Optogenetic stimulation of the cochlea—A review of mechanisms, measurements, and first models

ABSTRACT

This review evaluates the potential of optogenetic methods for the stimulation of the auditory nerve and assesses the feasability of optogenetic cochlear implants (CIs). It provides an overview of all critical steps like opsin targeting strategies, how opsins work, how their function can be modeled and included in neuronal models and the properties of light sources available for optical stimulation. From these foundations, quantitative estimates for the number of independent stimulation channels and the temporal precision of optogenetic stimulation of the auditory nerve are derived and compared with state-of-the-art electrical CIs. We conclude that optogenetic CIs have the potential to increase the number of independent stimulation channels by up to one order of magnitude to about 100, but only if light sources are able to deliver confined illumination patterns independently and parallelly. Already now, opsin variants like ChETA and Chronos enable driving of the auditory nerve up to rates of 200 spikes/s, close to the physiological value of their maximum sustained firing rate. Apart from requiring 10 times more energy than electrical stimulation, optical CIs still face major hurdles concerning the safety of gene transfection and optrode array implantation, for example, before becoming an option to replace electrical CIs.

KEYWORDS:

• Auditory system
• cochlear implants
• optogenetics
• spiking neurons

1. Introduction

The development of optogenetic stimulation methods (Nagel et al. 2005; Bamann et al. 2010; Deisseroth et al. 2006) was a major scientific breakthrough that has paved the way to stimulate large populations of neurons with high spatial resolution (Schoenenberger et al. 2008). Research groups now make this technology available for neuroprosthetic applications which suggests itself for retinal implants (Lagali et al. 2008; Ivanova et al. 2010). For cochlear implants (CIs), optical stimulation promises major advantages, in particular a substantial increase in the number of independent stimulation channels (Hernandez et al. 2014a; Jeschke and Moser 2015). This is very attractive for CIs, as present electrical stimulation with an intracochlear electrode array is severely hampered by the wide spread of electrical excitation which limits the number of independent stimulation channels to less than eight (Friesen et al. 2001). This low spectral resolution already provides surprisingly high speech recognition rates (Shannon et al., 1995), even in noise (Zirn et al. 2016). However, especially for music perception, this low spectral resolution is clearly far from sufficient to resolve pitch and timbre (McDermott 2004) as in an intact inner ear. In addition, the wide electrical crosstalk between electrodes limits their independent usage such that present stimulation strategies activate only one electrode after the other. Contemporary coding strategies stimulate at rates of 180 to 320 pps/channel (spectral peak coding), 800 to 1500 pps/channel in the case of continuous interleaved sampling strategy (CIS and CIS+), and for advanced combination encoders from 250 up to 3500 pps/channel (Arora 2012).

The root cause of the broad current spread in the cochlea lies in the fact that the stimulation electrodes are inserted into scala tympani which is filled with a high conducting ionic solution called perilymph. The neurons of the auditory nerve (the spiral ganglion neurons: SGNs) are hidden behind the thin shell formed by the osseous spiral lamina. SGNs are relatively far away from the electrode and therefore similar neuron populations are activated even when different electrode contacts are addressed (Frijns et al. 1995). The fundamental difference between electrical and optical stimulation is that electricity spreads uniformly into all directions in a conducting fluid, whereas light can be directed and even focussed. This is illustrated in Figure 1. Therefore, optical stimulation has the potential to target well-defined SGN populations which would allow a much higher number of independent stimulation channels (Hernandez et al. 2014a). The expectation is that this jump in frequency resolution improves the sound quality of CIs enormously.

Figure 1. Differences between electrical and optical stimulation in the inner ear. An optical implant can stimulate more locations in the cochlea, as optical spread can be confined by the nature of the light source or with additional focussing elements. Reprinted from Moser (2015) with permission from Elsevier. © Elsevier. Reproduced by permission of Elsevier. Permission to reuse must be obtained from the rightsholder.

During the last years, a number of optical mechanisms have been developed and explored that can be used to elicit an action potential in neurons. Callaway and Yuste (2002) describe the different methods that have evolved over time, starting with the invention of lasers and their use in neuronal research (Fork 1971). Direct stimulation of neurons with high-power infrared laser pulses has been described by Izzo et al. (2006). Although mechanisms like changes in the cell membrane capacitance caused by the temperature increase (Shapiro et al. 2012) or activation of TRPV4 heat-sensitive ion channels (Albert et al. 2012) have been identified, cochlear responses to lower energy infrared light pulses seem to rely rather on the optoacoustic effect that introduces pressure waves in the cochlea (Teudt et al. 2011). This excitation of the auditory nerve is probably caused by intact hair cells which survived in the cochlea (Thompson et al. 2015).

