Activation of ganglion cells and axon bundles using epiretinal electrical stimulation

Retinal preparation.

Electrophysiology data were recorded from primate retinas isolated and mounted on an array of extracellular electrodes as described previously (Jepson et al. 2013). Eyes were obtained from terminally anesthetized rhesus macaque monkeys (Macaca mulatta, male and female, ages 4–20 yr) used for experiments in other laboratories, in accordance with Institutional Animal Care and Use Committee guidelines. After enucleation, the eyes were hemisected and the vitreous humor was removed. The hemisected eye cups containing the retinas were stored in oxygenated bicarbonate-buffered Ames’ solution (Sigma) during transport (up to 2 h) back to the laboratory. Patches of intact retina 3 mm in diameter were placed in culture (Ames’ solution, gassed with 95% O₂ and 5% CO₂) with fluorescence-conjugated peanut agglutinin or isolectin GS-IB4 (AlexaFluor 568, catalog no. L32458 or I21412; Life Technologies) at 4.14 μg/ml (36.5 pM) for 5–10 h for vasculature labeling before proceeding with the dissection. The retina was then isolated from the pigment epithelium under infrared illumination and held RGC-side down on a custom multielectrode array (see below). Throughout the experiments, retinas were superfused with Ames’ solution at 34°C.

Electrophysiological recording and stimulation.

A custom 512-electrode stimulation and recording system (Hottowy et al. 2008, 2012) was used to apply electrical stimuli and record spikes from RGCs. The electrode array consisted of 512 electrodes in a 16 × 32 isosceles triangular lattice arrangement, with 60-μm spacing between electrodes within rows and between rows (Litke et al. 2004). Electrodes were 10 μm in diameter and electroplated with platinum. For recording, raw voltage signals from the electrodes were amplified, filtered (43–5,000 Hz), and multiplexed with custom circuitry. These voltage signals were sampled with commercial data acquisition hardware (National Instruments) at 20 kHz per channel. For stimulation, custom hardware (Hottowy et al. 2012) was controlled by commercial multifunction cards (National Instruments). Charge-balanced triphasic current pulses with relative amplitudes of 2:−3:1 and phase widths of 50 μs (total duration 150 μs) were delivered through one electrode at a time. Reported current amplitudes correspond to the magnitude of the second, cathodal, phase of the pulse. This pulse shape was chosen to reduce stimulation artifact in the recordings. Custom circuitry disconnected the recording amplifiers during stimulation, reducing stimulation artifact and making it possible to identify elicited spikes on the stimulating electrode and nearby electrodes (Hottowy et al. 2012; Jepson et al. 2013). For recording and stimulation, a platinum ground wire circling the recording chamber served as a distant ground.

Electrical images and cell type classification.

Recordings obtained with visual stimulation were analyzed to identify spike waveforms of distinct RGCs in the absence of electrical stimulation artifact, using spike sorting methods described previously (Field et al. 2007; Litke et al. 2004), which identified spike times of identified RGCs on the basis of relatively large, stereotyped spikes detected near the soma. The complete spatiotemporal signature of the spikes from each cell over all electrodes, or electrical image, was then computed by averaging the voltage waveforms on all electrodes at and near the times of its recorded spikes (Litke et al. 2004) (see Fig. 2). The electrical image of each cell provided a template of its spike waveform. This was used to identify the cells producing spikes in response to electrical stimulation (Fig. 1E).

Distinct RGC types were identified by their distinct responses to white noise visual stimuli. Briefly, a dynamic random checkerboard stimulus was presented, and the average stimulus that preceded a spike in each RGC was computed, producing the spike-triggered average (STA) stimulus (Chichilnisky 2001). The STA summarizes the spatial, temporal, and chromatic properties of light responses. Features of the STA were used to segregate functionally distinct RGC classes. Spatial receptive fields for each cell type (see Fig. 8) were obtained from fits to the STA (Chichilnisky and Kalmar 2002). For each identified RGC type, the receptive fields formed a regular mosaic covering the region of retina recorded (Devries and Baylor 1997; Field et al. 2007), confirming the correspondence to a morphologically distinct RGC type (Dacey 1993; Wässle et al. 1981), and in some cases revealing complete recordings from the population. The density and light responses of the four most frequently recorded RGC types uniquely identified them as ON and OFF midget, and ON and OFF parasol, which collectively account for 68% of RGCs in primates (Dacey 2004). Other RGC types were encountered but not identified. The regular mosaic structure of RGC receptive fields of each type (Chichilnisky and Kalmar 2002; Devries and Baylor 1997; Gauthier et al. 2009) was used to estimate the total number of cells present over the array (see Table 1). For the purposes of estimating cell type density, it was assumed that the ON/OFF density ratio was the same for parasol and midget cells. Analysis of other data sets (not shown) suggests a possible departure from this assumption: the ON/OFF density ratio appears to be closer to 1 for parasol cells. However, given that it has a small effect on the results, we do not attempt to estimate or use this differential in the present analysis. The total number of RGCs expected to be present over the array was estimated as [m₍₋₎ + m₍₊₎ + p₍₋₎ + p₍₊₎]/0.68, where m₍₋₎, m₍₊₎, p₍₋₎, and p₍₊₎ represent the number of OFF midget, ON midget, OFF parasol, and ON parasol cells over the array, respectively (Dacey 2004).

