Key words

1 Introduction

One of the main questions in modern neuroscience is how the activity of identified neuronal populations generates behavioral and mental states [1]. Recent advances in optical techniques that allow the simultaneous recording and manipulation of neurons with single-cell resolution [2, 3] combined with population analytical tools [4,5,6,7,8] suggest that neuronal ensembles are units of brain computation [9, 10]. Neuronal ensembles are groups of neurons with coordinated activity that may underlie sensations, perceptions, emotions, and memories. In order to prove causality between the activity of specific neuronal ensembles and learned behaviors, it is becoming clear that the ability to manipulate and record hundreds of neurons simultaneously needs to be guided by analytical tools that allow the identification of neurons that could have a deterministic impact in brain states. Holographic two-photon imaging [11], two-photon optogenetics [12], and analytical tools [13] offer the possibility to target neuronal ensembles to control behavioral states. The demonstration of the causal relation between neuronal ensemble activity and behavior has been achieved recently in different brain areas [4, 14,15,16,17,18], paving the pathway to design more sophisticated experiments to understand complex mental states in health and disease [19].

The optimal design of all-optical experiments to control learned behaviors with single-cell resolution requires the simultaneous reading and writing of neuronal activity. This could be achieved in many ways [20], but in this chapter we will focus mainly on scanning and parallel optical techniques guided by analytical tools. We describe the implementation of scanning two-photon imaging and parallel two-photon optogenetics using a spatial light modulator (SLM). We also describe the main concepts necessary to identify and target neurons with pattern completion capability [6, 21] that can recall neuronal ensembles related to a visually guided behavior [4].

2 Implementation of Simultaneous Two-Photon Imaging and Two-Photon Optogenetics

2.1 Background

2.1.1 Light-Sensitive Sensors and Actuators of Neuronal Activity

Optical methods are powerful to record and manipulate neuronal activity at single-cell resolution across a large population—a key requirement to investigate neuronal ensembles. Comparing with electrophysiology, optical methods can simultaneously sample and target a large group of neurons with high spatial specificity in a noninvasive manner and can be conveniently implemented for in vivo studies. The cornerstone of optical methods is the optical microscope, which uses light to record and manipulate the neurons that are illuminated. As the neurons in the brain do not typically respond to light, they can be loaded or transfected with light-sensitive sensors and actuators that can respectively report and manipulate their neuronal activity upon light illumination. Using these neuronal activity sensors and actuators, an optical microscope can simultaneously read and write neuronal activity across a large field of view with high spatial specificity.

Calcium [22,23,24,25,26] and voltage indicators [27, 28] are two commonly used sensors for neuronal activity. These indicators are embedded with fluorophores, which can absorb the illumination or excitation light, and emit fluorescence. The efficiency of this light-absorption, fluorescence-emission process is modulated by the intracellular calcium concentration and membrane potential respectively in calcium indicators and voltage indicators (See also Chaps. 1 and 2). Thus, the neuronal activity can be deduced from the time-lapse recording of the fluorescence in individual cells. The commonly used calcium indicators include GCaMP6 [26], jGCaMP7 [29], jRGECO [30], etc. Compared with calcium indicators, voltage indicators are less mature with smaller signal-to-noise ratio and prone to photo-bleach though much progress has been made in the past years [31] (Chap. 2). The counterparts of these activity sensors are actuators, which could be chemical-based neurotransmitter cage-compounds [32, 33] or opsins [34, 35] serving as light-sensitive ion channels. When absorbing the light, the cage-compounds could release the neurotransmitters (uncaging) and the opsins could open or close the ion channels (optogenetics) and thus control the membrane potential. This mechanism thus allows light control of neuronal activity in individual cells.

2.1.2 One-Photon and Multiphoton Excitation

Depending on how light is interacting with the light-sensitive sensors and actuators, optical methods can be classified into two categories: one-photon and multiphoton (i.e., two-photon [36, 37] or three-photon [38, 39]). In one-photon case, the interaction is linear. The fluorescence emission rate of the sensors and the actuation strength of the actuators are proportional to the intensity of the excitation light before saturation. While one-photon excitation is straightforward, it lacks spatial specificity in 3D. As the excitation would form a double cone-shape pattern along the axial direction (Fig. 1), all neurons with the sensors or actuators within the double cone may absorb the light. It is thus very challenging to just excite a single neuron in a 3D volume. Furthermore, in terms of imaging, the out-of-focus fluorescence emission typically becomes background on the image captured at the focal plane and thus reduces the image signal-to-noise ratio. As the light-absorption and fluorescence-emission cycles create phototoxicity, one-photon excitation pays a high price to image the focal plane by creating phototoxicity within the entire double cone volume. Multiphoton excitation can greatly overcome these limitations, as the interaction between the light and the sensors or actuators is nonlinear. In the two-photon case, the excitation effect is proportional to the intensity square of the excitation light before saturation, leading to a strong gradient of the excitation effect along the optical axis in the double cone. Thus, it is feasible to control the incident light so only the light intensity at the focal point (i.e., the tip of the double cone) is strong enough to excite the neuronal sensors or actuators (Fig. 1). This greatly improves the spatial specificity or resolution and reduces the phototoxicity in out-of-focus region. Another advantage of two-photon excitation is that the excitation light has a longer wavelength, which reduces the light scattering effect in scattering brain tissues. Thus two-photon excitation can penetrate much deeper in the brain compared with one-photon excitation. The drawback of two-photon excitation is that much more laser power is required due to a lower efficiency of multiphoton absorption. As the excitation light eventually turns into heat, more heat will be generated inside the brain. This is typically not a concern as in a typical imaging experiment, the laser power used is less than the brain damage threshold [40], even at very deep layers. In the past two decades, two-photon microscopes have become one of the workhorses in neuroscience. Recently, three-photon excitation has been successfully demonstrated for both calcium imaging [39] and optogenetics [41]. Comparing with two-photon excitation, three-photon excitation has an even higher nonlinearity and could further increase spatial specificity and penetration depth, though the laser power should be managed so it will not exceed the brain damage threshold.

Fig. 1
Two illustrations of photon excitation labeled a and b. A is a one-photon excitation with an hour glass structure. Below it is a 488 nanometer focal plane. B is a two-photon excitation with an hour glass structure, below it is a 960 nanometer focal plane.