Optogenetics, on the other hand, relies on light-sensitive proteins, called opsins, which mediate the conversion of a photon into an electrochemical signal. They have to be expressed through genetic targeting strategies directly in the auditory nerve, or alternatively in neurons located in higher brain areas. With selective targeting strategies, specific neurons and regions could be reached (Guo et al. 2015). Type I opsins (microbial opsins), in contrast to type II opsins (animal opsins), can do light sensing and ion conduction in a single-component system. Animal rhodopsins are also referred to as a specialized subset of G-protein-coupled receptors that can detect extracellular signals and were already discovered in the 1870s (Kühne and Foster 1878). They catalyze the GDP/GTP exchange which then mediates the opening of ion channels. This two-step process provides amplification and therefore high sensitivity, but at the cost of speed. So far, three classes can be distinguished for type I opsin genes (Yizhar et al. 2011): Bacteriorhodopsin, a light-gated proton pump; Halorhodopsin, a light-activated chloride pump; and Channelrhodopsin (ChR), a light-activated ion channel. Channelrhodopsin originates from green algae, where its depolarization is responsible for phototaxis (Ernst et al. 2014), for example. Channelrhodopsin-1 (ChR1) is selective to protons rather than other cations. The low number of protons that pass through at physiological pH values do not depolarize the neuron below firing threshold (Nagel et al. 2002, 2003). For Channelrhodopsin-2 (ChR2), the effective pore size seems to exceed that of the voltage-activated Na⁺ channel (Nagel et al. 2003). Its photocurrent shows an inverse relationship with the atomic radius of the cation. In addition to monovalent cations, ChR2 is also permeable for divalent cations, most notably for Ca²⁺ (Nagel et al. 2003). As the depolarizing currents are high enough to selectively excite neurons which have been genetically programmed to express this opsin, ChR2 has become a widely used tool in neuroscience. The discovery of channelrhodopsins led to a new field in neuroscience that allows scientists to research specific brain functions in in-vitro cultures but also in behavioral studies in animals (Tye and Deisseroth 2012; Nagel et al. 2005; Liewald et al. 2008; Zhang et al. 2006; Cardin et al. 2010; Hernandez et al. 2014a).

ChR2 and its variants have an organic, nonprotein component (an aldehyde derivative of vitamin A) that absorbs photons and is covalently or noncovalently bound to the protein. The larger the chromophore, the greater is the possible extent of electron delocalization.

This paper evaluates the potential that optogenetics holds to develop a novel generation of optogenetic CIs. From a technical point of view, their biggest possible advantage is a substantial increase in independent stimulation channels. A high number of channels can be realized only if the light sources are spatially selective. Therefore, one major part in this review analyzes the beam shape and scattering. The most limiting disadvantage of optogenetic stimulation is that opsins are relatively slow compared to electrical stimulation which will compromise the temporal precision of auditory nerve responses. To assess and balance these two effects, one has to consider four parts which we review in the following sections: (1) a model for the opsin and its kinetics, (2) a model for the transfection of neurons with opsins, (3) an excitable neuron model, and finally (4) a model of beam shape of the light source, light absorption, and scattering. This work should pave the way to develop and connect these models and it already provides some quantitative assessments of the advantages a light-based stimulation device might have over an electrical implant.

2. Opsin targeting strategies

Rendering SGNs light sensitive requires genetic manipulation in vivo. The genome of the opsins are usually transferred to the target neurons with viral transfection (see Table 1). Hernandez et al. (2014b) were able to provide a first proof of principle, expressing ChR2 in SGNs in transgenic mice and rats and also with transuterine injection of an adeno-associated virus into the embryotic oocyst. Meng et al. (2014) managed to render SGNs light-sensitive by injection of AAV2/1CgR2-mCherry into the cochlea of neonatal mice. Although Hight et al. (2015) were able to express ChR2 and Chronos with AAV2.8 virus transfection in the dorsal cochlear nucleus of adult mice, a demonstration for adult auditory nerve transfection is still missing.

Table 1. Transfection vectors and promoters adapted from Yizhar et al. (2011) with examples of applications in the auditory pathway.

Lentivirus
Promoter	Organism	Cell-type	Reference
hGFAP	rat, mouse	astrocytes	Gradinaru et al. [2009]
CaMKII	rat, mouse	cortex & hippocampus	Aravanis et al. [2007]
YFP	hESC	hESC-CM	Abilez et al. [2011]
Adeno-associated virus
Promoter	Organism	Cell-type	Reference
hGFAP	rat	astrocyte	Lee et al. [2010]
			Lawlor et al. [2009]
MBP (AAV8)	rat	oligodendrocytes	Lawlor et al. [2009]
HSYN	mouse	SGNs	Hernandez et al. [2014b]
CAG	mouse	DCN	Hight et al. [2015]
(AAV2.8-ChR2)

The technology of transfection is evolving rapidly, so it is probably only a question of time until different opsins can be delivered into the auditory nerve. For energy efficient CIs, it will be essential that the concentration of the opsins is dense enough to assure high light sensitivity. A completely uniform expression of the opsins across nerve fibers is not required. On the contrary, variation in opsin expression has the potential to mimic nerve fibers with different thresholds, similar to the in vivo condition (Yates, 1990). For electric stimulation, the dynamic range is only 10–20 dB (Zeng et al. 2002) into which the large acoustic range has to be compressed.

