Short pulses of epiretinal prostheses evoke network-mediated responses in retinal ganglion cells by stimulating presynaptic neurons

Experiments were performed following the institutional and national guidelines for animal use and care. An animal experiment protocol was approved by the Institutional Animal Care and Use Committees of KIST (KIST-2021-11-152). Wild-type (C57BL/6J) and rd10 (B6.CXB1-Pde6b^rd10/J) mice (postnatal days ranging from 56 to 70 for seven wild-type mice and 207, 213, and 383 for three rd10 mice) were anesthetized with vaporizing isoflurane and euthanized by cervical dislocation. After the enucleation of eyeballs, retina tissues were isolated and mounted on a filter paper, the photoreceptor layer facing down. Through a small hole (∼2 mm) made at the center of the filter paper, light stimuli were delivered to identify RGC types.

For electrophysiological recordings, patch pipettes (9–12 MΩ) filled with Ames medium were used at cell-attached mode. The inner limiting membrane was removed with patch pipettes prior to recording spiking activities from RGCs. We targeted RGCs that have large somata (>20 µm), which is a hallmark of alpha RGCs of the mouse retina (Pang et al 2003, Murphy and Rieke 2006). As ground electrodes, two silver chloride-coated silver wires were placed near the wall of the recording chamber, separated ∼6 mm from each other. Signals from the patch electrode were recorded and low-pass filtered at 2 kHz using an amplifier (MultiClamp 700B, Molecular Devices, Sunnyvale, CA). The recorded signals were digitized via a data acquisition card (PCI-MIO-16E-4, National Instruments, Austin, TX). Throughout the recording, retinal tissue was continuously superfused with oxygenated Ames medium at 4 ml min⁻¹ and the temperature was kept around 34 °C–36 °C.

Light stimuli were generated by a custom software written in MATLAB (MathWorks, Natick, MA) and LabVIEW (National Instruments, Austin, TX). An LCD projector (CineBeam PH550, LG) displayed light stimuli through a reflection mirror installed under the condenser of an upright microscope (Nikon FN1). Light stimuli made a focus at the photoreceptor outer segments of the mounted retina sample.

We classified RGCs into either ON or OFF types based on their spiking response to 1 s long stationary white spot flashes which were centered at the targeted RGC soma. We tested spot sizes ranging from 100 to 1000 µm; one single spot was presented on gray background each time. All light stimulation was repeated at least three times. In the present study, we recorded and analyzed six ON and six OFF RGCs from wild-type mice. Also, we recorded 11 RGCs from three rd10 mice but were not able to classify the rd10 RGCs due to their weak or almost no spiking responses to the spot flashes.

Electric current stimuli were epiretinally delivered to retinal samples using 10 kΩ platinum–iridium electrodes (MicroProbes, Gaithersburg, MD). The electrode was insulated other than its exposed tip. The height of the conical tip was ∼125 µm and the base diameter was 30 µm. Its conducting surface area is ∼5900 µm², which is comparable to that of a disk electrode of 87 µm in diameter. The position of the stimulating electrode was controlled by a micromanipulator; the electrode was placed ∼25 µm above the inner limiting membrane of the retina.

To comprehensively assess the effects of short pulse electrical stimulation, we varied number of pulses (N), pulse duration (D), current amplitude (A), and stimulation frequency (f) (figure 2(A)). All stimulation conditions we tested were summarized in table 1; all short stimuli consisted of cathodal-first biphasic pulses without inter-phase interval. We first tested 460 µs long pulses which were used in Argus II Retinal Prosthesis System. Some literatures reported the pulse duration of Argus I or II as 450 µs (Nanduri et al 2012, Ahuja and Behrend 2013, Beyeler et al 2019), however, we set the duration as to be 460 µs to follow Argus II Surgeon Manual (Second Sight Medical Products 2013). At a constant frequency of 20 Hz, we varied pulse counts (one, three, five, seven, and ten times) as well as current amplitudes (100, 200, 300, and 400 µA); clinical tests of Argus II delivered current amplitudes up to several hundred microamperes (Ahuja and Behrend 2013). Additionally, we tested different frequencies (i.e. 10 and 40 Hz). To further explore the RGC responses to even shorter pulses, we also delivered ten biphasic pulses of 100, 200, and 300 µs long stimuli at the fixed frequency of 20 Hz in various amplitudes (100, 200, and 300 µA). Lastly, we also applied a long pulse (monophasic cathodal current, 4 ms long, −100 µA) to compare with the results of other previous works reporting strong network-mediated responses (Im and Fried 2015b, 2016, Yoon et al 2020, Otgondemberel et al 2021). Previous work reported a minimal difference between spiking activities of a given RGC in responses to cathodal-only monophasic vs. biphasic pulses (Im and Fried 2016).