Comparison between one-photon and two-photon excitation. In one-photon excitation (a), a double cone is formed by focusing the visible excitation light (indicated by blue color); fluorophores (green) in the entire double cone could be excited. In two-photon excitation (b), while a double cone is still formed by the focus of infrared excitation light (indicated by red color), the fluorescence generation (green) is localized at the vicinity of the focal region. Reprinted with permission from [101], Springer Nature. Experimental illustration of the one-photon and two-photon excitation (0.16 NA for both cases) is shown in the bottom panel. (Reprinted with permission from Ref. [102], Springer Nature)

Full size image

2.1.3 Basic Setup of Multiphoton Microscope

Since the multiphoton absorption rate is generally low, and it has a nonlinear relationship with the light intensity, a femtosecond laser is required for multiphoton excitation. The femtosecond laser delivers a periodic pulse train, typically with a repetition rate of 1~80 MHz. The temporal width of a pulse is below 300 fs, yielding a very high instantaneous peak power and thus a high multiphoton absorption rate. Unlike one-photon excitation where a large area or volume of sample could be simultaneously illuminated and their fluorescence could be detected through a camera in the case of imaging, the illumination of multiphoton light is typically through a rapid scanning of the laser beam on the sample. In multiphoton imaging, a single pixel detector (versus pixel array in a camera) is used to record the emitted fluorescence (see also Chap. 2). By correlating the temporal signal from the detector and the scanning trajectory, the image can be built. The microscope setup (Fig. 2a) typically includes a raster lateral (xy) scanning system composed of a galvanometer scanner and a resonant scanner (30~60 Hz frame rate), or two galvanometer scanners (4~10 Hz frame rate). For volumetric imaging, an axial (z) scanning system is implemented by adding a piezo-electric controller on the objective lens or inserting axial focusing devices, such as electrically tunable lens [42], spatial light modulator [43], or remote focusing unit [44,45,46] (~ ms focus switching time). Different axial planes are sequentially scanned (Fig. 2b). While this configuration of lateral and axial scanning is straightforward, it may not be the most efficient as it blindly samples the brain tissue. There could be a large volume of “empty” extracellular space without neurons, and among all the neurons, typically only a subset of them are labeled with the calcium indicators. Random access scanning techniques [47,48,49,50,51] can overcome this issue. The laser focus spot can jump rapidly between different regions in the sample in 3D (<20 μs transition time between different regions) (Fig. 2c). This is enabled by acousto-optic deflectors (AODs). The challenges of this technique are to overcome the spatial and temporal distortion of the ultrashort laser pulses, caused by the angular dispersion from the AODs’ phase grating and the group delay dispersion of the AOD crystals, respectively. These limit the imaging field of view. While these distortions can be compensated [51], the optical setup is complex. AODs also have a limitation of high insertion loss. Recently, many advanced scanning mechanisms including beam multiplexing [43, 52,53,54,55] are proposed and demonstrated to increase the imaging throughput. For extensive reviews, see [56,57,58,59].

Fig. 2
Three illustrations of laser scanning microscope labeled a to c. A is a setup of laser scanning microscope. Some of its components are pocket cells, Z-scan system, scan lens, tube lens, P M T, and so on. B is a trajectory in volumetric imaging. C is a 3-D coordinate plane with a trajectory.

Basic setup of two-photon laser scanning microscope. (a) Schematics of a typical two-photon laser scanning microscope. The Pockels cell is used to modulate the laser light intensity. The xy-scan system and z-scan system are used to scan the laser focal spot laterally and axially in the sample, respectively. The dichroic mirror transmits the infrared excitation light and reflects the visible fluorescent light to the photomultiplier tube (PMT). (b) Schematics of the typical scanning trajectory in volumetric imaging. (c) Schematics of an exemplary scanning trajectory using random access scanners. The green regions indicate neurons of interest. (Reprinted with permission from Ref. [56], Springer Nature)

Full size image

The setup of multiphoton photostimulation is similar as multiphoton imaging, but without the detection module. The laser spot on the sample is typically controlled by a pair of galvanometer mirrors, and raster or spirally scanned across each target neuron (1~100 ms) before jumping to another neuron. Alternatively, the combination of beam shaping and temporal focusing techniques [60,61,62,63,64] can be implemented so that a disk pattern with tight axial confinement can be projected to the entire neuron. The scanners no longer perform raster or spiral scanning, but direct the disk excitation pattern to different neurons sequentially [3]. Using the holographic approach, illumination patterns could be simultaneously projected to multiple targeted neurons, so the scanners are no longer required [62,63,64,65,66]. We will discuss this topic in depth, as well as holographic illumination technique that can simultaneously photostimulate multiple neurons, in the next section.

In the following, we focus on our implementation of a microscope that can simultaneously perform calcium imaging to record neuronal activity and optogenetics to manipulate neuronal activity.

2.2 Simultaneous Two-Photon Imaging and Two-Photon Optogenetics

2.2.1 Overall Consideration

An all-optical method [20] refers to simultaneously recording and manipulating the neuronal activity through light. Here, we describe the combination of two-photon calcium imaging and two-photon optogenetics in a microscope. To flexibly image and manipulate neuronal ensembles, it is desired to have the following features in the microscope:

  1. (i)

    Calcium imaging and optogenetics should be independently controlled and performed.

  2. (ii)

    Two-photon imaging could image neuronal activity over a large neuronal population with cellular resolution and high temporal resolution (4~30 Hz for calcium imaging).

  3. (iii)

    Two-photon optogenetics could simultaneously photostimulate multiple neurons with cellular resolution and high temporal precision (sub-millisecond to milliseconds).

To simultaneously, while independently, perform calcium imaging and optogenetics, the microscope should have two independent beam paths, respectively, for calcium imaging and optogenetics (Fig. 3a) [2, 3, 66,67,68,69]. These two beam paths then merge together before the light is directed into the brain tissue. Each beam path should be implemented with an individual femtosecond laser with a different wavelength, which corresponds to the central excitation spectrum of the calcium indicators and opsins. It is important that they have a different excitation spectrum, so as to prevent crosstalk between imaging and photostimulation (see also Chaps. 1, 2, and 5). In other words, when performing calcium imaging, the imaging laser should not excite the opsin; when performing optogenetics, the photostimulation laser should not excite the fluorescence of the calcium indicator. There are two common options to choose the calcium indicator and opsin pair: green calcium indicators such as GCaMP [26, 29] with red-shifted opsins such as C1V1 [70], or red-shifted calcium indicators such as jRGECO [30] with blue-activated opsins such as ChR2 [34]. While both of these pairs have minimal overlap in the excitation spectrum, it is critical to keep the imaging laser power low and perform control experiments to ensure that the imaging laser does not increase the spiking rate of the neurons (see Note 1). In case the photostimulation laser creates fluorescence artifacts, the imaging data should be processed to eliminate the artifacts. For a successful experiment, it is also important to have a high co-expression rate of the calcium indicators and opsins on a large neuronal population. The cooperation of both constructs into a single virus could be a promising approach [14, 71]. An example of co-expression of GCaMP6s and C1V1 using two different viruses is shown in Fig. 3c.