In SGNs, the cell membrane with its nonlinear ion channels integrates electrical pulses. This causes complex subthreshold dynamics and adaptation after firing, such that their excitability depends on the stimulus and firing history. Therefore, the dynamic range compression is far from trivial, at least for electric stimulation. The analysis of the dynamics of the opsins together with the dynamics of SGNs due to their ion channel composition gives rise to the assumption that the firing characteristics will also exhibit complex behavior for optical stimulation. A larger spread of firing thresholds across SGNs could decouple these dynamics and increase the overall dynamic range despite the limited dynamic range of individual neurons. This could lead to more reliable coding of signals with large fluctuations like speech.

In summary, variations in expression density can be beneficial, as they increase the dynamic range of the neuron population to light stimulation. The drawback is that higher light intensities are required to stimulate the less sensitive neurons.

3. Opsins

The structural changes of channelrhodopsin when exposed to light were revealed with Fourier Transform Infrared (FTIR) spectroscopy (Kuhne et al. 2015) and by analysis of the crystal structure (Müller et al. 2011, Kato et al. 2012). All microbial rhodopsins show a bundle of seven transmembrane helices around the chromophore and the retinal binding structure is the most conserved element (Ernst et al., 2014). Channelrhodopsin mutants can therefore have different properties and they can even be optimized by genetic engineering.

To use opsins in any kind of model, their photocycle characteristics need to be known. For ChR2, these kinetics have been researched thoroughly by Hegemann et al. (2005) and Nagel et al. (2003) (see also Table 2). They demonstrated that the photocycle can be best approximated with a four-state model (see Figure 2).

Table 2. Time constants for different types of channelrhosopsins, adapted from Kleinlogel et al. (2011).

Opsin	(ms)	(ms)	Temperature	Reference
ChR2	0.2 ± 0.002	10 ± 1	8°C, 11.5°C, 14°C, 16°C, 22 - 24°C	Kleinlogel et al. [2011]
ChR2 H134R	0.6	19 ± 2	8°C, 11°C.5°C, 14°C, 16°C, room temp.	Kleinlogel et al. [2011]
CatCh	0.6 ± 0.003	16 ± 3	23°C	Kleinlogel et al. [2011]
ChETA	–	4.8 ± 0.6	32.5 ± 1°C	Kleinlogel et al. [2011]
ChIEF	–	9.8 ± 0.7	room temp.	Kleinlogel et al. [2011]
Chronos	2.3 ± 0.3	3.6 ± 0.2	22°C	Klapoetke et al. [2014]

Figure 2. The four state model for the light cycle of channelrhodopsin (Stefanescu et al., 2013) has two open states (o₁ and o₂) and two closed or dark states (c₁ and c₂). Light activations (light energy: ) are depicted as blue arrows cause transitions ( and ) to the open states. and indicate transition rates to the closed states and from c₁ to c₂. e₁₂ and e₂₁ are the transition rates between the open states. Modified from Stefanescu et al., (2013) with permission from Springer. © Springer. Reproduced by permission of Springer. Permission to reuse must be obtained from the rightsholder.

For a model of ChR2, the size of the chromophore defines the interface to the light source. After the absorption of a photon, the chromophore changes its conformation and opens the opsin’s ion channel (Hegemann et al., 2005; Müller et al., 2011). Therefore, the chromophore constitutes the effective surface for light activation. Hegemann et al. (2005) and Foutz et al. (2012) estimate the cross section for absorption in the order of and Grossman et al. (2011) approximate its size as . Considering the total area for light interaction, one has to consider the total area of the target, its channel density, and ultimately—for whole cell models—its 3D shape.

First models for channelrhodopsin photocycles have been proposed by Stehfest and Hegemann (2010), according to electrical measurements, as well as by Nikolic et al. (2009) and Grossman et al. (2011). Faster variants of ChR2 have also been implemented and compared to experimental data by Stefanescu et al. (2013). A computational model for ChR2 with four states has been introduced by Foutz et al. (2012), Arlow et al. (2013) for myelinated axons and by Williams et al. (2013) for cardiac models. They all share two open states (o₁ and o₂) and two closed (sometimes also called dark) states (c₁ and c₂) that all interact through transition rates. e₁₂ and e₂₁ are the transition rates between the open states, between the closed states, and , are the transition rates from the open state to the corresponding closed state (see Figure 2). and describe the light-induced channel activation from the closed states c₁ and c₂ to o₁ and o₂, respectively.