Zoom In Zoom Out Reset image size

Spike timing was detected from raw recording with the custom software written in MATLAB. To solely evaluate network-mediated responses of RGCs, the spikes obscured by artifact were not analyzed (highlighted with a gray band in figure 2(B)) because they are known to be results of direct activation of RGCs (Fried et al 2006, Sekirnjak et al 2008). Spikes were counted for 1 s from the offset of the last stimulus (as indicated with a red arrow in figures 2(A) and (B)) and averaged across multiple repeats. Then, the color-coded heat maps of indirect spiking activities (highlighted with a pink band in figure 2(B)) were created with 20 ms long bins and a 1 ms long moving step (figures 3, 7, and 8). To calculate spontaneous firing rate, we counted the number of spikes during 0.5 s of the pre-stimulus period and divided it by 0.5 (figure 3(D-iii)).

Zoom In Zoom Out Reset image size

Spike timing consistency was examined by calculating the spike time tilting coefficient (STTC) across repeats. The equation for STTC is as follows (Cutts and Eglen 2014):

$$\begin{align*}{\text{STTC}} = \frac{1}{2}\left( {\frac{{{P_{\text{A}}} - {T_{\text{B}}}}}{{1 - {P_{\text{A}}}{T_{\text{B}}}}} + \frac{{{P_{\text{B}}} - {T_{\text{A}}}}}{{1 - {P_{\text{B}}}{T_{\text{A}}}}}} \right)\end{align*}$$

where P_A is the proportion of spikes from the spike train A that lie within time window (±Δt) of each spike from the spike train B. T_A is the proportion of total recording time which has any spikes within ±Δt from the spike strain A. P_B and T_B were similarly calculated. In this study, we used Δt of 50 ms. We computed STTCs between every pair of responses elicited from multiple repeats of identical stimuli and visualized the trial-to-trial response consistency in a form of heat matrix (figures 4(A)–(C)). First, we created STTC matrices of responses arising from a single long pulse (4 ms long, −100 µA) in ON, OFF, and rd10 RGCs (first rows of figures 4(A)–(C)). Then, we created similar matrices of responses arising from short pulses (ten 460 µs long pulses, 400 µA, and 20 Hz) for those RGCs (second rows of figures 4(A)–(C)). The higher STTC value implies the greater spike timing consistency across repeats in a given cell. In addition, the STTC between every pair of both responses (i.e. responses to ten 460 µs long pulses and one 4 ms long pulse) of RGCs for each type were shown as violin plots to statistically visualize the differences across RGC types and stimulus conditions (figure 4(D)).

Zoom In Zoom Out Reset image size

Unless otherwise stated, all data in this work was shown as mean ± standard deviation (SD). Statistical analyses were performed using a one-way analysis of variance (ANOVA) with Holm–Sidak post-hoc comparisons or student's t-test in Origin software; p < 0.05 was considered statistically significant.

We questioned whether short (<1 ms) pulses do not indirectly activate RGCs (i.e. elicit direct responses only). In particular, we wondered if the Argus II stimulation condition evokes any network-mediated response. To quickly examine this, we recorded spiking activities of an ON RGC in response to a series of 460 µs long pulses; ten stimuli of biphasic current amplitude 400 µA were delivered at 20 Hz, which is described in Argus II Manual (Second Sight Medical Products 2013). As expected, the RGC responded with precisely stimulus-locked spikes per each pulse (shown in the gray band of figure 2(B)). Interestingly however, the RGC generated a strong burst of spikes after the last stimulus in each trial (highlighted in the pink band of figure 2(B)). This substantially (>∼100 ms) delayed burst is quite similar to the hallmark feature of network-mediated responses of ON RGCs to much longer stimulus (i.e. 4 ms long pulse) reported in some earlier works (Im and Fried 2015b, Lee and Im 2019). Also, this spike burst that appeared after the stimulus offset is highly likely to be resulted from activation of its presynaptic neurons (i.e. indirect activation of the RGC) because several researchers repeatedly demonstrated using pharmacological blockers that those delayed spiking activities are network-mediated responses (Eickenscheidt et al 2012, Boinagrov et al 2014). Definitely, the result of intra-stimulus recordings (gray band in figure 2(B)) does support previous belief about 'one or two-spike(s)-per-pulse' direct activation (Fried et al 2006, Sekirnjak et al 2006, Jepson et al 2014); but the result of post-stimulus recordings (pink band in figure 2(B)) suggests a possibility that short pulses can produce strong indirect responses, which have not been systematically examined. Thus, characterizing the intact 'normal' retina's network-mediated responses arising from short pulses would be a critical step to deeper understanding of response changes in degenerate retinas.