Identification of axon bundle activation.

To analyze electrically evoked activity over the entire array, voltages were recorded on all electrodes immediately following stimulation with 15–25 trials (repeats) of each electrical stimulus condition. Mean voltage waveforms recorded with the lowest stimulation current amplitude for a particular electrode were subtracted from data recorded with all higher stimulation amplitudes at that electrode, to reduce (but not eliminate) the electrical stimulus artifact. To then identify axon bundle activation, either human inspection or an automated algorithm was used. For human inspection, the observer viewed movies of the recorded activity following stimulation and identified the lowest current amplitude at which bidirectional propagation was visible. Analyzing the results from 6 preparations required viewing >15,000 movies of activity for all different electrodes and stimulus amplitudes. For automated bundle detection, the responses to electrical stimulation were mapped to a collection of weighted graphs, and graph partitioning and graph traversal algorithms were applied to identify bundle activity. The focus was on two characteristic features of axon bundle signals: bidirectional propagation, and growth of signal amplitude with stimulation current. The algorithm and an evaluation of its sensitivity with synthetic data are described in detail in the Appendix.

Identification of somatic activation.

Voltage traces recorded on the stimulating electrode immediately following stimulation were analyzed to find the lowest current amplitude that produced reproducible somatic activation. We use this term to refer to activation by an electrode that records a somatic spike from the cell, as revealed by the characteristic large voltage deflection on the stimulating electrode that begins with a dominantly negative component. This analysis was performed irrespective of the presence of axon bundle activation. Note that it is possible and indeed likely that the site of activation is actually the axon initial segment (Fried et al. 2009; Sekirnjak et al. 2008). To obtain the somatic activation thresholds in a semiautomated way, two features of electrical stimulation data were utilized: elicited spikes are stochastic for stimulation current levels near activation threshold for a given RGC, and the timing of elicited activation is consistent across trials (Jepson et al. 2014a; Sekirnjak et al. 2008). Therefore, recorded voltage traces following stimulation were divided into distinct groups when there was RGC activation on some but not all trials, as shown in Fig. 1A, top row. Principal component analysis (PCA) across trials was used at each stimulation amplitude, and clusters in principal component (PC) space were used to identify the amplitudes for which trials fell into distinct groups (Fig. 1A, bottom row). A fuzzy c-means clustering algorithm (MATLAB and Fuzzy Logic Toolbox release 2014b; The MathWorks) was used to find two clusters in the first two PCs of the data. In this algorithm, each PC data point was assigned a membership grade (between 0 and 1) for each cluster on the basis of distance to the cluster centers. A threshold on membership grades for all points in a cluster was used to identify which stimulation amplitudes produced two distinct clusters, indicating the presence of spikes on some trials but not others. For a cluster to be considered “distinct,” all members were required to have membership grades of 0.8 or higher, with 2 outliers allowed (out of a total of 25 trials). These parameters were chosen so that the analysis erred toward positively identifying clusters, minimizing false negatives. Clustering was confirmed by human inspection to remove false positives, and a cumulative Gaussian function was fit to the response probabilities to determine the 50% activation threshold (Fig. 1B).

After identification of RGC spikes for current levels near the threshold of activation, the remaining analysis steps were performed to determine whether the elicited waveforms were consistent with spikes from a single cell, and if so, to identify the cell type of the activated cell. These steps were performed by human inspection. First, the mean of the nonspiking trial responses was subtracted from the mean of the spiking trial responses for the stimulating electrode and surrounding electrodes. The resulting residual waveform was checked for consistency across stimulation amplitudes, to confirm the activation of a single cell (Fig. 1D). Responses producing multiple distinct residual waveforms were counted as failures of selective activation (see Figs. 6–9). Next, these residual waveforms (Fig. 1D) were compared with the template waveforms (electrical images) of cells recorded with visual stimulation (Fig. 1E), to identify the cell type of origin. Responses from cells not identified from visual stimulation templates were counted as “other” cell types (see Fig. 8).