The four-state model is governed by the following differential equations (adapted from Grossman et al. 2011):

Parameters of such models have been described by Stefanescu et al. (2013), Grossman et al. (2011), and Foutz et al. (2012). Figure 3 shows the open and closed states as a function of time for a 3ms light pulse with increasing irradiation. Only for the highest irradiation, the deactivation of the ChR2 appears, the probability for the state o₁ decreases while the light is still on. The probability for state o₂ increases, but as its conductance is lower, the photo current desensitizes for prolonged stimuli (Mattis et al., 2012).

Figure 3. Rise and decay of the optically induced ChR2 activation. The light source was on between 2 and 5ms (blue shaded area). The lines show the probabilities of the two open (green) and closed (red) states at three different light intensities which is indicated by the colour saturation.

As ChR1 shows less inactivation than ChR2, Lin et al. (2009) have engineered ChEF variants by crossing over the sequence from ChR1 to ChR2 to generate light-sensitive opsins with reduced inactivation (33% compared to 77% in ChR2) which increases channel efficacy and therefore light sensitivity. By introducing an additional point mutation, they also created the ChIEF variant, which exhibits a shorter time constant for channel closure. This modification enables faster stimulation and spiking patterns compared to ChR2 (Lin et al. 2009), because in the photocycle the time constants for channel closure are longer than for channel opening.

A more recently developed fast opsin, Chronos, has the fastest kinetics so far (Klapoetke et al. 2014). Hight et al. (2015) expressed Chronos in auditory brainstem neurons and recorded in a higher neuronal stage, the inferior colliculus. They found, compared with ChR2, driven rates were much higher for stimulation rates from 25 to 200 pps and also with significantly higher synchronization from 56 to 224 pps.

4. Modeling neuronal responses in transfected neurons

We can now introduce the light-sensitive ion channel in computational models of SGNs. This approach is straight-forward in Hodgkin–Huxley type models (Hodgkin and Huxley 1952; Abilez et al. 2011; Rattay et al. 2001), whereas the integration of light-induced currents is less clear in phenomenological models (Bruce et al. 1999a, b). Depending on the research question, models with the appropriate complexity can be chosen, like point neuron models (Nicoletti et al. 2013), multicompartment models (McNeal 1976), or full three-dimensional models.

Here we combined an auditory nerve model proposed by Nicoletti et al. (2013) and Rudnicki et al. (2015) and gained responsiveness to light by including a modeled ChR2-channel. Dynamic simulations were conducted with BRIAN (http://www.briansimulator.org) (Goodman and Brette 2009). We included parameters such as the expression level of ChR-variants with a channel-density parameter and the size of the chromophore. The optical activation of ChR2-channels depolarizes the neuron and action potentials are generated with a delay which depends on the expression density of the ChR2-channel and light intensity (see Figure 4). This delay, together with channel-noise, limits the temporal precision of optical stimulation.

Figure 4. Optical stimulation (1 ms light pulse, 4.5 mW/mm², blue bar) of a modeled neuron with ChR2 (expression densities: 20–120μm⁻²). Neurons with higher ChR2 expression depolarize faster and exhibit a shorter first-spike latency.

Whereas the activation of ChR2 is fast (, Nikolic et al. 2009), its slow closing dynamics (also visible in Figure 4) limit the maximal spike rate, which can be driven by increasing the frequency of light pulses. Here, faster opsin variants like Chronos promise performance improvements.

5. Light sources for optical stimulation

The main advantage of optical stimulation is the possibility to achieve a much sharper place-specific activation of SGNs compared to electrical stimulation. However, whether a sharply confined illumination at the location of the neurons can be achieved depends on the whole light path (see Figure 5). We have to analyze the light path in detail to achieve quantitative estimates of the spread of excitation and the required power for the light sources.

Figure 5. The general layout of recent illumination models. The light source properties describe the beam that exits the source. Attenuation accounts for scattering, absorption, and reflection in the space between the light source and the receiving tissue. Ultimately, the photon flux at the receiving tissue is calculated which drives the activation of the light-sensitive channels.

Therefore, the main goal of optical stimulation models is to predict the photon flux a light source creates at the position of the SGNs. While electromagnetic theory can describe the propagation of light, its direct application is too complex to describe absorption and scattering in tissue. For this task, usually a simpler approach is used in the form of radiation transport theory. This theory includes two analytical approaches, the diffusion approximation and the Kubelka–Munk model for absorption and scattering (Vo-Dinh 2003).

5.1. Properties of light sources

Light sources can be, depending on the type of experiment, glass fibers, micro-LEDs, laser diodes (especially vertical cavity surface emitting lasers: VCSELs), or spatial light modulators. Spatial light modulators are used predominatly in brain research to pattern light onto groups of neurons (Nikolenko et al. 2008, Pashaie et al. 2015). They allow the most flexible way to excite neurons by controlling light intensities with high spatial resolution which depends only on the optical properties of the system. However, such devices are too bulky for implants. Other methods such as wave guides connected to light sources; micro-LEDs or VCSELs are also used to illuminate tissue in fixed areas of interest. Modeling these light sources has to account for their specific properties like their diameter and radiation profile which is often described by their numerical aperture.