To more systematically explore network-mediated responses evoked by short pulses, we tested various conditions of stimulating parameters such as pulse duration (D), current amplitude (A), number of stimuli (N), and repetition frequency (f) (table 1). First, we investigated the effect of current amplitude: RGCs were stimulated with ten 460 μs long pulses delivered at 20 Hz in various current amplitudes. Raster plots of representative ON and OFF cells clearly showed their network-mediated response patterns in the two types (figures 3(Ai) and (Bi)).

In the case of the ON cell, the low current amplitudes (100 and 200 µA) generated no or weak network-mediated responses. However, quite strong responses were elicited by higher amplitudes (300 and 400 µA) (figure 3(Ai)): the peak firing rate of the ON RGC reached up to ∼500 Hz and the spiking lasted for >∼200 ms. To isolate the effect of a single stimulus, we also recorded responses to a single 460 μs long pulse of 400 µA in peak amplitude (first raster plot of figure 3(Aii)). For comparison, the same RGC was also stimulated by a 4 ms long (−100 µA) stimulus and its spiking responses were shown as the second raster plot of figure 3(Aii). In this particular cell, the multiple short pulses elicited a strong single burst (the last raster plot of figure 3(Ai)) but the single short pulse and long pulse evoked three bursts of spikes separated by silent periods in-between (both raster plots of figure 3(Aii)). Although the three bursts were in common, the majority of spiking activities appeared at ∼200 ms and ∼100 ms in responses to short and long pulses, respectively.

In most ON RGCs (n = 4/6), the single 460 μs long pulse (400 µA) generated slightly bigger responses than the 4 ms long pulse (100 µA) in terms of spike counts (figure 3(Aiii); note that this is not an equal-charge comparison). But, in all ON RGCs, the spike counts of those two responses seem largely similar. This is particularly interesting because, in each cathodal phase, the long pulse (4 ms, −100 µA) supplied ∼2.17 times more charges compared to the one single short pulse (460 µs, −400 µA). If the equal charges were delivered (4 ms, −46 µA), the long pulse was likely to evoke smaller number of spikes (Im and Fried 2016). Thus, it seems like the short pulse is more charge efficient than the long pulse when a single stimulus is delivered.

In contrast to the responses of ON cells, the responses elicited in OFF RGCs by repeated 460 μs long pulses were generally weaker than single short pulse and long pulse (compare figure 3(Bi) and (Bii)). However, the responses of OFF cells to a single 460 μs long pulse seems generally similar to those to a 4 ms long pulse (figure 3(Biii)), consistent with those of ON cells. The scatter plot of the OFF RGCs (figure 3(Biii)) supports our abovementioned claim that RGCs generated similar numbers of spikes in responses to both 460 μs and 4 ms long pulses (see section 4).

For more clinical implications, it would be essential to test whether our results observed in the wild-type mice are consistent in the retinal degeneration model. Accordingly, we have used three rd10 mice which were 207, 213, and 383 days old, respectively; previous literatures suggest that this rd10 animal was at a severely advanced/saturated stage of retinal degeneration including almost no visual responses as well as a significant alteration in retinal circuits (Chang et al 2002, 2007, Gargini et al 2007, Rösch et al 2014). Interestingly, however, some rd10 cells also evoked quite strong network-mediated responses in both absence and presence of spontaneous activities (figures 3(Ci) and (Cii) vs. (Di) and (Dii), respectively). Spontaneous firing rates of all rd10 RGCs are also plotted (figure 3(Diii)). In particular, the rd10 RGC with the highest spontaneous firing rate (15.71 Hz; marked with a gray arrow in figure 3(Diii)), clearly demonstrated stronger indirect responses with increasing current amplitude (figure 3(Di)). Due to the lack of stable light responses in rd10 RGCs, we were unable to identify their RGC type. The peak firing rate of this RGC in unknown type seemed weaker than those of both RGC types in the normal retina. For example, the maximum firing rates of ON and OFF RGCs were 500 and 350 Hz in responses to ten and single 460 μs long pulse(s) of 400 µA (last row of figure 3(Ai)) and first row of figure 3(Bii), respectively; while that of rd10 was 250 Hz in response to ten short pulses (figure 3(Ci); note the scale bars are different in each panel). The scatter plot also demonstrates similar spike counts between the single short pulse and the long pulse responses in the rd10 RGCs (figure 3(Ciii)).