Immunohistochemistry and imaging.

After stimulation and recording, wide-field fluorescence imaging was used to visualize axon bundles in the recorded patch of retina. Immunolabeling was performed as described previously (Li et al. 2015). Tissue was fixed with 4% paraformaldehyde in phosphate-buffered saline (PBS; 10 mM) for 45 min at room temperature and then washed in PBS 3 times for 10 min and left in PBS at 4°C for 6–48 h. Fixed tissue was kept for 2 h in blocking solution at room temperature and then incubated in blocking solution with a 1:200 dilution of primary antibody (rabbit monoclonal anti-βIII tubulin, Abcam catalog no. ab52623 or Covance catalog no. MRB-435P) for 2–3 days at 4°C on a shaker. The blocking solution consisted of 10% normal donkey serum and 0.25% Triton X-100 in PBS. In some cases, 0.05% sodium azide was added as a preservative. After incubation with primary antibody, the tissue was washed in PBS 3 times for 10 min. Tissue was then incubated in blocking buffer with fluorescence-conjugated (either AlexaFluor 488 or Cy3) secondary antibodies (goat anti-rabbit IgG H&L, either Abcam or Jackson ImmunoResearch) at a 1:200 dilution for 3–4 h at room temperature on a shaker, washed in PBS 3 times for 5 min, counterstained with 4,6-diamidino-2-phenylindole for 10 min, and mounted on slides in ProLong Gold antifade medium (Life Technologies). Coregistration of the electrode array to the fluorescence image (Fig. 2A) was performed as described by Li et al. (2015). An image of the electrode array and the retinal tissue with labeled vasculature was obtained during the experiment, and a second image of the labeled vasculature was obtained after fixation. Control points were chosen from these images, and a local weighted mean transformation was used to register the postfixation image to the live tissue image and array coordinates.

Bundle activation can be identified by its electrical signature.

To identify the electrical signatures of RGCs and axons overlying the electrode array, the electrical image of every RGC recorded during visual stimulation with spatiotemporal white noise was calculated (see materials and methods). The electrical image of a cell is the average spatiotemporal voltage pattern produced across the array during a spike (Greschner et al. 2014; Petrusca et al. 2007). All RGCs in each preparation had electrical images comprising signals that apparently initiated near their somas and propagated along their axons en route to the optic disk (Fig. 2B). As expected in peripheral recordings, the axons in the electrical images were approximately parallel to one another. An image of axon bundles obtained with antibodies to tubulin after recording (see materials and methods) revealed that the orientation of axon bundles corresponded closely to the orientation of the axons inferred from the electrical images (Fig. 2A).

Responses to electrical stimulation were then compared with these electrical images, in some cases revealing the activation of an individual RGC. This was observed by examining the mean voltage deflection recorded on each electrode after current was passed through a selected electrode, after subtraction of the artifact at the lowest stimulation amplitude (see materials and methods). In general, a large signal, composed of physiological signal and stimulus artifact, was observed near the stimulating electrode immediately after the stimulus (Fig. 2, C and D). A voltage deflection then propagated away from the stimulating electrode. In some cases, the starting point and unidirectional propagation of the elicited activity closely matched the electrical image of one of the RGCs obtained during visual stimulation, indicating that this specific cell was activated in isolation (Fig. 2C; single-cell activation was confirmed using the method described in Fig. 1). In other cases, the recorded signal originated at the stimulating electrode and propagated bidirectionally along the axon bundles (confirmed by imaging; Fig. 2A), indicating that passing axons of one or more RGCs had been activated by the electrical stimulus (Fig. 2D) (Abramian et al. 2015). In both cases, the speed of propagation was 1.1 ± 0.39 m/s (measured on 1,245 electrodes in 3 retinas), consistent with action potential propagation speeds recorded from single axons measured using the same experimental methods (Li et al. 2015). However, in the case of bidirectional propagation, the spatial pattern of the recorded signal did not match the pattern of any single electrical image obtained in the preparation (e.g., Fig. 2D). This mismatch suggests that the signal comprised axons of multiple RGCs and/or axons not identified during visual stimulation.