Fig. 3
Three illustrations of calcium imaging and holographic optogenetics labeled a and b. A is a setup for calcium imaging via imaging laser, simulation laser, spatial light modulator, scan lens, and tube lens. B has imaging and stimulation laser, on the left and 3-D stimulation pattern, on the right. C is a neuron field.

Simultaneous two-photon calcium imaging and two-photon holographic optogenetics. (a) Two-photon microscope setup with two independent beam paths for calcium imaging and photostimulation. An electrically tunable lens is equipped in the imaging path so the focal plane can be rapidly switched for high-speed volumetric imaging. The dichroic mirror 3 is used to spectrally separate the fluorescence into green and red channel. In the photostimulation path, a half-wave plate is placed before the spatial light modulator to align the polarization of the laser to the active axis of the spatial light modulator. The spatial light modulator then creates a hologram in the sample to photostimulate the regions of interest (e.g., a group of neurons). A pair of relay lens (4f system) is used to transfer the light field to a set of galvanometer mirrors which can spirally scan the holographic pattern. At the intermediate plane of this pair of relay lens, a zeroth-order beam block is used to block the residue light that is not modulated by the spatial light modulator. The imaging laser and photostimulation laser have different wavelengths, and they are combined through the dichroic mirror 1 before being delivered to the sample through the scan lens, tube lens, and objective lens. HWP half-wave plate, PMT photomultiplier tube. (b) Schematics for simultaneous volumetric calcium imaging and 3D holographic patterned photostimulation in mouse cortex. (c) An exemplary field of view showing neurons co-expressing GCaMP6s (green) and C1V1-mCherry (magenta). (Reprinted and adapted from Ref. [69])

Full size image

To characterize neuronal ensemble activity, it is necessary to image a large population of neurons, typically on the order of hundreds. For visual cortex, a minimal field of view of 200 × 200 μm2 is necessary, though 400 × 400 ~ 600 × 600 μm2 would be more desirable. Using low magnification objective lens, the field of view can go above 1 mm2. Furthermore, volumetric imaging can be performed to image neuronal ensembles in 3D and across different functional layers (Fig. 3a, b). Due to the sequential scanning nature of two-photon microscopy, there is a tradeoff between the field of view and spatiotemporal resolution. Thus, while it is desirable to increase the field of view so as to image more neurons, one should maintain cellular resolution with a good signal-to-noise ratio (i.e., able to distinguish the calcium transients), as well as a frame rate of at least 4–10 Hz. In the setup shown in Fig. 3a, we use a galvanometer and a resonant scanner pair for lateral scanning, and an electrically tunable lens for fast focal plane switching. Volumetric imaging of 3 planes could be achieved at a volume rate >6 Hz, with a field of view ~500 × 500 μm2 per plane [69].

A typical experiment starts with two-photon calcium imaging on the brain tissue. The animal is head-fixed and stabilized on the microscope stage and could perform behavior tasks. The spatial footprint as well as the neuronal activity of individual neurons within the image field of view can then be extracted from the recording. The neuronal ensemble can be identified from the activity pattern (see Subheading 3). Depending on the applications, the users may choose a specific population of neurons to perform optogenetics experiments. It is thus desirable to perform photostimulation on multiple neurons simultaneously. Two-photon holographic illumination enables photostimulating a group of user-selected neurons [2, 12, 14, 66, 68, 69, 72, 73]. We discuss the topic in the following section.

2.2.2 Holographic Illumination

Holographic illumination refers to projecting to the brain tissue a computer-generated holographic pattern. Neurons falling within this pattern can then be photostimulated. Computer-generated hologram is a light field of an arbitrary shape in 3D in an imaging volume, and this shape can be dynamically controlled and changed through a spatial light modulator (Fig. 4a). In the simplest imaging system, which contains a single lens, to generate a hologram in the imaging space (i.e., around the front focal plane of the lens), one can spatially modulate the light field at the back focal plane of the lens. This light field then propagates through the lens, and coherently interferes at the front focal plane and forms the holographic pattern (Fig. 4a). As the light field at the front focal plane and back focal plane of a lens forms a Fourier transform relationship, one can calculate how the light field should be spatially modulated at the back focal plane based on the desired pattern at the front focal plane (Fig. 4b, c). The typical spatial light modulator is based on liquid crystals and is configured to only modulate the phase of the light field. A Gerchberg-Saxton algorithm (Fig. 5) [74], which is an iterative approach, can be used to calculate the phase pattern on the spatial light modulator, given the amplitude of the desired pattern in the imaging space and the amplitude of the light field incident on the spatial light modulator. If the holographic pattern only contains a group of points with different weights, a superposition algorithm [55, 75], which is essentially a single iteration of the Gerchberg-Saxton algorithm, can be used to calculate the phase pattern (Table 1).

Fig. 4
Three illustrations for computer-generated hologram labeled a to c. A is a mechanism of 3D image formation via S L M and lens. B is a set of 4 focal spot patterns formed via nominal, lateral, axial, and lateral and axial shifts. C is a set of 4 photos titled target image, phase mask, calculated image, and measured fluorescence.

Computer-generated hologram. (a) Principle of computer-generated hologram. The collimated laser beam is incident onto the spatial light modulator (SLM), which spatially modulates the wavefront of the light. The light field then propagates through a lens and forms the desired 3D image in the imaging domain through interference. f, focal length of the lens. (b) By modulating the spatial phase profile at the back focal plane of the object lens, different focal spot patterns can be formed in the imaging domain. (Reprinted with permission from Ref. [103], Elsevier). (c) Example of a two-photon SLM hologram. The four panels illustrate the binary target image, the spatial light modulator phase hologram generated by Gerchberg-Saxton algorithm, the squared image (to mimic two-photon excitation) of the projected pattern back calculated from the phase hologram, and the experimentally measured two-photon fluorescence image generated by the SLM. A stylized picture of Cajal is used as a target image. (Reprinted from Ref. [11], Frontiers)

Full size image
Fig. 5
An algorithm to calculate the phase hologram. The components are as follows. Source amplitude, Fourier transform, amplitude of generated pattern at imaging space, ideal target amplitude, inverse Fourier transform, phase hologram for S L M, S L M plane and imaging plane.