It seems like both long- and short-duration pulse(s) well evoke the network-mediated responses (figures 3(A)–(D)), however, we noticed somewhat different spiking activities across repeats between the two stimuli conditions. To systematically analyze the trial-to-trial consistency of network-mediated responses, we calculated the STTC of responses evoked by a single 4 ms long pulse (first rows of figures 4(A)–(C)) as well as responses evoked by ten 460 µs long pulses in each cell (second rows of figures 4(A)–(C)). Although the inter-trial consistency of network-mediated responses is known to be dependent on stimulus amplitude (Lee et al 2013), we just plotted the STTC matrices for responses to the highest current amplitude (400 µA) for 460 μs long pulses. Because of the absence of spiking in some OFF and rd10 RGCs, STTC values of those RGCs were neither computed for heat matrices (shown in black color in figures 4(B) and (C)) nor included in violin plots. The heat matrices show most ON,OFF and rd10 RGCs evoked highly correlated spiking patterns in response to both long- and short-duration pulse(s) (figures 4(A)–(C)). To explore this in more detail, we also visualized all STTC values in a form of violin plots (figure 4(D)). The STTC_AVG of responses to the short pulses for ON and OFF cells were significantly lower than those of responses to the long pulse. It seems that long-duration stimulation is more suitable to evoke the consistent network-mediated responses for ON and OFF types.

In the case of rd10 RGCs, the STTCs of long pulse responses were more inconsistent than those for ON and OFF RGCs in the normal retina (p < 0.001; bottom two pairs of statistical comparisons of figure 4(D)). This is consistent with the previous work which demonstrated reduced magnitude and consistency of network-mediated responses (Yoon et al 2020) in the degenerate retina. As argued in that earlier work, the reduction in the trial-to-trial consistency in responses may hinder appropriate perception of evoked artificial vision (see section 4). However, there were no statistical significance between short pulse responses for ON, OFF and rd10 RGCs (two pairs of statistical comparisons marked with n.s. in figure 4(D)). It would be worth to note that theses statistical comparisons may be different if rd10 cells are further classified into ON, OFF, or other physiological types. Although we were not able to identify their cell types due to the lack of light responses, extra anatomical experiments can enable it in future studies (see section 4). The RGCs of rd10 retinas may be grouped depending on the levels of spontaneous activities to further explore its effects on electric responses. For example, the rd10 cells we recorded can be classified into three types (figure 3(Diii)): (a) no spontaneous spiking (n = 4/11), (b) spontaneous firing below their average (n = 4/11), and (c) spontaneous firing above their average (n = 3/11). Due to the stochastic nature of spontaneous activities, STTCs may be lowered but we did not see any significant differences across the abovementioned three groups. Also, those groups were not separately plotted due to the small numbers of samples in each group.

Our earlier results (figures 3(Aii)–(Cii)) indicate that one single short stimulus also activates presynaptic networks but differently across RGC types. For instance, when the number of stimuli reduced from 10 to 1, the ON cell responses became weaker while the OFF cell responses got stronger (compare the last rows of figures 3(Ai) and (Bi) vs. the first rows of figures 3(Aii) and (Bii)). Thus, we examined how a smaller number of short stimuli alters network-mediated response patterns: we presented various numbers of stimuli (i.e. 10, 7, 5, 3, and 1) delivered at 20 Hz. Also, we tested current amplitudes ranging from 100 to 400 μA. Figures 5(A)–(C) summarize average numbers of elicited spikes as a function of current amplitude for ON and OFF RGCs in the healthy retina, and rd10 RGCs, respectively. In all three groups, the magnitudes of responses elicited by a 4 ms long pulse were very similar with those elicited by single 460 μs long pulse (compare the rightmost columns of figures 5(Av)–(Cv) and (Avi)–(Cvi)). Consistent with our earlier plot (figure 3), the ON cells evoked generally higher spike counts than the OFF cells for all stimulation conditions (compare figures 5(A) and (B)).

Zoom In Zoom Out Reset image size

The short-pulse-eliciting indirect responses differ between ON and OFF RGC types. For example, the network-mediated response amplitude of the ON cell increased with the increasing current amplitude (figure 5(Ai)–(Av)). The responses elicited from 100 μA were always significantly smaller than those elicited from 400 μA for all pulse repetitions ranging from 1 to 10 (figure 5(Ai)–(Av)). In contrast, except an outlier OFF cells (marked with a star symbol in figure 5(B)), the influence of stimulation amplitude on response was not that strong in OFF RGCs compared to ON RGCs (compare figures 5(B) to (A)). When one or three short stimuli were applied, the maximum spike counts of OFF cells were observed at 400 μA (11.6 ± 3.8 and 4.3 ± 2.9, respectively; figure 5(Bv) and (Biv)). But, when five or more short pulses were presented, the responses of the OFF cells generated slightly higher spike counts at a current amplitude of 300 μA (2.2 ± 2.8, 3.6 ± 4.3, and, 3.2 ± 3.1, respectively; figure 5(Bi) and (Biii)) than other current amplitudes. However, there was no statistical significance found.

