Large-scale voltage imaging in the brain using targeted illumination

Recent improvements in genetically encoded voltage indicators enabled optical imaging of action potentials and subthreshold membrane voltage dynamics from single neurons in the mammalian brain. To perform high speed voltage imaging, widefield microscopy remains an essential tool for recording activity from many neurons simultaneously over a large anatomical area. However, the lack of optical sectioning makes widefield microscopy more prone to background signal contamination, and thus far voltage imaging using fully genetically encoded voltage indicators remains limited to simultaneous sampling of a few cells over a restricted field-of-view. We here demonstrate a strategy for large scale voltage imaging using the fully genetically encoded voltage indicator SomArchon and targeted illumination. We implemented a simple, low-cost digital micromirror device based targeted illumination strategy to restrict illumination to the cells of interest, and systematically quantified the improvement of this microscopy design theoretically and experimentally with SomArchon expressing neurons in single layer cell cultures and in the brains of awake mice. We found that targeted illumination, in comparison to widefield illumination, increased SomArchon signal contrast and reduced background cross-contamination in the brain. Such improvement permitted the reduction of illumination intensity, and thus reduced fluorescence photobleaching and prolonged imaging duration. When coupled with a high-speed, large area sCMOS camera, we routinely imaged tens of spiking neurons simultaneously over minutes in the brain. Thus, the widefield microscopy design with an integrated targeted illumination system described here offers a simple solution for voltage imaging analysis of large neuron populations in behaving animals.


Introduction 37
Recent advances in genetically encoded voltage indicators (GEVIs) have enabled neuroscientists to directly 38 measure membrane voltage from individual neurons in the mammalian brain 1-6 . In particular, a few recent GEVIs, 39 including SomArchon, QuasAr3, Voltron, ASAP3 and Ace2N have achieved sufficient sensitivity to capture 40 individual action potentials from single neurons recorded from behaving mice. Of these high performance GEVIs,41 several are fully genetically encoded, whereas others are hybrid sensors that require exogenous chemicals 7-13 . 42 One class of fully genetically encoded indicators detects voltage dependent fluorescence of fluorophores fused to 43 voltage sensitive peptide domains derived from voltage gated ion channels, voltage sensitive phosphatases or 44 rhodopsins 7-12 . For these GEVI designs, changes in cell membrane voltage induce confirmational transitions of the 45 voltage sensitive domains, which subsequently alter the intensity or the efficiency of Forster resonance energy 46 transfer of the tethered fluorophores. A recent example is ASAP3 that measures voltage dependent fluorescence 47 of a circular permutated GFP fused to the voltage sensing domain of G. gallus voltage-sensing phosphatase 4 . 48 Another class of fully genetically encoded indicators are single compartment and directly detect the intrinsic 49 voltage dependent fluorescence of engineered rhodopsins, such as QuasAr3, Archon and SomArchon 3,5,14 . To 50 improve fluorescence signals, bright chemical fluorophores have also been explored in the designs of GVEIs, 51 yielding a class of high-performance hybrid GEVIs that requires both exogeneous chemical dyes and the 52 corresponding voltage sensing protein counterparts 2,13,15 . 53 54 With rapid and continued improvements of GEVIs, voltage imaging offers great promise for direct analysis of 55 neuronal voltage dynamics in the brain. To capture fast membrane voltage fluctuations, especially action 56 potentials that occur on the millisecond and sub-millisecond time scale, fluorescence voltage imaging needs to be 57 performed at a near kilohertz sampling speed. Point scanning techniques, such as multiphoton microscopy, have 58 minimum signal cross-contamination and out-of-focus background due to confined excitation volumes 16 , but are 59 generally limited to video-rate acquisition speed as they rely on mechanical scanners. Fast random access scanning 60 using acousto-optic deflectors has been demonstrated with kilohertz sampling rates 4 , although these devices 61 require a complicated setup, are sensitive to motion artifacts, and more importantly, can only record very few 62 pre-selected cells at once. More recently, kilohertz frame rate two-photon imaging over a field-of-view (FOV) of 63 50 × 250 µm 2 has been demonstrated by means of passive pulse splitting from a specialized low-repetition rate 64 laser 17 . However, its stringent alignment requirements, high cost, and concerns regarding long-term system 65 stability remain a major obstacle for its widespread use for neuroscience studies. 