In the case of bidirectional propagation, several additional aspects of the recorded voltage signal provided information about the cells and axons activated. The average amplitude of the recorded waveforms increased progressively over a range of stimulation currents (Fig. 3C), as did the amplitude of recorded waveforms on individual trials. These observations are consistent with the recruitment of multiple axons and inconsistent with the all-or-none signals that would be produced by a spike in a single neuron. Furthermore, stepwise increases in response amplitude (Fig. 3C) suggested that different axons were recruited at different stimulation current levels. In some cases, cross-sectional voltage profiles (orthogonal to the direction of axon propagation) exhibited different forms at different current levels (Fig. 3B), suggesting the recruitment of distinct axon bundles at different current levels. These observations were consistently observed at the large majority of stimulating electrodes tested in eight preparations.

The number of axons in the evoked signal was estimated by comparing the response obtained at high stimulus amplitudes to the mean single-axon response amplitude. The single-axon amplitude was estimated using data from the electrical images obtained without electrical stimulation in the same preparation. The amplitudes of the waveforms on electrodes that recorded axonal activity (identified by a triphasic waveform on that electrode) were averaged for each recorded RGC in a preparation, and the average of all RGCs was used as the estimate for the single-axon signal. For example, in one case (Fig. 3C, right), the peak voltage deflection observed in response to the maximum amplitude stimulation (3.7 μA), which produced a saturating response, was 27-fold larger than the mean voltage deflection associated with individual identified axons in the same preparation. Over all electrodes along the activated path, shown in yellow in Fig. 3A, the mean recorded signal was 13-fold larger than the average single-axon signal amplitude. In the same preparation, the mean of the signals recorded along the activated paths produced after stimulation at each of the 512 electrodes was 12-fold (±5-fold, 1SD) larger than a single-axon signal. Note that these estimates of the number of axons could be biased by the spatial arrangement of axons stimulated and recorded, and the selection of cells for which electrical images were obtained and used for single-axon amplitude estimates. With this caveat, the results suggest that tens of axons from several nearby bundles were activated at the maximum stimulus amplitude and that the stepwise activation of axons probably reflected single axons or small groups of axons, perhaps comprising different bundles, being recruited at specific current levels.

On the basis of these observations, in what follows the evoked bidirectional signal will be referred to as axon bundle activation, although it is possible that in some cases only one axon was activated at low stimulus current levels (see Fig. A3).

Axon bundle activation thresholds vary over electrodes and preparations.

The form of the current-response curve permitted the identification of a threshold for bundle activation. Specifically, in most cases the recorded signal amplitude exhibited a discrete change in slope at a particular stimulation current level (e.g., 1.2 μA in Fig. 3, B–C). The threshold amplitude for the activation of bundles was identified for all electrodes (excluding array borders) in each preparation by examination of the voltage recordings after a stimulation pulse, focusing on the presence of bidirectional propagation (see materials and methods). The threshold for activation was defined as the midpoint between successive stimulation amplitudes at which bidirectional activity was and was not observed. Bundle threshold typically varied severalfold across electrodes (Fig. 4). For six preparations from five retinas, the mean axon bundle activation threshold was 1.34 ± 0.58 μA (mean ± 1SD). This value is comparable to the thresholds for somatic stimulation of individual RGCs (see below) (Jepson et al. 2013).

Different recordings from peripheral retina exhibited substantially different axon bundle thresholds. For each of the six preparations described above, the observed thresholds were 1.87 ± 0.58, 0.96 ± 0.35, 1.48 ± 0.51, 1.41 ± 0.50, 0.84 ± 0.26, and 1.14 ± 0.41 μA (mean ± 1SD across electrodes for each preparation). The differences between preparations were statistically reliable. This was determined by pooling the thresholds obtained in all six recordings, resampling with replacement from this pool to create six distinct groups, each containing the same number of observations as one of the preparations, and computing the RMS difference between the means of the groups. The observed value of 2.07 in the real data was well outside the range of the resampled data, 0.16 ± 0.11 (mean ± 2SD), indicating that the thresholds from different preparations were not drawn from a single population. Diverse factors could contribute to this variation, including spatial variations in the axon layer thickness and pressing of the retina against the electrode array.