We also plotted the same data as a function of pulse counts for each cell (figures 6(A)–(C)). Contrary to current amplitudes, the number of pulses seems not having influence on the spike counts of ON RGCs (figure 6(A)): regardless of stimulation amplitude (e.g. from 100 to 400 μA), spikes elicited from one, three, five, seven, and ten stimuli are not statistically significant for all pairs. In the case of the OFF RGCs, the single stimulus clearly generated much stronger responses than multiple stimuli (figure 6(B)). Statistical analysis across spike counts elicited by different numbers of short stimuli in 400 μA showed the spike counts from a single short pulse were significantly higher than those from multiple stimuli (p < 0.001; figure 6(Biv)). This result is consistent with the previous paper (Im and Fried 2016) that the spike counts of OFF RGCs were decreased as the number of pulse count increased. Our results also indicate the more biased activation of ON over OFF cells can be achieved by optimal choice of current amplitude and number of pulses. For example, the ON over OFF response ratios, which was computed by the ratio of average spike counts between ON over OFF cells, 1.12 and 20.21 for a single short pulse of 100 μA and ten short pulses of 400 μA, respectively (compare the left-most columns of figures 6(Ai) and (Bi) vs. the right-most columns of figures 6(Aiv) and (Biv); see section 4). Moreover, in the case of ten short pulses of 100 μA, no responses of OFF cells were elicited, which would be ideal for completely selective activation of ON cells; however, the responses of ON cells were too small (3.0 ± 3.6 spikes), which may not be enough to evoke artificial visual percepts.

Zoom In Zoom Out Reset image size

Lastly, we performed similar analyses for rd10 RGCs. In general, the number of spike counts increased with increasing current amplitude which is not different from healthy ON and OFF RGCs (figure 5(C)). But, rd10 RGCs showed independence between spiking counts and number of stimuli (figure 6(C)): no statistical difference among all possible pairs given the same current amplitude. Interestingly, only at 400 μA stimuli, one of the cells generated increasing spike counts as pulse count increased (pointed with an arrow in figure 6(Civ)). It suggests a possibility that this particular cell was a different subtype. However, it was not possible to conduct further analysis (e.g. sub-classification by its light responses) and differentiate it from other rd10 RGCs due to the advanced degeneration.

The frequency of repeating stimuli is one of the important parameters determining presynaptic activation properties (Sekirnjak et al 2006, Freeman et al 2010) and prosthetic visual percepts (Nanduri et al 2012). To investigate the effect of differing frequencies, we applied ten short pulses at 10 and 40 Hz in addition to 20 Hz which was tested for earlier plots. In the healthy retinas, ON and OFF RGCs showed distinct response characteristics to those three different stimulation frequencies (figure 7).

In the case of the particular ON RGC shown in figure 7, its network-mediated responses started ∼100 ms from the offset of the last stimulus for both 10 and 20 Hz (figures 7(Ai) and (Bi)) but slightly delayed further (figure 7(Ci)) for 40 Hz. More interestingly, the ten pulses delivered at 10 Hz showed the biggest enhancement in spike counts of network-mediated responses with increasing current amplitudes (figure 7(Aii)). However, at 40 Hz, the response magnitudes were not considerably increased as a function of the current amplitude; no statistical difference was found across responses arising from different current amplitudes (figure 7(Cii)). As a result, the ON RGCs generated higher spike counts at a lower frequency: for example, the average spike counts elicited by 400 μA were 23.9 ± 19.1, 18.2 ± 6.0, and 12.4 ± 7.4 for 10, 20, and 40 Hz, respectively (the rightmost columns of figures 7(Aii)–(Cii)).

