (Fig. 5 I and J). In addition, the coincidence of highly connected cells with ipRGCs was lower than chance in the light, as observed in β2KO mice (Fig. 5K, **P < 0.01). Finally, although some recovered waves were present in the dark in some retinas, the frequency of waves increased significantly in the light (Fig. 5C, ***P < 0.001). Together, these observations show that an auxiliary wave-generating circuit is latent in WT mice, where it is normally suppressed by a combination of nAChR activation together with D1R signaling. Light stimulation of ipRGCs facilitates the activation of this auxiliary circuit, indicating that ipRGCs contribute to the recovery of correlated spontaneous firing patterns in the absence of cholinergic waves.

Discussion

In this study we demonstrate that ipRGCs are used in degenerate circuit mechanisms to maintain correlated spontaneous activity in the developing retina. These observations show that retinal wiring diagrams are dynamic and malleable during development and suggest a unique function for ipRGCs in mediating retinal wave plasticity that underlies the maintenance of correlated activity.

We found that in the absence of cholinergic waves during retinal development, a distinct light-modulated wave circuit was activated (Figs. 1 and 5). Cx36 was necessary for waves in β2KO mice (Fig. 2), and because neither light nor dopamine antagonists induced waves in β2-cx36 dKO mice, we postulate that gap junctions are the likely target of the observed light and dopamine modulation of noncholinergic waves. Similar modulation of gap junctions has been observed in many retinal circuits, implicating gap junctions as sites of plasticity in both developing and adult retinal circuits. For example, light and dopamine modulation of coupling has been extensively described for horizontal cells (36), AII amacrine cells (34, 37), and alpha-ganglion cells (33, 35).

Based on the observation that ipRGCs likely signal to the retina via dopaminergic signaling [(14, 15) but see ref. 30], we propose the following model. During cholinergic waves, the dominant circuit is starburst amacrine cell release of acetylcholine onto ganglion cells and other starburst amacrine cells (Fig. 6A). Upon cholinergic block, ipRGCs increase their contribution to network dynamics and the dominant circuit becomes a gap junction coupled network regulated by ipRGCs acting through modulation of dopamine release (Fig. 6B). In this model, we depict the ipRGC–DAC synapse as a reciprocal connection, as ipRGCs themselves express dopamine receptors (38). How a gap junction coupled network is activated in the absence of cholinergic waves remains to be delineated. However, one possibility is that RGCs experience a change in gap junction coupling in response to reduced cholinergic input. This has previously been shown to occur in rat adrenal medulla, where acute pharmacological block of nAChRs leads to an increase in dye coupling of adrenal chromaffin cells together with an increase in junctional currents (39).

Fig. 6.

Model of overlapping wave circuits. (A) During cholinergic waves, the dominant circuit is nAChR activation by acetylcholine, which is spontaneously released from starburst amacrine cells. Activation of D1 receptors inhibits the gap junction circuit. (B) Following cholinergic block, ipRGCs may modulate dopamine release from DACs, which, via activation of D2 receptors, modulates the strength of gap junction coupling between ganglion cells (not pictured) or among a network of amacrine and ganglion cells. SAC, starburst amacrine cell; AC, amacrine cell; DAC, dopaminergic amacrine cell; RGC, retinal ganglion cell; ipRGC, intrinsically photosensitive retinal ganglion cell.

Further, we found that reduced D1 receptor signaling was required to recover wave-like events in the absence of cholinergic waves in WT retinas (Fig. 5), suggesting that dopamine signaling may normally suppress the auxiliary wave circuit during cholinergic waves. In addition, we found that recovered waves in β2KO mice were suppressed by activation of D1 receptors but stimulated by activation of D2 receptors (Fig. 4B). Because D2 receptors are approximately 10-fold more sensitive to dopamine than D1 receptors (34), a balance of the opposing effects of D1 and D2 receptor activation can likely be achieved in vivo with a low concentration of dopamine. Interestingly, although blockade of D2 receptors blocked waves, it did not decrease the mean firing rate of individual RGCs (Fig. 4C). Hence, we postulate that light stimulation of ipRGCs does not increase wave frequency by a general increase in network excitability, but rather via further increases in gap junction coupling, thus lowering the threshold for wave events. Together, these observations suggest that the absence of cholinergic waves might result in changes in dopamine signaling, which facilitate activation of recovered waves and sculpt their dynamics.