Gerchberg-Saxton algorithm. This algorithm is used to calculate the phase hologram φs for the spatial light modulator (SLM) based on the amplitude of the target pattern At in the imaging space and the amplitude of the incident light field As onto the SLM

Full size image
Table 1 Analytical expression of the phase hologram and the Zernike polynomials and Zernike coefficients in superposition algorithm
Full size table

In two-photon microscopes, we could not directly modulate the phase at the back focal plane of the objective lens, as the back focal plane is typically inside the objective lens housing. Relay lens pairs (4f system) can create the conjugate planes of the back focal plane of the objective lens, and the spatial light modulator can be placed in one of those conjugate planes (Fig. 3a). Furthermore, as there is residue light that is not modulated by the spatial light modulator (termed zeroth order beam), it will create a strong focus at the imaging space (see Note 2). The relay lens pairs help to resolve this issue as a small beam block can be placed at the intermediate plane (which is conjugate to the focal plane at the imaging space) to block the zeroth order beam (Fig. 3a). Using the holographic illumination, various patterns can be generated in the imaging space (Fig. 4b, c).

Holographic illumination essentially spatially multiplexes the excitation beams, and allows multiple neurons being photostimulated simultaneously. However, this comes with an increase of laser power on the brain tissue. To alleviate this issue, a laser with a lower repetition rate can be used. When the pulse repetition rate is reduced, the energy in each laser pulse is increased, while keeping the overall average power the same. As the two-photon excitation efficiency is proportional to the square of the laser peak power, the increase of excitation efficiency due to the increase in pulse energy outweighs its reduction due to the reduced number of pulses per unit time. Thus, using a laser with lower repetition rate, a higher overall excitation efficiency can be achieved while keeping the same average power. In other words, to achieve the same excitation efficiency, the average laser power can be reduced. In the holographic photostimulation system described, we used a laser with a repetition rate of 1 MHz [69]. This reduces the overall laser power by 80 times compared with a commonly used 80 MHz femtosecond laser.

Due to the chromatic dispersion and spatial discretization of the SLM pixels, the SLM has a spatially varying diffraction efficiency [75]. The diffraction efficiency drops as the deflection angle increases, limiting the addressable 3D field of view of SLM. To alleviate this effect, the diffraction efficiency can be first measured, and then compensated in the Gerchberg-Saxton algorithm or through the weight factor Ai (Table 1) in the superposition algorithm [55, 75]. The laser intensity across the field of view can then be made uniform. By using an XY galvanometer set to provide a lateral offset to the centroid of the SLM’s addressable field of view, the effective lateral field of view can be further extended. While neurons across this enlarged field of view cannot be targeted simultaneously, groups of neurons located at different sub-fields can be targeted by switching the offset of the XY galvanometer and the phase pattern on the SLM [76].

2.2.3 Spiral Scan Versus Scanless Approach in Holographic Photostimulation

As mentioned in Subheading 2.1, photostimulation can be performed by raster or spirally scanning the laser focal spot across the neuron, or by projecting a disk pattern which matches the morphology of the neuron so the entire neuron can be stimulated at once. The same applies in holographic photostimulation. In the first approach, a group of focal spots are created in the hologram, and each focal spot lies on the centroid of the individual targeted neuron [2, 12, 69]. A set of galvanometer scanners then spirally scan the entire group of focal points (with certain repetitions), so each focal spot spirally scans across the corresponding targeted neuron. In the second approach, a group of disk patterns are created in the hologram and projected to the brain tissue [72, 77, 78]. Each disk spatially overlaps with a targeted neuron. Since the hologram already generates multiple disks for all the target neurons, this method can work without scanners, and thus this is a scanless approach. In two-photon excitation, these two approaches have their own advantages and limitations. The spiral scan approach spatially concentrates the laser intensity into focal spots; as two-photon excitation effect is proportional to light intensity square, this approach has a high excitation efficiency. As the entire neuron is not photostimulated at the same time, to reach the threshold of evoking action potentials, it relies on the accumulation of the excitation effect along the spiral trajectory. The decay constant of the opsin kinetics (tau-off) should thus be longer than the duration of each spiral (i.e., the opsin channels can stay open during the one single spiral scan, which could be less than 1 ms); otherwise the excitation effect cannot be accumulated. The scanless approach, on the other hand, disperses the light intensity across the entire neuron, and thus it requires a higher average power to reach the same excitation strength as the spiral scan approach. However, as the entire neuron is being stimulated at once, it does not pose a limitation on the kinetics of the opsin. It is also expected that the jitter of the delay between the onset of photostimulation and onset of action potential is smaller. As the opsin decay constant (tau-off) is typically larger than the duration of a single spiral scan (which can be <1 ms), the scanning approach is in favor as it takes lower laser power to photoactivate the neurons. In our experiments, we used C1V1 as the opsin. Using spiral scan, it takes about 1.8× less power to photoactivate the neurons than the scanless approach, with the same photostimulation duration (Fig. 6) [69]. For opsins with faster kinetics, a faster spiral scan or the scanless approach may be preferred.

Fig. 6
Three graphs labeled a to c. A is a set of 12 line graphs with high and low peaks arranged in 3 rows and 4 columns. B is a line graph that plots normalized delta F over F versus modality, relative power, and duration. C is a box plot, that plots normalized delta F over F versus modality, relative power, and duration.

Comparison between spiral scan and scanless holographic approaches for photostimulation. In the scanning approach, the laser spot is spirally scanned over the cell body; in the scanless approach, a disk pattern is generated by the SLM, covering the entire cell body at once. (a) Photostimulation triggered calcium response of a targeted neuron in vivo at mouse layer 2/3 of V1, for different stimulation modalities. For each modality, the multiplication of stimulation duration and the square of the laser power was kept constant over four different stimulation durations. The average response traces are plotted over those from the individual trials. (b) ΔF/F response of individual neurons on different photostimulation conditions (layer 2/3 of V1, over a depth of 100 ~ 270 μm from pial surface; one-way ANOVA test). For each neuron and each stimulation duration, the laser power used in the scanless disk modality is 1 and 1.8 times relative to that in the spiral scan. For each neuron and each modality, the multiplication of the stimulation duration and the square of the laser power was kept constant over four different stimulation durations. (c) Boxplot summarizing the statistics in (b). The central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points (99.3% coverage if the data are normal distributed) not considered outliers, and the outliers are plotted individually using the “+” symbol. In this experiment, the mice were transfected with GCaMP6f and C1V1-mCherry. Repetition rate of the photostimulation laser is 1 MHz. The spiral scan consists of 50 rotations to cover the neuronal cell body, and the scanning speed is adjusted to make different stimulation durations. (Reprinted with permission from Ref. [69])

Full size image

2.2.4 Detailed Implementation of the Microscope for Simultaneous Two-Photon Imaging and Two-Photon Holographic Photostimulation

In this section, we explain the details of how the microscope can be constructed with two beam paths for two-photon imaging and two-photon holographic photostimulation (Fig. 3a). We aim to help the readers understand the inside of a commercial microscope system and meanwhile provide basic guidelines for those who want to home-build the microscope. Here, we use GCaMP6 as calcium indicators, and C1V1-mCherry as the red-shifted opsin. The mCherry can be used to indicate if the opsin is expressed in each neuron. We combine the holographic illumination with spiral scan approach to minimize the laser power onto the brain.