66 67 Alternatively, widefield microscopy, especially when equipped with the newly developed high-speed large-area 68 sCMOS cameras, remains a cost-effective and easily implementable solution for wide FOV, kilohertz frame-rate 69 imaging. This ability to image a large FOV at high spatiotemporal resolution is particularly critical to resolving 70 morphological details of individual neurons, and to correct for tissue movement associated with physiological 71 processes (i.e., heart rate, breathing) that are unavoidable when imaging the brains of awake behaving animals. 72 However, a major limitation of widefield microscopy is the inability to reject out-of-focus and scattered light 18 , 73 making it prone to signal contamination and background shot noise caused by non-specific excitations 19 . To 74 address this problem molecularly, recently developed GEVIs have utilized soma targeting peptides to restrict the 75 expression of GEVIs to the soma or the proximal dendrites. For example, SomArchon, QuasAr3, Voltron, and 76 ASAP3-kv all include the axon initial segment targeting motif of the potassium channel Kv2.1 20 , which are critical 77 for their success in the brains of behaving animals 2-6 . Restricting the expression of GEVIs to a sparse subset of 78 neurons can also help reduce background signal contamination in vivo, and this strategy was recently used to 79 achieve simultaneous imaging of tens of neurons using the hybrid sensor Voltron 2 . 80 In parallel with the molecular targeting of GEVIs, targeted illumination has also been developed to enhance image 81 contrast and signal-to-noise ratio (SNR) 5,6,21 . By using a digital micromirror device (DMD) or a spatial light 82 modulator to pattern the illumination light, targeted illumination confines fluorescent excitation to pre-selected  83  areas of interest based on reference images previously obtained from the same sample. It has been implemented  84  in extended-depth-of-field microscopy for video rate calcium imaging 21 , and was recently demonstrated to  85 significantly improve voltage imaging performance with a widefield microscope 5,6 . However, for voltage imaging 86 applications, this approach thus far has been limited to simultaneous sampling of a few cell, and requires 87 sophisticated microscopy setups. 88 We integrated a simple, low cost DMD-based targeted illumination module into a standard widefield microscope, 89 and directly compared SomArchon imaging performance of the same neurons under targeted versus widefield 90 illumination, in both cultured neuron preparations, and from the brains of awake behaving mice. We found that 91 illumination targeting reduced nonspecific background fluorescence and fluorescence signal cross-contamination, 92 leading to increased SomArchon spike signal-to-background ratio. The improvement of SomArchon fluorescence 93 contrast allowed us to decrease the total excitation power over the imaging FOV with reduced fluorescence decay. 94 As a result, we were able to perform routine SomArchon voltage imaging from tens of neurons over a large 95 anatomical area of 360 × 180 µm 2 while maintaining SomArchon fluorescence contrast, and over a prolonged 96 recording duration of several continuous minutes. These results demonstrate that targeted illumination with a 97 DMD represents a simple, low cost, and practical strategy for large scale voltage imaging of tens of neurons over 98 an extended period of time in awake behaving animals. 99

Results 101
Modeling and testing the effects of targeted illumination on optical crosstalk in widefield optical imaging 102 Motivated by the unique advantage of widefield microscopy in performing optical voltage imaging with high 103 spatiotemporal resolution over large FOVs, we considered a targeted illumination approach to further enhance 104 signal quality by reducing out-of-focus background signals. We first developed a theoretical model to estimate 105 how targeted illumination might minimize signal crosstalk due to out-of-focus excitation or tissue scattering from 106 nearby neurons that are not actively being imaged. We considered contributions from both out-of-focus 107 fluorescence and tissue scattering, by modeling light propagation through scattering media using the radiative 108 transfer equation in the forward scattering limit 22 (see Methods and Fig. S1). In our model, we characterized 109 crosstalk values from non-targeted neurons at distance laterally and axially from a region of interest (ROI) 110 under widefield versus targeted illumination conditions (Fig. 1a, b, and Fig. S3). We found that in simulated 111 fluorescence images, targeting illumination to a specific neuron substantially reduced the overall background in 112 the imaging plane, and therefore reduced the strength of crosstalk from neighboring neurons (red circle: Fig. 1d, 113 e). Additionally, the level of crosstalk contamination from a non-overlapping axially displaced neuron is strongly 114 affected by the distance between the out-of-focus neurons relative to the imaging plane (Fig. 1c). These findings 115 confirm that, for widefield microscopy, targeting illumination to a neuron of interest can improve signal quality 116 by reducing the overall fluorescence background, and limiting signal contamination from neighboring out-of-focus 117 neurons. These computational results highlight that targeted illumination is a viable approach for low-background, 118 high-contrast imaging of voltage signals in the brain using widefield microscopy. 