Two problems which are relevant for larger distances between light source and neuron (distance > light emitting area) arise from large beam divergence: 1) loss in spatial selectivity and 2) loss of radiation per area. Optical fibers have medium numerical apertures but large losses can occur when a light source is coupled into a fiber. LEDs are very efficient light sources but they have very high numerical apertures of close to one and therefore a wide beam divergence. Despite this wide divergence, a first attempt for the fabrication of an optical cochlea implant consists of a GaN-based micro-LED array on a flexible substrate, achieving a radiant emittance of 6mW/mm² with an efficiency of about 17% (Goßler et al. 2014). As recent commercially available blue LEDs have a much better efficiency (e.g., 63% for the TR260 from Cree), improvements for micro-LED arrays on flexible substrates can be expected for the near future. To reduce the divergence, micro-LEDs with microlenses can be used. To achieve a narrow focus area at larger distances, adjustable microlenses are available, as described by Buchegger et al. (2009), for example. In order to circumvent the need for focussing elements, it could be beneficial to use VCSELs instead of LEDs. Their low diverging output beams would enhance the spatial selectivity and efficiency by reducing misdirected light power. Lasers in general have the advantage that they can be coupled efficiently to fibers (Lu et al. 2009). Such lasers have been created with wavelengths down to 462 nm, efficiencies of few percent at 0.7 mW maximum output power, and 8 μm aperture diameter (Kasahara et al. 2011). They still would have to be adopted for optogenetic stimulation, as their output power would be rather too high.

Technology is progressing rapidly in this field and many interesting novel devices appear, such as multi-LED arrays with attached polymer waveguide (Kwon et al. 2013), spatially addressable optical fibers (Pisanello et al. 2014), or high-density silicon probes (Scharf et al. 2016). A review of all light sources would go beyond the scope of this paper.

5.2. Modeling light sources

Starting with the light source irradiance , the irradiance at a certain point in space is defined by in a cylindrical coordinate system (Foutz et al. 2012) (z: distance from light source, r: distance from optical axis), where is a combination of a Gaussian beam model emitted by a light source, conical spread of unfocused light, and a Kubelka–Munk model for scattering and absorption.

In a first approximation, the Gaussian beam form is thought to expand with the conical spread. If we consider, for example, the case of a passive implant, such as an optical fiber connected to a light source, the light reaching the transfected neurons is estimated by the following equations using a cylindrical coordinate system (Vo-Dinh 2003).

5.2.1. Geometrical spread

The optical fiber emits a beam with the radius at its end. Due to the geometrical spread (divergence angle: ), the radius of the light cone increases with distance z from the end of the fiber:

(1)

Due to the conservation of energy, the following relation between radiant power P and the irradiance can be calculated:

(2)

Therefore the intensity I decreases due to geometrical spread as a function of z:

(3)

which leads to the transmittance due to the geometrical spread with:

(4)

5.2.2. Gaussian beam model

The beam profile follows a Gaussian shape, which expands with the geometric spread (Yizhar et al. 2011; Huber et al. 2008). It can be approximated as a transmittance function :

(5)

However, in addition to the characteristics of the light source, the light transmission through tissue, especially bone, has a major influence on the light flux at the target. Tissue properties have large effects on absorption, scattering and refraction. Here, modeling approaches are less straight-forward.

5.3. Light absorption and scattering

5.3.1. Scattering and absorption

The Kubelka–Munk equations provide a simplification of scattering and absorption. They are guided by the coefficients S for scattering and K for absorption:

where i indicates the intensity of light that is transmitted in the sample, j the intensity of light propagating in the backscattered direction, and z the distance from the nonilluminated side of the sample. The sample is assumed to be planar, homogeneous, and an ideal diffuser that is illuminated from one side with diffuse monochromatic light Vo-Dinh (2003).

As shown by Kubelka and Munk (1931) and Vo-Dinh (2003), scattering and absorption in diffuse scattering media can be described as another transmittance M:

with:

where K is the absorption coefficient and S the scatter coefficient both in mm⁻¹.

The light flux at a point (r, z) is then:

(6)

Together with the single photon energy and the cross section of the light-activated area in the channelrhodopsin, a flux across the single chromophore can be computed into a flux that is required for the activation of channelrhodopsin (compare Figure 3).

5.3.2. Monte Carlo Simulation

Instead of these analytical equations, light scattering can also be simulated with the Monte Carlo method, where single photons are traced on their way through the tissue. In this method, probability distributions are replaced with random numbers to simulate absorption and angular deflection during scattering for many photons. An infinite amount of photons would approximate the exact solution of the radiation transport equations even for three-dimensional scattering and multiple layers of tissue (Wang et al. 2008).