The responses of OFF RGCs also appeared to increase at all frequencies with increasing current amplitude (figures 7(Dii)–(Fii)). However, the responses were not statistically different between each current amplitude, probably because the numbers of spikes were smaller than 10. The onset timings of the strongest responses (i.e. the latency of delayed burst of spikes) differed across the OFF cells: some OFF RGCs showed the latencies of ∼200 ms, similar to those of the ON RGCs (n = 4/6; figures 7(Di)–(Fi)). In contrast, a couple of OFF cells showed the latencies of >∼500 ms (n = 2/6; figure 8), which are substantially more delayed than those of other ON or OFF cells. For instance, one of the OFF cells, which was considered as an outlier in our earlier analyses (marked as a star symbol in figures 3(Biii), 5(B), 6(B) and 7(Dii)–(Fii)), generated a remarkably strong burst of spikes that begins at ∼490 ms (figure 8(Ai)). The other cell also demonstrated a burst of spikes with a similar temporal delay (∼580 ms) which arose only to the current amplitude of 300 μA (figure 8(Bi)). This burst was recorded consistently, suggesting it was not spontaneous activity. The response characteristics of these cells were quite interesting because a single short pulse and/or a single long pulse elicited much different spiking patterns (figures 8(Aii) and (Bii)). Although we did not anatomically characterize these cells our results suggest that there are subtypes of OFF RGCs which are distinct in responses to a series of short pules (see section 4).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Up to this point, RGCs were stimulated by biphasic pulse(s) with a fixed duration of 460 μs for each cathodal/anodal phase. To further explore whether similar responses arise from even shorter pulses, we applied ten biphasic pulses in durations of 100, 200, and 300 μs. When delivered at 20 Hz, the spike counts of ON and OFF RGCs decreased remarkably as the duration decreased (figure 9), probably due to insufficient applied charge to fully activate presynaptic neuronal circuits. However, it is worth to note that network-mediated responses to 300 μs long pulses were still significant at least in the ON RGCs. For example, with the current amplitude of 300 μA, the 460 μs and 300 μs long pulses elicited 15.3 ± 11.3 and 6.9 ± 7.0 spikes in average, respectively. In comparison, the 200 and 100 μs long pulses produced 3.3 ± 4.8 and 0.7 ± 1.2 spikes, respectively. Contrastingly, the OFF RGCs showed puny responses to all those shorter stimuli (1.1 ± 0.7, 0.04 ± 0.08, and 1.1 ± 0.7 spikes for the 300, 200 and 100 μs long pulses in 300 μA). Although we have not tested amplitudes >300 μA for those short durations, it might have been possible to enhance responses by further increasing current amplitude for both ON and OFF types. In particular, in case of 300 μs long pulses, both types demonstrated somewhat elevated spike counts to 300 μA pulses (figures 9(Aii) and (Bii)). This suggests more thorough investigation for a wider range of current amplitudes may unravel the lower limit of stimulus duration for network-mediated responses.

Zoom In Zoom Out Reset image size

It has often been claimed that epiretinal prostheses directly activate RGCs only without activating presynaptic neurons, due to the electrode location they implant and/or the stimulus duration they typically use (Jensen et al 2003, Fried et al 2006, Sekirnjak et al 2008). Also, it has been previously reported that direct selectivity, which was defined as direct response over network-mediated response, was highest (∼4) for the stimulus duration <0.5 ms (Boinagrov et al 2014). Surprisingly, however, the present study clearly demonstrated that short pulses delivered from the epiretinal side can strongly activate presynaptic neurons. Our results have clinical meanings because we tested the identical pulse duration which has long been used in a commercialized epiretinal implant system, Argus I & II (Ahuja et al 2011, Humayun et al 2012, da Cruz et al 2013, Kotecha et al 2014).

It should be emphasized that the present work is not the first report about the network-mediated responses elicited by stimulus pulses <1 ms as previous work reported indirect responses evoked such short pulses (Tsai et al 2009, Eickenscheidt et al 2012, Lee et al 2013, Boinagrov et al 2014). But, it might be no surprise because most of those studies used subretinal stimulation which has long been believed to activate presynaptic neurons (Zrenner 2002, Jensen and Rizzo 2006, Palanker et al 2020). Several previous researchers regarding epiretinal stimulation have focused on avoiding axon bundle activation (Jensen et al 2003, Behrend et al 2009, Grosberg et al 2017, Tong et al 2020) because short pulses that activate axon fibers are known to create streaks, blobs, and wedges (Nanduri et al 2012, Beyeler et al 2017, 2019). But actually, short pulses can not only directly activate RGCs/axons but also evoke network-mediated responses, suggesting a necessity of extra investigation of short pulses from the indirect activation point of view. It would be important to pay attention to the latencies of strong indirect responses which were distinct across the cell types. In the case of ON RGCs, the spike occurred within 200 ms at most, and after 500 ms in the case of some OFF RGCs (figures 3 and 8). This might be caused by the different retinal circuit structures of ON and OFF RGCs. Also, it seems that more information about the latency and spiking pattern of ON and OFF cells can be collected by further classifying subtypes which show slightly different response patterns even to electrical stimulation (Werginz and Fried 2019). However, we did not classify subtype of RGCs in the present study.