119 120 To experimentally evaluate the improvement of targeted illumination, we integrated a DMD into a custom-built 121 widefield microscope configured for dual-color GFP and SomArchon imaging (Fig. 1h). We performed voltage 122 imaging of SomArchon expressing neurons, in both cell cultures and in the visual cortex and the hippocampus of 123 awake head fixed mice. Since SomArchon protein is fused to the GFP reporter, static GFP fluorescence images 124 were first taken to identify SomArchon expressing neuronal soma. The GFP fluorescence images were then used 125 to generate references for targeting illumination to the identified SomArchon positive neurons. Consistent with 126 what was observed in our computational models, we found that restricting illumination to the soma reduced the 127 overall background fluorescence and accordingly enhanced the contrast of SomArchon fluorescence in individual 128 cells (Fig. 1f, g). 129 action potentials. Since SomArchon can detect subthreshold voltage dynamics 3 , the actual photon shot noise is 156 mixed with real biological subthreshold voltage fluctuations. Therefore, it is difficult to quantify the actual noise 157 level and therefore accurately evaluate SomArchon signal qualities by calculating the absolute SNR. We thus 158 calculated the spike signal-to-baseline ratio (SBR), defined as the amplitude of the spike divided by the variance 159 of the experimentally measured baseline fluctuations (Vm), as an estimated performance metric of SomArchon in 160 recording individual spikes. The estimation of Vm variation however is affected by the presence of suprathreshold 161 spike events. We therefore developed a spike-insensitive SBR estimation method that does not require prior 162 identification of spikes (see Methods). To test the properties of the SBR algorithm, we simulated membrane 163 voltage using the Izhikevich-type neuron model that exhibit both action potentials and biological subthreshold 164 voltage (Fig. S4a). Additionally, to model experimentally measured noise signals, we added different levels of 165 Gaussian white noise (Fig. S4b, d, e). We calculated the theoretical SBR as spike amplitude divided by the 166 experimentally measurable Vm that contains both the biological subthreshold voltage and the white noise. We 167 further calculated the theoretical SNR as the spike amplitude divided by the variance of the white noise only. We 168 found that our SBR estimation substantially underestimated the theoretical SNR, but it better reflected the 169 theoretic SBR that considers biological Vm variation (Fig. S4c). Thus, though the spike SBR measure is an 170 underestimation of SomArchon molecular performance, it provides an intuitive and spike-insensitive measure of 171 the optical voltage signal quality, especially for in vivo recordings where conditions can vary substantially. 172

173
With targeted illumination, we detected a spike SBR of 5.7 ± 2.0 (mean ± standard deviation, from 226 neurons in 174 16 FOVs), significantly greater than that observed from the same neurons under the widefield illumination 175 condition (4.9 ± 1.4, Fig. 2h). Since the spike identification algorithm relies on a custom spike SBR threshold, we 176 investigated whether the increase in spike SBRs depends on the threshold used to identify spikes. We found that 177 across several chosen SBR threshold values, targeted illumination consistently resulted in greater spike SBRs than 178 widefield illumination (Fig. S5, Table S3). Furthermore, targeted illumination resulted in more detected spikes than 179 that detected from the same neurons measured in the widefield condition (Fig. 2i). 180 181 We next examined fluorescence decay, calculated as the percent reduction of fluorescence intensity over time, 182 an important parameter that limits the duration of fluorescence imaging in general. We found that with targeted 183 illumination, SomArchon showed a slight fluorescence decay of 2.15 ± 2.66% (mean ± standard deviation, n = 226 184 neurons) over a 20 second recording period, significantly smaller than that observed under widefield illumination 185 (2.99 ± 1.04%, Fig. 2j). Together, these results demonstrate that targeted illumination significantly improves 186 SomArchon performance in terms of spike SBR and fluorescence decay, even in cultured neurons where out-of-187 focus background is minimal.  