Our data suggest that the circuits that mediate both cholinergic and recovered gap junction waves exist in WT retina as opposed to emerging after the prolonged activity blockade that accompanies genetic deletions. Specifically, a short-term cholinergic block resulted in periodic activity and correlated wave-like activity emerged when D1 receptor signaling was low (Fig. 5). Hence, the wiring diagram of the developing retina many include several “overconnected circuits,” in which some circuits are closed and others activated depending on the internal state of the system (2), such as classically described in the stomatogastric ganglia (1) and recently described in Caenorhabditis elegans (40). The data presented here indicate that the developing retina may use a similar overconnection strategy as a means of maintaining spontaneous firing patterns. Such a strategy has the advantage of allowing the network to rapidly change its properties without having to construct new circuits. It is interesting to postulate whether this is a general mechanism used by other developing networks.

Materials and Methods

Animals.

Recordings were performed on mice aged postnatal day P4–P7 from C57BL/6 WT (Harlan Laboratories, Indianapolis), β2KO (A. Beaudet, Baylor University, Waco, TX) (41), β2-Opn4 dKO (David Copenhagen, University of California, San Francisco), Opn4-GFP (P. Kofuji, Minnesota University, Minneapolis) (29), β2KO/Opn4-GFP, and β2-cx36 dKO (24). Animal procedures were approved by the University of California, Berkeley Institutional Animal Care and Use Committees and conformed to the National Institutes of Health Guide for the Care and Use of Laboratory Animals, the Public Health Service Policy, and the Society for Neuroscience Policy on the Use of Animals in Neuroscience Research. Animals were anesthetized with isofluorane, decapitated, and the eyes were enucleated in a dark room with dim red ambient light. Retinas were removed from eyecups in 95% O₂–5% (vol/vol) CO₂ bicarbonate buffered Ames’ solution (purchased from Sigma-Aldrich) under infrared optics.

MEA Recordings.

Isolated pieces of retina were placed RGC side down onto a 60-electrode commercial MEA arranged in an 8 × 8 grid excluding the four corners, with 10 μm diameter electrodes at 100 μm interelectrode spacing (Multi Channel Systems). The retina was held in place using a dialysis membrane weighted with a ring of platinum wire. The recording chamber was superfused with Ames’ solution bubbled with 95% O₂ and 5% CO₂ and maintained between 33 and 35 °C, pH 7.4. Each preparation was allowed to equilibrate for 20 min in the dark before starting data acquisition. Spontaneous firing patterns were recorded for 30 min in the dark followed by 30 min of unfiltered broad-band full-field light, delivered by a tungsten-halogen lamp with irradiance (in photons s⁻¹⋅cm⁻²) of 2.4 × 10¹² at 480 nm and 2.9 × 10¹³ at 600 nm. This corresponds to a photon flux comparable to that experienced by newborn pups through closed eyelids (16). A second series of dark-light recording conditions was repeated to ensure that any changes in firing patterns were not due to a change in recording conditions over extended periods of time. Raw data were filtered between 120 and 2,000 Hz, and spikes sorted offline to identify single units using Plexon Offline Sorter software. The mean firing rate of all units over the duration of the recording was calculated and units with a mean firing rate less than 10% of the overall mean firing rate were excluded from further analysis. Spike-sorted data were analyzed in MATLAB (MathWorks).