Setup of the Two-Photon Imaging Path

  • (1.1) The imaging laser is typically a wavelength tunable Ti:Sapphire laser with a repetition rate of 80 MHz, and a pulse width of 70~140 fs. For GCaMP imaging, the wavelength can be set to 920~940 nm. A Pockels cell is set up after the laser to modulate the laser power. An optional pulse compressor could be set up after the Pockels cell to optimize the pulse width at the sample.

  • (1.2) An optional electrically tunable lens could be set up as an axial focal control device. A set of XY galvanometer mirrors, or a resonant scanner paired with a galvanometer mirror, can be set up to raster scan the laser focal spot on the sample. The electrically tunable lens and the scanning mirrors are located at the conjugate plane of the back focal plane of the objective lens. 4f systems are used to relay these planes and to magnify the beam size so that the beam can illuminate a large portion of the aperture of the electrically tunable lens and scanning mirrors.

  • (1.3) A dichroic mirror is set up to combine the imaging path and photostimulation path and direct the beam to scan lens and tube lens, which forms a 4f system to magnify the beam at the back focal plane of the objective lens. This beam size should be optimized for specific objective lens so as to achieve a good excitation numerical aperture (NA) for cellular resolution. For population imaging, the objective lens is typically chosen to be 16~40×, and the excitation NA is typically 0.45~0.6. The collection NA should be as large as possible so as much as the emitted fluorescence could be collected.

  • (1.4) A dichroic mirror is set up before the objective lens to transmit the infrared excitation light and reflect the fluorescence emission to photomultiplier tubes (PMTs). A dichroic mirror can further separate the fluorescence spectrum into different channels (typically green and red).

Setup of the Two-Photon Photostimulation Path

  • (2.1) The photostimulation laser can be a high repetition rate fiber laser or a low repetition rate pulse-amplified laser (including optical parametric amplified laser). The latter laser is preferred for holographic photostimulation as its higher pulse energy allows lowering the average photostimulation power per cell, thus enabling photostimulating a large number of neurons with the overall available power budget. For red-shifted opsin such as C1V1, the preferred laser wavelength is 1040~1080 nm. Similar as the imaging path, a Pockels cell and an optional pulse compressor are set up after the laser.

  • (2.2) A spatial light modulator is set up at a conjugate plane to the back focal plane of the objective lens. Following the Pockels cell, a half-wave plate is set up to rotate the polarization of the laser so the light polarization aligns with the active axis of the spatial light modulator. A 4f system is used to magnify the beam size so it fills up the active region of the spatial light modulator. A 4f system is placed after the spatial light modulator with a zeroth-order beam block set up at the intermediate imaging plane. After this 4f system, a set of XY galvanometer mirrors are set up, which is also at the conjugate plane of the back focal plane of the objective lens. Here, the XY galvanometer can not only perform spiral scanning but also provide a lateral offset to the centroid of the SLM’s addressable field of view, thus effectively enlarging the overall field of view, as discussed in earlier sessions. Thus, the XY galvanometer is preferred over galvo/resonant scanner pair, as the latter cannot provide a static beam shift in the resonant scanner’s scanning direction.

  • (2.3) The photostimulation path then combines with the imaging path at the dichroic mirror and is then directed to the sample through the scan lens, tube lens, and objective lens.

Control Electronics and Software

  • (3.1) The electronic control system typically includes analog inputs/outputs to control the Pockels cells, electrically tunable lens, spatial light modulator, and scanners in both beam paths, as well as high-speed digitizers to record the imaging data from the PMT. For home-built microscopes, ScanImage [79] is one commonly used control software.

  • (3.2) The phase pattern on the spatial light modulator can be calculated through a superposition algorithm or a Gerchberg-Saxton algorithm as detailed in Subheading 2.2.2 so a user-defined excitation pattern can be projected to the sample. An open source control software can be found in [80].

Coordinate Calibration

  • (4.1) It is critical to register the photostimulation beam’s target coordinate with the imaging laser’s image coordinate. An autofluorescent plastic slide can serve as the sample, and a 2D holographic pattern can be projected to burn spots on the surface of the autofluorescent plastic slide. The imaging laser can then visualize the burned spots. This can also calibrate the axial focus offset between the two beam paths. An affine transformation can be extracted to map the coordinates between the hologram generation algorithm and the actual imaging system. This can be repeated for a few defocusing depths set in the spatial light modulator, and a linear interpolation of the mapping can be applied for the depths in between.

  • (4.2) Due to the chromatic dispersion and finite pixel size of SLM, the SLM’s beam steering efficiency, also called diffraction efficiency, drops with larger angle, leading to a lower beam power for targets further away from the center field of view (in xy), and nominal focus (in z). This efficiency drop can be analytically calculated [55, 75] or experimentally measured by raster scanning the photostimulation beam on the autofluorescence slide and detecting the fluorescence strength. A linear compensation can be applied in the weighting coefficient among different points in the target pattern to counteract this non-uniformity (see Note 3).

Optogenetics Experiment

  • (5.1) Before each set of experiments on animals, it is a good practice to verify the system (laser power, targeting accuracy, power uniformity among different beamlets from the hologram) by generating groups of random spots through holograms, burning the spots on an autofluorescent plastic slide, and comparing the resultant image with the desired target.

  • (5.2) The calcium imaging is first performed followed by the extraction of the spatial footprint and temporal dynamics of each recorded neuron. To confirm the co-expression of calcium indicators and opsins in an individual neuron, one can image the fluorophore linked to the opsin. A desired group of neurons can then be selected for optogenetics experiments. Laser power and stimulation time should be adjusted to produce responses observed physiologically (see Note 4).