5.4. Comparison with measurements

Models and measurements coincide (Flock et al. 1989), although stronger beam widening occurs in highly scattering media (brain). In the cochlea, scattering and absorption can be estimated with a combined model for light spread in saline and bone. However, scattering in the bone and a potential light-pipe effect caused by its anisotropy may have a strong influence on the photon count at the SGNs. Also, alignment of the beam toward the SGNs hidden in the modiolus is crucial for optogenetic stimulation in the cochlea.

The beam simulation by Hernandez et al. (2014b) was fit for a mouse model, where the distance of the micro-LED array to the modiolus was only 100 μm, which seems almost impossible for humans. In the case of fiber-optic stimulation, the experiments of Huber et al. (2008) (see Figure 6) match a potential configuration in the human cochlea quite well. They had an LED (500 μm) positioned with a distance of 500 μm above the brain surface and included scattering by a brain tissue of thickness 250 μm. They measured the profile of the light intensity and found relatively broad (intensity dropped by 50% outside a diameter of 2 mm) excitation. If the condition would be similar in the human cochlea, such a spread would provide only less than 15 optical channels.

Figure 6. Lateral light spread as a function of distance from a 200 μm optical fiber (numerical aperture = 0.37) in saline solution (left) and rat gray matter (right) for blue (473nm; left) and yellow (594 nm; right) light. Contour maps show where light intensity decays by 50% (red), 10% (orange), 5% (yellow), and 1% (cyan) with increasing distance from the fiber end. Light spreads conically in saline, mostly due to fiber properties. In brain tissue, scattering is much larger and causes high attenuation and also widening of the light beam. © Elsevier. Reproduced by permission of Elsevier. Permission to reuse must be obtained from the rightsholder.

6. Discussion of potential benefits and outlook

Models are very important tools to evaluate new technologies, such as optogenetical neuroprostheses. With quantitative models, technical results achieved in one species can be extrapolated to other species or from one system to the other. This allows educated estimates of possible benefits of a new technology without taking the risks of its application. Such a procedure is especially important in the case of optogenetic stimulation, where the risks involved in gene transfection have to be carefully balanced against the possible benefits. Here we assess the most important technical benefits.

6.1. Number of independent channels

In the case of CIs, the spread of spatial excitation limits the number of effective (independent) stimulation channels. Experimentally, optogenetic stimulation of the auditory nerve so far was only achieved by Hernandez et al. (2014a). In these experiments, the spatial spread of stimulation was only indirectly accessible with recordings in the inferior colliculus. With light models, at least estimations of the spatial extent of neuronal excitation in humans can be achieved. However, it has to be considered that light models are either solved with Kubelka–Munk or a Monte Carlo simulation that do not take into account the anisotropic properties of the bone in the cochlea. Light conduction in the bone can have an influence on the excitation spread and needs to be further investigated. The same is true for sheath formation and tissue growth around the electrode (Nadol and Eddington 2004). Its effect on photo absorption and light scattering is unclear and requires further investigation. A solid knowledge of the light reaching SGNs is crucial to understand where and how selectively these neurons are excited, especially if we consider that they have an extended structure and run from the osseous spiral lamina to the center of the modiolus and then from apex to base inside the modiolus. When we transfer the results of Hernandez et al. (2014a) to the conditions in the human cochlea, the extent of spatial excitation is expected to be about 250 μm only if the LEDs can be placed directly on the modiolar bone (215 μm distance from the LED to the nerve). If it would be possible to insert a dense LED array up to 25 mm into the human cochlea, we could realize about 100 independent stimulation channels which is one order of magnitude more than what can be achieved with electrical stimulation (Friesen et al. 2001). If the LEDs would be further away from the modiolus after implantation, this would cause a massive increase of the illuminated area due to the broad radiation angle of LEDs. In addition to the larger crosstalk caused by this effect, an even worse consequence would be the diminished light flux reaching the neurons. To compensate this effect, it would be necessary to increase the power of the LEDs which would increase unwanted heating of the cochlea and reduced battery life time. LEDs with microlenses could counteract this effect; however, the optrode would become much more bulky. In addition, a more focussed light source means that it has to be aimed more precisely at the target neurons. It is unlikely that this issue can be resolved easily for optical sources in the cochlea.

VCSELs would not require optics, as their beam width is narrower than that of LEDs. They are widely known in telecommunication applications and a wide variety of products is available. Die sizes are below 250 × 250 × 150 μm, which would allow their insertion directly into the cochlea, at least in theory. However, optogenetic applications in the cochlea would require the development of custom-built devices (wavelength, power, power dissipation). A big challenge would be the development of flexible optrodes which can be inserted deeply into the cochlea without damage.