An automated algorithm for bundle activation detection (see materials and methods and appendix) provided estimates of activation threshold similar to those identified by manual analysis. Comparison of the manual and automated analysis revealed a 0.93 correlation between the methods over 1,885 electrodes examined in 5 preparations from 4 retinas (Fig. 5A). Sixty-two percent of the thresholds identified were identical, and 88% of the thresholds agreed within ±1 experimental stimulation amplitude step (each step corresponded to a 10% increment in stimulation current; Fig. 5B). For this analysis, stimulation electrodes at the border of the array were ignored because of the difficulty of detecting bidirectional propagation at the border in manual analysis. Importantly, the algorithm captured the growth of the recorded signal with stimulation amplitude (as shown in Fig. 3C) and used it to help identify bundle activation, instead of relying on bidirectional propagation alone. The algorithm sometimes defined a bundle threshold lower than the value obtained by manual analysis. In 80 of 100 such cases examined, subsequent inspection (by viewing the same movies of the recorded activity, but plotted on a more sensitive scale) revealed the presence of an apparent bundle signal not initially detected with manual analysis. For each of the five preparations included in the comparison (preparations 2–6 in Fig. 4B), the thresholds identified by the algorithm were 0.93 ± 0.34, 1.46 ± 0.52, 1.39 ± 0.50, 0.86 ± 0.31, and 1.17 ± 0.44 μA (mean ± SD across electrodes for each preparation). Note that thresholds manually identified by different observers also differed somewhat (Fig. 5C). On a subset of the data reported in Fig. 5A, the correlation between bundle thresholds defined by two observers was 0.97, a value comparable to the correlation observed between the automated and manual methods.

Somatic activation is sometimes possible without axon bundle activation.

Electrical stimulation at current amplitudes near bundle activation threshold often permitted the selective activation of individual RGCs. Activation of RGCs was determined by examining the voltage recordings on the stimulating electrode and surrounding six electrodes, immediately following stimulation, for the presence of large-amplitude spikes with a spatiotemporal waveform closely matching spikes from an individual RGC identified separately during visual stimulation (see materials and methods). The eliciting of these easily identified spikes by electrical stimulation will be referred to as somatic activation, although the actual site of activation may well be the axon initial segment (Fried et al. 2009; Sekirnjak et al. 2008) or elsewhere. In each preparation, somatic activation thresholds for RGCs were in approximately the same current range as bundle activation thresholds (Fig. 6). Given the similar activation thresholds for axon bundles and RGC somas, the fundamental problem for development of high-resolution epiretinal prostheses is whether individual RGCs can be activated at their somas without activating bundles. This was tested by evaluating the fraction of electrodes on which the threshold for activation of an individual RGC was lower than that for bundle activation. On average across 3 preparations at temporal eccentricities of 48.2°, 58.1°, and 58.1°, 97% of electrodes were able to elicit some electrical activity, as revealed by bundle activation (see Fig. 4). Analysis was restricted to the 243 (47%), 150 (29%), and 199 (39%) electrodes on which somatic activation of a single RGC was detected on the stimulating electrode and its neighbors (see above). Of these electrodes, 115 (47%), 73 (49%), and 79 (40%) were able to activate a single RGC at threshold using a stimulation current level at or below bundle activation threshold for that electrode. Thus, in aggregate, 45% of all electrodes that were able to activate a single RGC (i.e., 17% of all electrodes on the array) were able to do so without activating bundles.

The above comparison of thresholds provides only a partial view of how effectively RGCs can be activated without activating bundles, because thresholds have different meanings in the two cases. For RGCs, the threshold indicates the current at which a spike is elicited in 50% of trials, and the dependence on current is sigmoidal and gradual over the current levels tested (Jepson et al. 2013). For bundles, the threshold indicates the estimated highest current level at which the probability of activation is zero, and the dependence on current is nearly steplike above threshold over the current levels tested (Fig. 3C). Thus no single number can convey the ability to activate RGCs without bundles. As an example, consider the case of attempting to stimulate RGCs with high probability, rather than at threshold. Of the electrodes examined above, 35 (14%), 26 (13%), and 23 (15%) were able to activate a single RGC with probability 0.95 with the use of a stimulation current level at or below bundle activation threshold, substantially fewer than the number of RGCs that could be activated with probability 0.5. Given the reduced number of cells that can be activated, it is not obvious a priori what RGC activation probability level is most appropriate to target. This issue is more fruitfully evaluated within the context of the neural code of the retina (see discussion).