  • (5.3) While the excitation wavelengths between the calcium indicators and opsins are different, the calcium indicator may still generate fluorescence background during photostimulation, particularly when a large number of neurons are photostimulated simultaneously. Imaging processing should be implemented to remove or suppress this background.

3 Identification and Targeting of Neuronal Ensembles Related to Behavior

3.1 Background

The implementation of simultaneous two-photon imaging [11] and two-photon optogenetics [12] allowed the recording and manipulation of tens of cells simultaneously [14, 66, 69]. However, the identification of the neurons that could have a robust impact on the overall population activity also represents a critical step in the design of experiments aiming to control learned behaviors [4, 21, 81].

3.1.1 Multidimensional Reduction Techniques Applied to Population Recordings

It has been demonstrated that the activity of neuronal ensembles could be defined as an array of multidimensional population vectors, where the dimensionality of the array is given by all recorded cells in the field of view [5, 8, 82]. Similar population vectors could be visualized by diverse computational techniques that define clusters in a reduced dimensional space, where each cluster depicts a neuronal ensemble [5, 83]. A neuronal ensemble represents a group of neurons with coordinated activity that repeats at different points in time [10]. The characterization of brain states using multidimensional population vectors is independent of the recording length and could be implemented in chronic experiments where the activity of identified neuronal ensembles could be compared at different days.

3.1.2 Targeting Visualized Neuronal Populations with Two-Photon Optogenetics

Two experimental designs are possible to control behavior targeting visualized neuronal populations (see Note 5). One solution is to target all available neurons that could respond to a given behavioral cue with single-cell precision [14, 16, 18]. The other solution is to target neurons with pattern completion capabilities that could recall physiologically relevant neuronal ensembles related to behavior [4, 6, 10, 19, 21].

In cortical circuits, the synchronous activation of several neurons could simultaneously result in two unwanted scenarios: (i) the generation of epileptiform activity or (ii) the forced engagement of GABAergic circuits. In both scenarios, the physiological recalling of a specialized neuronal ensemble is compromised (see Note 5).

On the other hand, the identification and targeting of neurons with pattern completion capabilities could keep the balance between excitation and inhibition inherent to the microcircuit under study, allowing the recalling of physiological neuronal ensembles that work as attractors [84] evoking a given behavior [84, 85].

3.2 Implementation of Analytical Methods to Recall Neuronal Ensembles Relevant to a Learned Behavior

In this part of the chapter, we describe the procedures and concepts to analyze population activity extracted from calcium imaging recordings. We will focus on the steps after individual neurons have been identified from imaging recordings and their activity has been inferred from calcium transients.

3.2.1 Motion Correction, Identification of Neurons, and Spike Extraction

Simultaneous population recordings allow the characterization of population activity with single-cell resolution but generate large datasets representing an analytical challenge. A common practice used for analysis of simultaneous population recordings is to create a binary representation of neuronal activity where 1’s indicate firing and 0’s indicate silent periods [82, 86]. Recently, several independent research groups have released open-source code to preprocess calcium imaging recordings to extract activity information from a series of images [87,88,89,90].

There are four main steps that need to be considered before performing population analysis on binary arrays (Fig. 7):

  1. (i)

    Motion correction: Lateral displacements are commonly corrected using TurboReg [91], fast Fourier transform [92] or field approaches [93]. Translations in depth should be avoided or the data should be discarded.

  2. (ii)

    Identification of neurons: Automatic identification of regions of interest (ROI’s) can be performed using different approaches [87, 88].

  3. (iii)

    Determination of changes in fluorescence: After the identification of ROI’s, fluorescence traces for each neuron should be extracted.

  4. (iv)

    Activity inference from calcium transients: The first time derivative, deconvolution, or fitting could be used to infer spiking activity from calcium transients [82, 87, 88]. Note that inferring co-activity between pairs of neurons from raw calcium transients introduces an artifact from the decaying phase of the changes in fluorescence. Such common mistake generates spurious correlations between neurons as demonstrated before [7].

Fig. 7
Four illustrations labeled a to d. A is a photon microscopy photo titled simultaneous dual plane imaging, b is titled extracted R O I contour with dense circular structures, c is a line graph with high and low peaks titled calcium transit, d is a line graph with high and low peaks titled event detection.

Spike inference from holographic calcium imaging recordings. (a) Representative experiment where two focal planes were recorded simultaneously using holographic two-photon microscopy. (b) Regions of interest (ROIs) detected from the recordings in a. Each ROI corresponds to an identified neuron. (c) Representative changes in fluorescence obtained from ROIs shown in b. (d) Binary arrays representing the activity of one neuron obtained from inferred spikes. (Modified from Ref. [5, 10])

Full size image

3.2.2 Binary Arrays from Inferred Activity

To construct binary arrays, inferred activity should be threshold usually 3 S.D. above noise level (Fig. 7d). This procedure has been shown to match bursting activity of neurons above 2 action potentials [82]. The onsets from inferred activity can be used to represent the overall population activity. A binary matrix [N × T] is constructed, where N denotes the total number of neurons in the field of view and T represents the total number of time points recorded. Each row in the binary matrix depicts the activity of one neuron and each column in the binary matrix represents the activity profile of a neuronal population (Fig. 8a). To visualize the overall network activity, the binary matrix can be plotted as a raster plot where 1’s are dots. The population synchronicity can be extracted from the time histogram of the raster plot [5, 8].

Fig. 8
Two illustrations of population vectors labeled a and b. A has time frames titled T 1, T 2, up to T n, on the left and a population activity matrix on the right. A row titled neuron and a column titled population vector are highlighted. B is an n-dimension vector diagram with angle theta between a and b population vectors, to its right is similarity index.

Neuronal ensembles represented as multidimensional population vectors. (a) Schematic representation of different frames where active neurons are shown in black-filled circles (left). Binary matrix illustrating the overall population activity, where each row is a neuron and each column denotes a population vector. The total number of recorded neurons gives the dimensionality of the array. (b) Neuronal ensembles representing recurrent groups of neurons that are active at different times can be understood as clusters of similar population vectors in a multidimensional space. The cosine similarity could be used as a metric to calculate the angle between population vectors. If the angle is close to 0°, the population vectors are similar, therefore almost the same neurons fired at different times. If the angle is close to 90°, the population vectors are different. (Reprinted from Ref. [10])

Full size image

3.2.3 Multidimensional Population Vectors Defining Neuronal Ensembles

To identify groups of neurons with coordinated activity, population vectors from a physiologically meaningful time window are constructed; time windows from 100–500 ms have been shown to reflect physiological functions and generate similar results. Population vectors capture the coordinated activation of specific groups of neurons. Once the population vectors are defined, comparing the distribution of coordinated activity that could appear randomly against the observed coordinated activity can show population vectors above chance levels [5, 7]. Only population vectors with more active cells than the ones expected by chance should be considered for further analysis. Significant population vectors capture network activity defining a multidimensional space in which the number of dimensions is dictated by the total number of identified neurons with coordinated activity above chance levels [4,5,6, 8, 10]. It is important to highlight that to characterize population activity, each dot in a multidimensional space should be a population vector instead of an individual neuron (Fig. 8b).