It has been reported that the synaptically-mediated response is effectively stimulated by long stimulus because of the slow activation speed of voltage-gated calcium channels than voltage-gated sodium channels (Freeman et al 2010, Twyford and Fried 2016). Indeed, there are some previous studies that used relatively long-duration pulse (>1 ms) to activate bipolar cells (Greenberg 1998, Jensen et al 2005, Fried et al 2006, Jensen and Rizzo 2006, Freeman et al 2010). Therefore, it would be surprising that 460 μs long pulses can strongly activate presynaptic neurons as well, resulting in indirect activation of RGCs. Due to the well-known activation kinetics of voltage-gated sodium and calcium channels, the short pulse is highly likely to primarily activate sodium channels. Given the fact that photoreceptors have no sodium channels (van Hook et al 2019), those network-mediated responses shown in the present study are likely to be resulted from the activation of voltage-gated sodium channels expressed in bipolar cells (Puthussery et al 2013) and/or different type(s) of neuronal classes such as amacrine (Maguire 1999, Heflin and Cook 2007) and/or horizontal cells (Shingai and Christensen 1983, Malchow et al 1990, Golard et al 1992). Recently, however, Werginz et al (2020) reported calcium channels can also be activated by short pulses (0.1–2 ms) at high current amplitude, which is in parallel with our results that the spike counts increased with the higher current amplitude (figures 5(Av), (Bv) and (Cv)). Revealing the exact mechanism underlying the short-pulseinitiated network-mediated responses would be helpful to better understand how retinal implant users who were stimulated with short pulses (i.e. such as Argus II users) perceived artificial visual precepts.

The network-mediated responses of ON cells showed a higher correlation with current amplitude than OFF cells (figures 5–7 and 9). For example, the spike counts in ON RGCs were greatest at the maximum current amplitude of 400 μA regardless of the number of stimulation repetitions (figures 5 and 6). This increasing response of ON RGCs with increasing current amplitude is not surprising as shown in many previous papers (Stett et al 2000, 2007, Tsai et al 2009). Recently, it has been reported stronger responses arise from the higher current amplitude at a fixed charge condition (Im et al 2018). Another previous work also demonstrated peak current amplitude is important in resulting spike counts in network-mediated responses of ON RGCs (Lee and Im 2018). Taken all together, our results suggest the peak current amplitude plays an important role in optimizing network-mediated responses in ON RGCs. However, since we used a much smaller stimulating electrode than Argus II, it may be difficult to directly compare our results with clinical results with the actual implant. Because a previous paper reported the threshold current for direct response is dependent on the electrode side (Sekirnjak et al 2006), further studies seem to be needed to investigate the relationship between the electrode size and the threshold for indirect response.

Contrast to ON RGCs, the maximal responses of OFF cells at the intermediate current amplitude (i.e. 300 μA; figures 7(Dii)–(Fii)) may need to be further explored. In a previous study, the highest spike counts evoked from 3 ms long pulse than 10 and 50 ms long pulse at the same current amplitude (Lee et al 2013). In other words, even though more charges were supplied, it does not mean always resulting in stronger activation of presynaptic terminals. There is another possibility that the rate of charge injection may be important. To tease out those two, additional experiments should be performed.

In addition, it is important to know the minimum amount of charge to effectively generate indirect responses. When we stimulated the ON and OFF RGCs using pulses shorter than 460 μs (i.e. 100, 200, and 300 μs), the spike counts decreased rapidly according to the decreasing pulse duration. Therefore, it is important whether or not each short pulse has enough charge to activate cells. Previous literature reported that the spike counts did not increase at all with the extremely short duration even if the amplitude was increased (Lee et al 2013).

The number of spikes in ON RGCs were maintained with the increasing number of stimuli (figures 5(A) and 6(A)). In contrast, the indirect responses of OFF RGCs were decreased (figures 5(B) and 6(B)), probably due to the desensitization (Jensen and Rizzo 2007, Freeman and Fried 2011, Im and Fried 2016). Although we did not further analyze intra-stimulus responses (e.g. highlighted in gray in figure 2(B)), it is likely that the first stimulus might have evoked more than a single spike in other cells, similar to what was reported by the earlier work (Im and Fried 2016).

The performance of retinal prosthetics is highly likely to be dependent on the different levels of degeneration, particularly in the outer retinal structures such as outer and inner segments of photoreceptors. However, even before characterizing behaviors of electrically evoked responses in the degenerate retina, it would be critical to thoroughly understand how RGCs in the 'normal' retina respond.

It has long been believed that short pulses (e.g. several hundred microseconds) directly activate RGCs without activating presynaptic neurons. However, our results from the wild-type mice indicate that short pulses can actually evoke network-mediated responses induced by presynaptic neurons including bipolar cells and/or photoreceptors of the healthy retinas. Also, it is important to note that the rd10 cells of the severely degenerate retinas showed somewhat strong network-mediated responses (figures 3–6) although rd10 mice >P200 are known to have almost no photoreceptor layer (Chang et al 2007, Rösch et al 2014). This result indicates that Argus I and II which use the cathodic-first biphasic pulse with 460 μs of pulse duration at 20 Hz may also elicit the synaptically-mediated responses. Thus, there is a possibility that indirect responses contribute to artificial visual percepts.