both Vm and spikes. We found that both Vm-Vm and spike-spike correlation decreased slightly with increasing 211 anatomical distance between simultaneously recorded neuron pairs (slopes for linear regression between Vm-Vm 212 correlation and distance are -5.3e -4 and -3.7e -4 for targeted illumination and widefield illumination respectively, 213 and between spike-spike correlation and distance are -1.7e -4 and -7.2e -5 respectively, Fig. 3a,b). The regression 214 slopes of Vm-Vm correlation and spike-spike correlation over anatomical distance under the targeted illumination 215 condition are both greater than that observed during the widefield illumination condition (Vm-Vm correlation, p 216 = 5.3592e -6 , z score = -4.5502, permutation test; spike-spike correlation, p = 0.0039, z score = -2.8865). 217

218
To further evaluate changes in Vm-Vm and spike-spike correlation across different anatomical distances between 219 the two illumination conditions, we binned the correlation values of neuron pairs every 30 µm. Consistent with 220 the improvement of spike SBR observed under the targeted illumination condition, spike-spike correlation was 221 slightly greater under targeted illumination than widefield illumination condition across neuron pairs within 180 222 µm, although no difference was observed for neuron pairs over 180 µm (Fig. 3d). When we examined Vm-Vm 223 correlation, we found no difference between targeted illumination and widefield illumination conditions for 224 neurons pairs within 120 µm, though a slightly smaller correlation value was obtained under targeted illumination 225 for neurons over 120 µm away (Fig. 3c). The similar Vm-Vm correlations under widefield and targeted illumination 226 is consistent with our numerical models when the sample is only a monolayer of cells absent of significant 227 contributions from out-of-focus and scattered fluorescence (Fig. S2). 228 229 To quantify the effect of targeted illumination in the brains of awake animals, we examined SomArchon expressing 244 neurons in the superficial layers of visual cortex. Mice were head-fixed and able to freely locomote on a spherical 245 treadmill. For each FOV, we alternated 10-second long voltage imaging sessions between targeted illumination 246 and widefield illumination conditions. We found that targeted illumination significantly reduced the decay of 247

(a,b) Pearson's correlation values between pairs of simultaneously recorded neurons decreased over anatomical
SomArchon fluorescence (11.02 ± 3.08% over 10 seconds), approximate half of that observed with widefield 248 illumination (20.38 ± 3.05%, p = 4.74 -14 , paired t-test, df = 20, Fig. 4k). Targeted illumination also resulted in a 249 significant increase in SomArchon spike SBR, achieving 4.6 ± 0.7, significantly higher than the 4.1 ± 0.44 obtained 250 with widefield illumination (p = 0.023, paired t-test, df = 18 neurons, Fig. 4l), and similar to that observed in 251 cultured neurons (Fig. 2h). This significant increase in spike SBR for the targeted illumination condition accordingly 252 led to a greater frequency of spikes identified when spike SBR threshold is used for spike identification (Fig. 4m). 253 254 To examine how targeted illumination impacts correlation measurements between simultaneously recorded 255 neuron pairs, we computed spike-spike and Vm-Vm correlations as detailed above in cultured neuron 256 experiments. Unlike in cultured neurons, here due to tissue scattering and fluorescence from out-of-focus 257 neurons, we observed that targeted illumination significantly reduced Vm-Vm correlation values (Fig. 4n). 258 However, spike-spike correlation values remained largely consistent under both conditions (Fig. 4o). Since spikes 259 are only produced when Vm depolarization reaches sodium channel activation threshold for action potential 260 generation, joint synaptic inputs that produce correlative low amplitude Vm changes between neuron pairs that 261 are subthreshold in theory will not be captured by spike-spike correlation measures. The fact that Vm-Vm 262 correlation is reduced by targeted illumination highlights that Vm signals contain a higher proportion of 263 background signal crosstalk than spiking signals.  