3.2.4 Similarity Measurements on Population Vectors

It has been shown that the use of significant population vectors can be used to discriminate similar patterns of activity repeated at different times [4,5,6, 8]. The representation of network activity as population vectors allows an exhaustive comparison of all population vectors to visualize different experimental conditions. Similarity maps of population vectors represent a valuable tool to visualize recurrent activity of neuronal ensembles. To construct similarity maps, all possible combinations of vector pairs need to be computed. Since population vectors representing the activity of a given group of neurons point to a similar place in a multidimensional space, the angle between population vectors could be used to create a similarity map [8, 82]. The cosine of the angle between a pair of population vectors is defined by their normalized dot product: cos(θ)=V1 • V2 / llV1ll llV2ll. Thus, if the angle is close to zero, the vectors are similar whereas if the vectors are different, they tend to be orthogonal (Fig. 8b).

3.2.5 Identification of Neuronal Ensembles from Population Vectors

Similarity maps constructed from all possible vector pairs could be understood as a low-dimensional representation of the original network activity, where the angles between all population vectors as a function of time are highlighted. Thus, high similarity values in the same row denote recurrent groups of neurons firing together. A cluster of population vectors pointing in a similar direction gives the definition of a neuronal ensemble using the population vector approach. Similarity maps can be transformed to a binary matrix S of size [T × T] (Fig. 9a, b). The factorization of S using singular value decomposition (SVD) factorizes S as the multiplication of three factors: V, ∑, and VT, where V and VT are orthonormal basis and ∑ contains the singular values [5, 8]. To detect neuronal ensembles from recurrent patterns of activity observed in the matrix S, the rate of singular values decay is determined (Fig. 9c). Taking the singular values above chance levels can reproduce the original matrix S with high accuracy (Fig. 9d). Then each factor of the SVD decomposition represents a linearly orthogonal component and each factor defines a neuronal ensemble (Fig. 9e). In the case of primary visual cortex, each neuronal ensemble represents a different orientation of drifting-gratings, a different natural scene or the Go signal from a visually guided Go/No-Go task [4, 5, 7]. Another approach to identify neuronal ensembles is the projection of the multidimensional population vectors into a low dimensional space, in such reduced dimensional space clusters of population vectors depicted with cluster analysis define neuronal ensembles or network states (Fig. 8b) [82]. The advantage of using SVD factorization is that the total number of neuronal ensembles could be systematically detected from the magnitude of the singular values. Thus, recurrent population vectors can be assigned to a given neuronal ensemble and population vectors with low repetition rate are excluded.

Fig. 9
Five illustrations of population vectors using S V D labeled a to e. A is a monochrome map with patterns all over it. B is a matrix with similar patterns in a. C is a line graph that plots magnitude versus singular value, d is a matrix with similar patterns in a. E is a set of 5 visual stimuli with patterns on it.

Neuronal ensembles defined by similarity maps of population vectors using singular value decomposition (SVD). (a) Similarity map illustrating the angles between all possible pair combinations of population vectors. Note that the recurrent activation of a given group of neurons is visualized as increased similarity index in the same row. An angle between two vectors close to 0° represents a similarity value close to 1, whereas an angle between two vectors close to 90° represents a similarity value close to 0. (b) Binary matrix computed from the similarity map in a, representing significant patterns of activity factorized by SVD. Black patterns in the same row indicate recurrent coactive neurons at different times. (c) Magnitude of singular values used to define the number of neuronal ensembles repeated above chance levels. Red line indicates the double of singular values from shuffled data. The cutoff indicates the number of neuronal ensembles that account for >90% of the data. In general microcircuits of ~100 neurons stimulated with four drifting-gratings can be defined by ~6 neuronal ensembles. (d) Binary matrix of the first 6 factors obtained from SVD of b. (e) Factors from SVD that reproduce the overall response of the imaged focal plane to visual stimuli. Bars on top indicate orientations of drifting-gratings. Empty squares represent spontaneously active neuronal ensembles that appeared in the absence of visual stimuli. (Modified from Ref. [5])

Full size image

3.2.6 Pattern Completion Properties of Neuronal Ensembles

In the neuronal ensemble framework, pattern completion refers to the ability to recall a whole neuronal ensemble by the activation of few neurons with strong functional connectivity [6]. It has been recently shown that the stimulation of the same neuronal population for several times (~100) was able to imprint an artificial neuronal ensemble in layer 2/3 of primary visual cortex of awake head fixed mice [6]. Imprinted ensembles were composed by neurons that had low probability to fire together, but after the imprinting protocol, the stimulated neurons fired spontaneously in the absence of stimuli (Fig. 10a). The mechanism governing the creation of artificial neuronal ensembles could be explained by Hebbian synaptic plasticity [94], where the connectivity between neurons firing together was strengthened (Fig. 10b). Indeed, neurons responding to a given drifting-grating or belonging to a neuronal ensemble have increased probability to be connected [8, 95]. Imprinted ensembles could also be recalled by the stimulation of individual neurons with pattern completion capabilities (Fig. 10), suggesting that pattern completion could be a general property of different brain areas [21].

Fig. 10
Two illustrations of artificial neuronal ensembles labeled a and b. A is a set of 3 spatial maps titled ensemble training, imprinted ensemble, and pattern completion, with dense circular structures. B has natural ensemble on the left and recalled ensemble on the right.