Moreover, it is worth to note that, in the clinical trials, the artificial perception was not much changed even with a stimulation frequency change from 20 to 80 Hz (Nanduri et al 2012), which were likely to evoke direct spikes with inter-spike intervals (ISIs) of 50 and 12.5 ms, respectively. However, distinct clinical outcomes between the two frequencies are rather expected because the spikes with ISIs > 30 ms were not only extremely few in light responses of the healthy retinas (Im and Fried 2015b) but also known to have no measurable influence on the lateral geniculate nucleus (Usrey et al 1998, Rathbun et al 2007). The similar visual percepts arising from those two stimulation frequencies (i.e. 20 and 80 Hz) may be more reasonably explainable with indirect responses because similar bursts of spikes were consistently evoked at various stimulation frequencies ranging from 2 to 100 Hz (Im and Fried 2016).

Also, it has long been doubted why retinal prosthetics users perceived 'bright' phosphene although both ON and OFF types of RGCs are known to be indiscriminately activated by electric stimulation. Previous work suggested a possibility of better matching between visually- and electrically-evoked responses in ON RGCs may contribute in biased perception of bright sensation (Im and Fried 2015b). Our results presented here suggest another possibility that responses of the ON pathway may be better perceived due to bigger network-mediated responses than the OFF pathway (compare figures 7(A)–(C) vs. (D)–(F)). The different response magnitudes of ON vs. OFF RGCs to repeating short stimuli improves differential activation of ON vs. OFF RGCs: compared with a single pulse, ten pulses enhances/reduces network-mediated responses in ON/OFF types, respectively (compare last rows of figures 3(Ai) and (Bi) vs. first rows of figures 3(Aii) and (Bii) or compare figures 5(Ai) vs. (Bi)). During short pulse repeats, both ON and OFF types transmit direct spikes to downstream visual centers (highlighted in gray of figure 2(B)), suggesting a challenge regarding discern neural signals in those two pathways. But, post-stimulus indirect responses arise from short pulse repeats (highlighted in pink of figure 2(B)) with greater numbers of spikes in ON than OFF RGCs.

In the degenerate retinas, we were not able to unambiguously identify RGC types into either ON or OFF due to the lack of their light responses. Although our results of rd10 cells somewhat resemble those of ON cells of the normal retinas in some aspect (figure 6), additional sophisticated experiments are needed because our rd10 cells are likely to be heterogeneous in terms of physiological types (i.e. ON, OFF, or direction-selective and so on). For example, to verify whether our findings for each type in the healthy retinas persist after severe degeneration, RGCs should be classified by examining their dendritic stratification depths after physiological recordings (Margolis et al 2008). Previous studies reported the retinal degeneration increases number of OFF cells compared to that of ON cells (Marc et al 2007, Jones et al 2011, 2016), then the delivery of multiple short pulses is likely to reduce the overall (i.e. ON and OFF RGCs as a whole) retinal responses (see figures 5 and 6). Also, future studies are essential to investigate the effects of short vs. long pulses depending on the level of spontaneous firings. Our results showed the network-mediated responses arising from both short and long pulses are somewhat obscured by increased spontaneous activities (compare figures 3(Ci) and (Cii) vs. (Di) and (Dii)), which is known to happen during retinal degeneration (Stasheff 2008).

One of the clinical challenges that need to be addressed is that even when the same retinal prosthetic is implanted on degenerate retinas, the performance of the device varies from patient to patient. Consistent with previous work (Yoon et al 2020), the network-mediated responses arising from long pulses were somewhat more consistent across repeats of identical stimuli than those arising from short pulses in both ON and OFF types of RGCs in the healthy retinas (blue and red violin plots of figure 4(D)). However, in the degenerate retinas, this benefit of long stimuli in terms of the response consistency seems disappeared; rather, the consistency was slightly better in responses to short than long pulses (p < 0.001; purple violin plots of figure 4(D)).

Like it had been argued in the earlier work, this reduced consistency compared to the normal retina may have been related with the difficulty in perception of artificial vision, maybe due to the reduced neural information (Kang et al 2021, Kim et al 2022). The performance of the clinical device may be affected by the level of surviving retinal circuit of each patient. Therefore, in order to successfully apply a retinal prosthesis, it is important to measure the remaining retinal layers to find an appropriate patient group with a highly accurate method such as a spectral-domain optical coherence tomography (Pfau et al 2022). Based on the more accurate anatomical information about degenerate retinas of potential prosthetic users, it may be possible to more efficiently select proper stimulating methods such as direct/indirect activation.