A potentially important factor in targeting RGCs without bundles is spatial resolution. If RGCs and bundles are highly sensitive to the location of the stimulating electrode, then more precise control over electrode position could provide more selective RGC targeting. To probe this possibility, activation was examined on the basis of the estimated distance from the electrode to the site of RGC activation. For each RGC, the electrical image (see Fig. 2) was examined to identify the two electrodes with the greatest recorded somatic spike amplitude. The center of mass of these two electrode locations (locations weighted by signal amplitude) was computed and was interpreted to reflect the position of the axon inner segment, where sodium channel density is highest and where electrical activation appears to occur in these conditions (Sekirnjak et al. 2008). The number of electrodes for which bundle threshold was larger than or smaller than RGC activation threshold was then examined as a function of distance between the estimated activation site and the electrode nearest that activation site. The results (Fig. 7) reveal that the electrodes for which this distance was lower were more likely to have a lower RGC threshold than bundle activation threshold (P < 1e-4, 0.0543, and 0.0004; observed difference in distribution means compared with differences in means obtained by resampling from the pooled population; 1-tailed). Thus the proximity of the electrode to the activation site is correlated with the ability to activate RGCs below bundle threshold. Although not definitive, this analysis suggests that spatial resolution may be a limiting factor and that denser electrode arrays could produce more selective activation of RGCs without bundles.

The impact of these results on the ability of a prosthesis to reproduce the neural code of the retina was further examined by analyzing the proportion of RGCs of each type that were activated selectively at or below bundle threshold. As an example, consider the activation of ON parasol RGCs from one preparation (retina 1, Fig. 8; temporal eccentricity 48.2°). In this preparation, light responses were obtained from 116 ON parasol cells. Based on the mosaic structure of RGC receptive fields, this represented 97% of the total number of ON parasol cells presumably overlying the electrode array, assuming a healthy retina. Among these cells, 78 (67%) were activated selectively, i.e., without other detectable somatic activation. The receptive fields of the selectively activated cells, which represent their normal encoding of visual space, densely sampled the area covered by the electrode array (Fig. 8A, blue). Among these cells, 37 (47%) exhibited somatic activation thresholds less than or equal to bundle activation threshold on one or more electrodes that stimulated it most effectively. Correspondingly, the receptive fields of these cells sampled the area of retina over the electrode array more sparsely (Fig. 8B, blue). The fraction of recorded cells that could be activated without activating bundles varied widely across RGC types and preparations (Table 1; see discussion). In summary, the collection of RGCs that could be activated selectively with stimulation currents below bundle threshold formed a patchy representation of the visual scene.

Central recording.

The applicability of these results for a clinical prosthesis in the central retina was evaluated by recording and stimulating in the raphe region, which occupies 0°–20° eccentricity on the temporal horizontal meridian. The raphe is avoided by peripheral axon bundles that travel in an arc around the fovea as they head toward the optic disk, and thus exhibits low axon bundle density compared with other central locations (Fig. 9A). Following the analysis above, in one raphe recording (7.3° eccentricity), 421 (98%) of 430 analyzable electrodes produced some degree of electrical activation, as assessed by bundle activity. Selective somatic activation of individual RGCs was observed with 93 (22%) of the electrodes, compared with 39% of electrodes that could achieve selective somatic activation in the periphery. Individual RGC activation was more rare in the raphe than in the periphery, presumably because RGCs are more closely spaced in the raphe. Among the electrodes that produced selective somatic activation, 70 (75%) did so at or below bundle threshold (Fig. 9C), compared with 45% of electrodes that did so in the peripheral recordings (Fig. 6B). Thus, in the raphe, 75% of all electrodes that were able to activate a single RGC (i.e., 16% of all electrodes on the array) were able to do so without bundle activation. Separate analysis was not performed on different RGC types (e.g., Table 1), because the light response data collected were insufficient for reliable cell type classification.

Application.

To apply these findings in a clinical prosthesis and accurately reproduce the neural code in a subset of RGCs would require a system architecture quite unlike that of present-day devices. First, it would be necessary to calibrate the stimulation current levels of the device to target the subset of RGCs that can be selectively activated. This would presumably involve incorporating circuitry into the device to record the activity elicited by electrical stimulation, processing the recordings in a manner similar to the analysis presented here, and then adjusting stimulation current levels to the optimal values for selective RGC activation. Second, reproducing the neural code of the retina would require identification of distinct RGC types recorded and the use of models of visual response to determine the appropriate firing pattern of each RGC on the basis of its type and location with respect to the input image (Pillow et al. 2008). This would require developing methods to identify RGC types using only intrinsic properties of cells (Richard et al. 2015), because light responses would not be available for this purpose in the blind patient. Third, spatially patterned stimulation (Jepson et al. 2014a) or optimized stimulus waveforms could potentially be used to optimize the selectivity of stimulation. Finally, this precisely calibrated stimulation approach would likely need to be adjusted periodically over time, as the implant shifts relative to the tissue or if an immune response occurs near the implant (Polikov et al. 2005). This vision of a future device raises many technical issues, including close apposition of the implant to prevent current spread, stability over time, and size and power dissipation constraints associated with high-bandwidth electrical recording and mimicking of natural retinal responses. Although no neural interfaces with such capacities have yet been developed, the relatively accessible retina and its well-understood function may provide the ideal setting for a first attempt to interface to nervous system circuitry at cellular resolution.