Imprinting and recalling of artificial neuronal ensembles in layer 2/3 of primary visual cortex. (a) Spatial map of neurons activated with two-photon optogenetics several times (~100). Red neurons represent repeatedly responding neurons (left). Scale bar: 50 μm. Imprinted ensemble (red neurons) shows spontaneous activity in the absence of any stimuli (middle). The activation of one neuron with pattern completion capability (blue arrow) was able to recall the imprinted ensemble (red neurons). (b) Cartoon illustrating how imprinted and recalled ensembles could be explained by the strengthening of connections of pre-existing ensembles and photostimulated neurons. (Modified from Ref. [10])

Full size image

3.2.7 Recalling Neuronal Ensembles Related to Behavior

To prove the causal relation between neuronal activity and a learned behavior, it is necessary to identify representative members of neuronal ensembles that could trigger the behavior [81]. During the behavioral training phase photostimulation should be avoided (see Note 6). It has been recently shown, in a visually guided Go/No-Go task, that neuronal ensembles responding to drifting-gratings in layer 2/3 of primary visual cortex increase their reliability for the Go signal whereas reduce their reliability for the No-Go signal [4]. Increased reliability is reflected as strengthened functional connectivity of the Go ensemble allowing neurons with pattern completion capabilities [21] to recall the whole Go ensemble and trigger the perception of the Go signal evoking the learned behavior (Fig. 11).

Fig. 11
Two illustrations of go ensembles labeled a and b. The components of a are go signal visual stimuli, go ensemble layer 2 over 3 V 1, and licking. The components of b are P C neurons opto stimuli, recall ensemble layer 2 over 3 V 1, and licking.

Controlling behavior using pattern completion properties of Go ensembles. (a) In mice trained in a visually guided Go/No-Go task, the Go signal activates a neuronal ensemble related to the correct execution of the task (green neurons). The activation of the Go ensemble produced licking. (b) Activation of as few as two pattern completion neurons (red neurons) belonging to the Go ensemble is able to recall the whole Go ensemble and produce licking even in the absence of visual stimuli or behavioral cues. (Reprinted from Ref. [4])

Full size image

4 Considerations for the Implementation of a Visually Guided Go/No-Go Task

  • Training and experiments should be performed at the same hour of the day.

  • That same person should handle the animals.

  • Training should be performed every day, avoiding skipping training sessions on weekends or holidays.

  • The weight of the animals should be systematically documented and maintained at the same level overnight. In general, 1 mL of water should increase weight by 1 g.

  • Changes in weight ±2% of the total mass could significantly alter behavioral performance.

  • To avoid continuous licking, plain water should be preferred over sweetened water.

  • The lick port should be placed 1.5 mm from the mouth. A lick port too close to the mouth could cause increased non-task related licking.

5 Notes

  1. 1.

    A proper calcium indicator and opsin pair should be chosen to minimize crosstalk. It is also critical to control imaging laser power so that neurons are not depolarized by the imaging laser during imaging [Subheading 2.2.1].

  2. 2.

    The zeroth-order beam block in the photostimulation path should be positioned at the intermediate plane of the 4f system, right after the spatial light modulator. While it can block the zeroth-order beam, it also makes the neurons in the center field of view inaccessible to the photostimulation beam. Thus, the zeroth-order beam block should have an intermediate size: if too small, the zeroth-order beam cannot be totally blocked, but if too large, there could be a large number of neurons in the center field of view not accessible by the photostimulation laser [Subheading 2.2.2, Fig. 3a].

  3. 3.

    As the opsin expression level may differ between different neurons, the weight factors for different target neurons should be adjusted in the hologram calculation [55, 75] so that their calcium transients evoked by two-photon optogenetics have a similar response in amplitude [see Subheading 2.2.4. (4.2)].

  4. 4.

    The amplitude of calcium transients in individual neurons evoked by two-photon optogenetic stimulation should be similar to the amplitude of calcium transients evoked by physiological stimuli, otherwise the effects described could be due to over stimulation of neurons [see Subheading 2.2.4. (5.2)].

  5. 5.

    The responses induced by the activation of large numbers of neurons in a given optical field could be attenuated by the engagement of GABAergic circuits. The total number of simultaneously photostimulated neurons in a given optical field should be maintained in the same range as the number of active neurons evoked by physiological conditions (sensory stimuli) [see Subheading 3.1.2].

  6. 6.

    During the training phase of the behavioral task, pairing of sensory stimuli and optogenetic stimuli should be avoided. Otherwise, the effects attributed to optogenetic stimuli could be due to mice learning to identify the optogenetic activation of neurons [see Subheading 3.2.7].

6 Outlook

A general development direction of all-optical methods is to increase the throughput (i.e., field of view and spatiotemporal resolution) of both imaging and photostimulation, while reducing the laser power incident onto the brain. Beam multiplexing techniques are promising approaches to increase the imaging throughput [56,57,58,59]. Larger spatial light modulators with higher pixel count could increase the beam diffraction efficiency and thus the field of view [14]. Faster spatial light modulators could allow rapid switching of the hologram and thus could enable faster photostimulation of neuronal ensembles in specific spatiotemporal patterns. As the simultaneously photostimulated neurons across a 3D volume increase [14, 66, 69, 72], there could be off-target effects through photostimulating the dendritic arbors crossing the cell bodies of target neurons. Somatic restricted viruses where the opsin only expresses in the cell body could alleviate this issue [14, 66, 72, 96].

To access deeper brain regions noninvasively, adaptive optics [97, 98] and three-photon excitation [38, 39, 41, 99, 100] could be used. A spatial light modulator in the photostimulation path could implement adaptive optics, and a spatial light modulator in the imaging path could allow for similar corrections. Three-photon calcium imaging has also been successfully demonstrated [39, 41, 99, 100]. One current challenge is that it needs high power for three-photon optogenetics in deep brain regions. The development of high-efficiency opsins optimized for three-photon excitation could overcome this challenge.

The development of all-optical techniques to simultaneously read and write patterned activity in neuronal populations could help the understanding of perception, memory formation, behavioral states, and pathological conditions with single-cell resolution in the next decades. The systematic documentation of the causality between neuronal ensemble activity and brain states requires further development of analytical tools to visualize and target specific neurons that can control the overall network activity [81]. The adaptation of artificial intelligence algorithms used in big data to understand population activity could be used to design and guide in vivo experiments [13].

Scaling of these techniques to several brain areas to measure and control thousands of neurons in different parts of the brain will require the identification and targeting of neurons with pattern completion capability [21, 81]. This approach could control thousands of neurons by the activation of a small percentage of them, reducing sample heating and deterioration of spatial resolution that comes with increased number of parallel stimulations.

Finally, the development of transgenic mice with genetically encoded calcium or voltage indicators and opsins that use different wavelengths in molecularly identified subpopulations of neurons [20] will allow chronic recordings and stimulation of several brain areas with single cell and molecular identity precision, which represent the next challenge for neuronal ensemble research.