The same is valid for optical fibers. Their physical properties and manufacturing methods are well known and their flexibility allows the insertion into bent cavities. Still, the need for reliable coupling with a light source, the space required to place the light source and the task to aim the light from multiple fibers toward the neurons would pose many challenges for the development of optical CIs.

As with all optical emitters, the alignment of the beam with the target is crucial, as deviations from the main axis result in a drop in intensity. This way thresholds might not be reached or the required power to drive the light source could by far exceed its electrical counterpart. Self-aligning and adjustable microlenses could help solve this challenge (Buchegger et al. 2009).

6.2. Temporal properties

The photocycle of opsins involves relatively slow channel dynamics which seems to be a large disadvantage compared to electrical stimulation. However, a closer look at the model (see Figure 2) reveals that only the channel closing time constants are slow, whereas the activation of ChR2 takes only about 0.2–1 ms (Nikolic et al. 2009). While the value of the activation time constant is in the order of the membrane time constant, the deactivation of ChR2 currents is much slower (above 10 ms, Kleinlogel et al. 2011). Therefore the deactivation time constant limits the maximal firing rates which can be achieved when ChR2 is expressed in neurons (measurements show maximal sustained firing rates up to 20 spikes/s, Boyden et al. 2005). Opsins are often characterized in cells with slow membrane time constants, which makes it harder to predict how they function in auditory neurons, which usually have very fast membrane time constants. For example, Klapoetke et al. (2014) characterized the dynamical properties of a number of channelrhodopsins in human embryonic kidney cells. Channel dynamics measured with voltage clamp was characterized by the time to reach 90% peak current and a monoexponential fit of the channel closing rate. Values for ChR2 were 5.92 ± 0.26 ms (onset to 90%) and 14.6 ± 1.4 ms (channel turn off kinetics). Chronos reached values of 2.2 ± 0.28 ms and 3.59 ± 0.21 ms, respectively, and was the fastest opsin investigated so far. However, in the experiments with human embryonic kidney cells, the membrane resistance was in the range of 500 M and the membrane capacitance above 40 pF, yielding a membrane time constant above 20 ms. In their measurements, at a stimulation rate of 60 Hz the spike probability started to decline. As this was also the case with electric stimulation, this limit was rather due to the cell’s membrane time constant and not so much a limitation of the opsin itself (Klapoetke et al. 2014, supplemental figure 11). The first spike latency was in the order of 2 ms and its standard deviation was below 0.1 ms with Chronos (Klapoetke et al. 2014, supplemental figure 10).

The measured temporal properties improve when the opsins are expressed in neurons with faster dynamics. Gunaydin et al. (2010) could drive hippocampal interneurons expressing ChR2 reliably up to 10 spikes/s with 5 ms light pulses. With neurons expressing ChETA, which has a deactivation time constant of about 5 ms, they achieved maximal sustained firing rates of 200 spikes/s (Gunaydin et al. 2010). This comes already close to the physiological value of the maximum steady-state firing rate of SGNs which is around 300 spikes/s (Kiang 1965). Such a rate would probably be sufficient for optical CIs, because Fu and Shannon (2000) have shown in electrical CIs, stimulation rates above 150 pps do not improve speech recognition. Another advantage of ChETA over ChR2 is its higher temporal fidelity: spike patterns of neurons expressing ChETA to poisson-distributed light pulses followed the stimulus pattern much better than neurons with ChR2. Neurons expressing ChR2 did not fire for short inter-stimulus intervals but on the other hand, they frequently generated extra spikes (Gunaydin et al. 2010), which was not the case with ChETA. Hight et al. (2015) managed to transfect auditory brainstem neurons in mice with ChR2 and Chronos and made recordings in the inferior colliculus. They found strong excitation to 1 ms laser pulses up to a stimulation rate of 448 Hz with Chronos. In addition, they proved that the temporal fidelity of the light stimulation with Chronos was higher than with ChR2. However, they could measure responses not directly in the stimulated neurons but only in the next neuronal stage, where they evaluated the synchronization of the neurons to the stimuli. On the other hand, Hernandez et al. (2014b) were able to transfect the neurons of the primary auditory nerve with ChR2 or CatCh (a variant of ChR2 with higher light sensitivity) and also to record them directly. They found that first spike latency decreased (approximately inversely proportional) to light intensity and reached values of 4.6 ± 0.3 ms (ChR2) and 2.8 ± 0.2 ms (CatCh) in SGNs for strong (22 mW) light stimuli. The temporal precision of optical stimulation, quantified by the trial-to-trial variance in FSL, was 80 ± 60 μm/s² (ChR2) and 20 ± 0.2 μm/s² (CatCh). This corresponds to a standard deviation of 8.9 ms and 4.5 ms, respectively. For electrical stimulation, Shepherd and Wicke (1997) quantified the temporal jitter with the standard deviation of the period histogram peak and found values between 33 and 41 μs (mid-dynamic range to saturation) which is still two orders of magnitude less than for optical stimulation. The larger jitter is at least partly due to the fact, that for optical stimulation, much longer pulses were required (10 or 5 ms), compared to electrical stimulation, where the phase duration was 100 μs (Shepherd and Wicke 1997). With shorter pulses, neurons depolarize faster and therefore fire with less jitter. But with the development of highly efficient opsins and high-power light sources with very small spatial spread, there is still room to further improve the temporal precision of optogenetic stimulation, such that it could even reach the precision of the synaptic transmission in the intact inner ear.