To identify wave events, we used a modified Poisson Surprise algorithm, outlined below (9, 42). The recording was divided into 1-s bins and the firing rate of each single unit in each 1-s bin was determined. From this, the probability of chance occurrence of the firing rate in each bin given a unit’s mean firing rate was determined using the Poisson distribution, where the probability of c spikes occurring in a time bin, t = 1 s, for a mean firing rate r is

A cell was considered to be bursting if P_c < 10⁻⁴ in any given bin. We then identified the time bins in which more than 5% of all cells in the recording were bursting with P_c < 10⁻⁴, and hence computed a pair-wise correlation index, CI, as a function of distance between two cells for all spikes in these bins, where

N_AB(∆t) corresponds to the number of spike pairs for which unit B fires within a time window ±∆t from unit A; N_A(T) corresponds to the total number of spikes fired by unit A during the total recording time, T (and similarly for N_B(T)) (17). We used a correlation time window of ∆t = 100 ms. Thus, only spikes in bins that displayed a decreasing nearest neighbor correlation index were accepted as waves and considered for analysis of wave properties. Waves detected using this algorithm agreed with those determined by eye. Upon identification of waves, the wave frequency, burst duration during a wave, and firing rate during a wave were computed and averaged for each unit.

The correlation index was calculated for all cell pairs in each retina. The distance between cells was approximated as the distance between the electrodes of cell pairs. Pairs were then grouped according to intercellular distance and the medians computed over all cell pairs. The median correlation index was then plotted as a function of increasing intercellular distance. To establish connectivity maps, we defined two cells as being connected if their correlation index fell in the top 5% of all correlation index values (26). We then extracted cells that were connected to at least 15% of other cells (red units in Fig. 3 B and C). We mapped these units back onto the electrodes on which they were recorded, and computed the distances to their connected units (blue lines in Fig. 3 B and C).

To identify ipRGCs, we computed the difference in a unit’s mean firing rate in the 60 s following light onset and its mean firing rate in the 60 s preceding light onset. Most cells followed a narrow normal distribution centered about 0 Hz difference (Gaussian fit parameters, μ = 0.07 Hz; σ = 0.35 Hz), which were classified as nonipRGCs. Units that showed an increase in mean firing rate of at least 0.75 Hz (45 additional spikes in a 60-s window) in the light were classified as ipRGCs. This classification came from targeted cell-attached recordings from a transgenic mouse line in which GFP is expressed under the melanopsin promoter (Opn4–EGFP mouse) (29). Cell-attached recordings showed that GFP+ RGCs exhibited an increase in mean firing rate following light onset of at least 0.75Hz, whereas the difference in mean firing rates for non-GFP+ cells fell into a cluster centered around 0 Hz (refer to Fig. S2). Each MEA unit was inspected manually to verify that ipRGCs classified in this manner showed a light response. These units were mapped back on to the electrodes on which they were recorded (turquoise units in Fig. 3 B and C).

Pharmacology.

DHβE (8 μM), d-AP5 (50 μM), DNQX (20 μM), SCH23390 hydrochloride (10 μM), and raclopride (8 μM) were added to Ames’ media as stock solutions prepared in either distilled water (DHβE, d-AP5, DNQX, and SCH23390) or DMSO (raclopride). Antagonists were purchased from Tocris.

Electrophysiology.

Isolated retinas were mounted RGC side up on filter paper over a small viewing hole. Retinas were superfused with Ames’ solution bubbled with 95% O₂ and 5% CO₂ and maintained between 33 and 35 °C, pH. 7.4. Retinas were visualized with differential interference contrast optics on an Olympus BX51WI microscope under a LUMPlanFL 60× water-immersion objective. ipRGCs were identified by GFP signal under epifluorescent illumination at 488 nm. A hole was pierced in the inner limiting membrane of the retina using a glass recording pipette to access the RGC layer. RGCs were targeted under control of a micromanipulator (MP-225, Sutter Instruments). Recording pipettes were pulled with a tip resistance of 4–5 MΩ (Sutter Instruments) and filled with filtered NaCl (150 mM). Data were acquired using pCLAMP 10.2 recording software and a Multiclamp 700B amplifier (Molecular Devices), sampled at 6 kHz and filtered between 120 and 2,000 Hz.