Neural code.

Comparison of bundle and somatic activation thresholds (Fig. 6) was used to assess the visual field coverage by several RGC types that can be reproduced by electrical stimulation (Fig. 8). However, this simple assessment fails to address several important factors relevant for reproducing the normal visual signal. First, although somatic threshold is often lower than bundle threshold, in many cases the difference is small, implying that precise calibration of the stimulation current would be required for selective activation of somas (see Application above). Second, in many cases activation of a RGC would occur with probability <1 if the highest current level that does not activate bundles were used, because somatic threshold is defined as the stimulation current producing activation probability of 0.5, and the increase of RGC activation probability with current is relatively shallow (Jepson et al. 2013). For example, in a preparation in which 47% of electrodes produced RGC activation probability >0.5 at bundle threshold, only 14% of electrodes produced RGC activation with probability >0.95 (see results). It is possible that this limitation could be addressed by stimulating multiple times in rapid succession to ensure the production of a spike in the RGC. Specifically, the threshold definitions for RGC somas and bundles differ such that repetitive stimulation would be expected to increase the probability of RGC activation without eliciting bundle activation. This approach could also improve activation for cases in which bundle threshold is lower than somatic threshold (i.e., RGC firing probability in response to stimuli at bundle threshold is <0.5). However, this strategy has not been tested experimentally. Third, and more generally, it remains unclear how faithfully a partial representation in a subset of cells of each type could convey a visual image to the brain, because it is unknown how the different RGC types contribute to different aspects of visual perception and behavior. For example, even if the collection of available cells is patchy (Fig. 8), it is possible that patients could perceive the visual scene more completely by exploiting spatial correlations in natural images, eye movements, and multiple cell types. Given these factors, a full understanding of the consequences of the partial visual signal will require substantial experimental and computational investigation.

Central retina.

In much of the central retina, the thick layer of axons overlying RGC somas would almost certainly make selective activation of RGCs without activating axon bundles more difficult than in the peripheral recordings presented (Fig. 6–8). Given these factors, the raphe region of the central retina (Fig. 9), with its low axon density, may represent the ideal target location for a high-resolution epiretinal prosthesis. The results of recording in the raphe support this possibility (Fig. 9), demonstrating that the fraction of electrodes that could produce selectively activated RGCs was comparable to or higher than the fraction observed in the peripheral retina. Recent emphasis with clinical devices has been on targeting the fovea, rather than the raphe or other more peripheral areas, because of superior clinical outcomes (Ahuja et al. 2011; Humayun et al. 2012; Stingl et al. 2013a). However, the comparisons in the clinic have been performed using devices that activate RGCs indiscriminately and simultaneously over large areas. This kind of activation produces a visual signal that is never observed in nature: in the healthy retina, RGCs of 20 types signal information in their precise and distinct spatiotemporal firing patterns (Field et al. 2007). The present results suggest an entirely different approach to epiretinal electrical stimulation: reproducing the neural code cell by cell and spike by spike, in a subset of RGCs, in a central but perhaps not foveal region of the retina. Thus the unanswered clinical question is, would a partial but precise artificial visual image in the raphe region provide clinical outcomes superior or inferior to unnatural, indiscriminate stimulation of RGCs in the fovea? Answering this question will probably require development of the technology and testing in vivo, although computational modeling currently underway may provide some clues.

Imaging.

A possible alternative approach to avoiding axon bundle activation would be imaging of bundles before device implantation, using optical coherence tomography (Kocaoglu et al. 2011; Shemonski et al. 2015). This could potentially yield a simpler or faster way to identify bundle locations and the electrodes that would most likely activate them. In the present data, images of bundles (e.g., Figs. 2A, 8C, 8F, 8I) obtained with immunohistochemistry after recording were compared with bundle activation thresholds (e.g., Fig. 4). The correlations between image intensity summed over a 15-μm radius of each electrode and bundle threshold at that electrode were weak (0.34) but statistically significant for one preparation with low bundle density (e.g., Fig. 2A) and near zero for preparations with higher bundle density (e.g., Fig. 8C). Thus, although bundle position estimated from imaging seems to influence bundle activation thresholds in some cases, other factors such as thickness of the inner limiting membrane or proximity of the tissue to the electrode array are probably also important. Given that bundle activation thresholds and somatic activation thresholds are similar, these observations suggest that bundle avoidance will be much more effective with electrical calibration of the kind performed in the present study than with imaging.