6.3. Summary of advantages and risks

While this review focussed on technical issues concerned with optogenetic stimulation, it is clear that this technology also introduces biological risks. Obvious risks are due to the transfection of neurons with new genetic material which are reviewed in Mingozzi and High (2013) and Ivanova and Pan (2009). In addition, the relatively slow time constants involved in the closure of light-sensitive ion channels introduces additional stress in the transfected neurons to restore and sustain the required ionic gradients. In this respect, unwanted side effects due to H⁺ and nonspecific Ca²⁺ influx could affect the neurons (Mahn et al. 2016) and, in the worst case, shorten their life. Direct tissue damage is also caused by high light power (>100mW/mm², Cardin et al. 2010), but in most cases excitation thresholds are smaller (Hernandez et al. 2014a) and using short light pulses also minimizes tissue damage.

From a technical point of view, implants using optogenetic stimulation with ChR2 would require 10 times more energy than electrical stimulation (2 μJ compared to 0.2 μJ, Hernandez et al. 2014a). This disadvantage could be diminished with novel, more light-sensitive opsins which would probably also allow shorter light pulses and thus improve the temporal precision of optogenetic stimulation. On the other hand, electrical stimulation involves so much crosstalk that it requires interleaved stimulation startegies. Because of that, the high temporal resolution of electrical stimulation measured with single channel stimulation is effectively severely reduced in contemporary CI coding strategies. If concurrent and independent optical stimulation would be possible due to the confined optical stimulation of neurons, the temporal resolution would come closer to what CIS strategies achieve with electrical stimulation now. Ultimately, the total amount of information that can be coded into the auditory nerve is proportional to the number of independent channels, their temporal resolution and the number of resolved amplitude steps. Right now, it seems possible that optogenetic stimulation has the potential to outperform electrical stimulation, especially due to the larger number of independent stimulation channels. However, this is only true if the optical stimulation system provides a highly selective stimulation. This might not be possible with LEDs that lie at the radial edge of the scala tympani, at least not without focussing elements.

It is necessary that we assess the probable potential of optogenetic stimulation critically which definitely requires more experiments in combination with models. The key question will be if the optical stimulation can be locally confined such that it permits independent stimulation of a much larger number of neurons than what is possible with electrical stimulation. The critical technical points will be 1) how the light beam can be directed toward the target, 2) how broad and 3) how intense the beam is after diverging, absorption, and scattering in the cochlear fluids but especially in the bony modiolus, 4) if independent groups of SGNs, which run through the modiolus from apex to base, can be excited, 5) if there is a sufficient number of surviving SGNs available to convey information from a high number of channels to the brain, 6) how this information is best coded, and 7) if it is possible to fabricate robust and long-lived optical stimulation systems with reasonable power requirements. From a clinical point of view, the safety of gene transfection and the question of whether active or passive optrode arrays can be implanted safely is of utmost relevance. Another advantage of optogenetic stimulation, which could well become important in future systems, is that it would produce smaller artifacts for concurrent (electrical) recordings of neuronal activity. This would make a closed-loop stimulation system easier (Pashaie et al. 2015), meaning a system where the stimulation is adjusted based on the elicited neuronal activity. Closed-loop systems are ubiquitous in the neuronal system and the technical realization of a feed-back loop in the stimulation strategy has the potential to enhance the coding of relevant information. Still, if all technical, biological, and medical hurdles are solved, we will have to critically assess if the possible advantages of optogenetic stimulation still outweight its disadvantages and risks.

Acknowledgments

We want to thank our reviewers for their valuable comments.

Funding

This work was supported by the German Federal Ministry of Education and Research within the Munich Bernstein Center for Computational Neuroscience (reference No. 01GQ1004B) and by the DFG “Ultrafast and temporally precise information processing: normal and dysfunctional hearing” SPP 1608 (HE6713/1-1 and 1-2).

Additional information

Funding

This work was supported by the German Federal Ministry of Education and Research within the Munich Bernstein Center for Computational Neuroscience (reference No. 01GQ1004B) and by the DFG “Ultrafast and temporally precise information processing: normal and dysfunctional hearing” SPP 1608 (HE6713/1-1 and 1-2).