Spatial resolution.

The present results were obtained using stimulation through individual electrodes on the multielectrode array. However, spatial manipulations could improve the selective recruitment of cells over axons. In fact, multielectrode patterned stimulation has been shown to allow selective activation of a target cell without activating nearby cells, when single-electrode stimulation failed (Jepson et al. 2014a), using the same experimental methods. In addition, the 60-μm pitch of the electrode array limited the resolution with which stimulation locations could be probed. It is possible that higher selectivity could be obtained with denser arrays (Radivojevic et al. 2016). These considerations suggest that the present results represent a conservative estimate of how effectively single RGCs can be activated without evoking bundle activity.

Temporal modulation.

Previous results (Boinagrov et al. 2014; Freeman et al. 2010; Weitz et al. 2013, 2015) have shown that changes in the waveform of electrical stimulation can target bipolar cells over RGC axons, in a manner that substantially affects the recruitment of axon activity, due to the different integration times of different cells and compartments. Although the long pulses that elicit network-mediated responses are not appropriate for fine-grained reproduction of the temporal structure of retinal spike trains, the general concept of improved temporal stimulation patterns remains open to further exploration and again suggests that the current findings represent a conservative estimate of how selectively activity can be evoked in RGCs.

Degeneration.

The effect of retinal degeneration on the application of these methods is unknown. In animal studies, there have been varied results, showing either that thresholds for direct activation of RGCs increase with degeneration (Cho et al. 2016) or that thresholds stay the same throughout degeneration (Sekirnjak et al. 2009). It is unclear how these changes would differentially influence various cell types, or axonal vs somatic activation, in a human patient. Other changes that occur in the degenerate retina, such as changes in the morphology and density of RGCs, may also have an effect on somatic and bundle activation thresholds (Chan et al. 2011; Cho et al. 2011). Changes in the RGC spontaneous firing patterns (Margolis et al. 2008; Sekirnjak et al. 2011; Ye and Goo 2007) may affect the application of the present algorithm for detecting bundle activation, because it could increase the overall electrical activity in the recording, effectively adding more noise.

Caveats.

Ultimately, understanding the full artificial signal created by electrical stimulation, including the complete set of cells that can be activated without producing unwanted bundle activity, would require recording reliably from every cell and axon in the vicinity of the electrode array. This goal represents a substantial challenge for current laboratory recording technology, let alone implantable hardware. Calcium imaging approaches increasingly provide an attractive technology option (Weitz et al. 2015) in the laboratory, although at present they are limited to detecting bursts of spikes with low temporal precision. The techniques used in the present study probably represent the closest approximation to recording complete spike trains from complete neural populations in the mammalian nervous system (Field et al. 2010; Frechette et al. 2004; Segev et al. 2004) and in principle could be performed using the electrodes of a prosthesis. However, significant uncertainties remain. Based on the known mosaic organization of the major RGC types in primate retina (Chichilnisky and Kalmar 2002; Dacey 2004; Frechette et al. 2005), a substantial fraction of cells overlying the electrode array were not recorded, to a degree that depended strongly on cell type (Table 1). It is possible that activation of these cells or their axons was not detected by the present methods. Indeed, previous work using the same methods but smaller electrode arrays likely underestimated the influence of axons by analyzing only somatic spikes without access to propagating signals over a large array (Jepson et al. 2014b). In the present work, control analysis with synthetic data (see appendix) revealed that activation of single axons was usually detectable (this would likely not be the case with calcium imaging; Weitz et al. 2015). However, the axon signals used for synthetic data were obtained from the cells that were identified during recordings with visual stimulation; it is possible that axons with smaller recorded signals were activated by electrical stimulation. Furthermore, it is possible that selective somatic activation of a single RGC would be overestimated by the present methods, if other RGCs produced somatic spikes too small to detect. On the other hand, it is possible that RGCs with small spikes were selectively activated at current levels below bundle threshold but could not be identified with confidence using the present methods, particularly because of the large electrical stimulation artifact. It is difficult to assess the collective impact of these factors, which could bias the estimated selectivity upward or downward. In principle, if the factors that limit the efficiency of recording similarly limit the efficiency of activation (e.g., electrode and local tissue properties), then the present results on the fraction of electrodes and cells for which selective activation is possible may represent a reasonable approximation to the in vivo situation relevant for a clinical device.