photobleaching and thus allowed for recording over an extended duration under targeted illumination. Figure 5  287 represents an example of a continuous recording of 5 minutes, revealing that excellent spike SBRs can be 288 maintained throughout the recording duration. Of the two simultaneously recorded neurons, the spike SBR for 289 neuron 1 was 4.13 ± 1.1 (mean ± standard deviation, n = 1366 spikes), and for neuron 2 was 4.81 ± 1.29 (mean ± 290 standard deviation, n = 261 spikes). However, we did notice a reduction in spike SBR over time ( Having established the significant advantage of targeted illumination, we deployed targeted illumination to image 301 multiple neurons in the dorsal hippocampus CA1 region. Combining targeted illumination with a high-speed, large 302 sensor sCMOS camera, we sampled a FOV of 360 × 180 µm 2 , often containing tens of neurons. We performed 6 303 recordings of 17 or more CA1 neurons (37 ± 22 neurons per session, mean ± standard deviation), while mice were 304 awake and head-fixed navigating on a spherical treadmill. Across these recording sessions, we recorded a total of 305 222 spiking neurons, with a spike SBR of 4.16 ± 0.5 (mean ± standard deviation, n = 222 spiking neurons, Fig. 6, 306 Fig. S6). In one recording, we were able to record 76 neurons simultaneously, and detected spikes in 58 of those 307 neurons, over a 90-second long recording period (Fig. 6). The mean spike SBR of these neurons was 3.94 ± 0.4 308 (mean ± standard deviation, n = 58 neurons). recorded CA1 neuron pairs revealed that both Vm-Vm and spike-spike correlations substantially decreased with 320 anatomical distance (Fig. 7, Kruskal-wallis, p = 5.16e -7 , df = 5109 for Vm-Vm correlations; p = 3.18e -6 , df = 4771 for 321 spike-spike correlations). These results are consistent with that observed in cultured neurons and reflect the 322 general understanding that nearby neurons tend to receive more temporally aligned synaptic inputs relative to 323 neurons further apart 23 . 324 We found that by restricting illumination to neuronal cell bodies, we were able to significantly increase SomArchon 343 signal quality in terms of spike SBR and reduce out-of-focus background fluorescence. These improvements were 344 more substantial for imaging neuron in the brain than in single layered neuron cultures, and were consistently 345 observed across the two brain regions tested that have varying labeling density, including the superficial layers of 346 visual cortex with sparsely labeled neurons and the hippocampus with densely labeled neurons. With such 347 improvements in SomArchon signal quality, together with a high-speed large sensor size sCMOS camera, we were 348 able to record optical voltage signals from over 70 neurons simultaneously over a wide FOV of 360 × 180 µm 2 at 349 500 Hz. 350 351 One advantage of using targeted illumination is the reduced power density of excitation light, from both direct 352 ballistic excitation photons and backscattered photons from tissue scattering. In this study, the ballistic excitation 353 power density used for in vivo recordings was measured at 3 -5 W/mm 2 , which equals to 0.7 -1.1 mW per neuron 354 (assuming a 15 x 15 µm 2 square excitation region). However, for in vivo imaging, the actual excitation power will 355 be further affected by tissue scattering. Photons targeting a cell can be scattered away from the ROI, whereas 356 photons targeting non-ROI regions could eventually reach an ROI due to forward and backward scattering. 357 Therefore, although the same excitation power density was applied in both targeted illumination and widefield 358 illumination conditions, neurons under widefield illumination conditions were actually exposed to higher 359 excitation power, increasing photobleaching therefore causing greater observed fluorescence decay. In addition, 360 under targeted illumination, the reduced background and consequently improved spike SBR also allowed us to 361 reduce the ballistic excitation power. As a result, we were able to perform continuous recordings over several 362 minutes in duration, with only moderate reductions in spike SBR. While the performance of fluorescence based 363 activity indicators are always limited by photobleaching, deploying trial-based study designs without excitation 364 illumination during inter-trial-intervals should allow SomArchon to measure membrane voltage over many trials, 365 and potentially over a greater cumulative period of time than demonstrated here using continuous illumination. 366

367
Other than large FOV and long-term imaging, our DMD-based targeted illumination widefield microscope also 368 presents several additional advantages that would help to provide greater access to general voltage imaging 369 applications. With the use of SomArchon, only a single-color excitation source is required to be patterned through 370 the DMD. This greatly simplified our targeted illumination module that the DMD surface can be directly imaged 371 onto the sample through a tube lens and an objective, as opposed to alternative systems that requires dual-color 372 excitation 5 or more complicated holographic targeting 6 . For wide FOV SomArchon voltage imaging, large area 373 excitation with a power density on the order of a few W/mm 2 is necessary. While typically LEDs have insufficient 374 power density, it can be easily satisfied with a multimode laser diode array because of the widefield nature of our 375 DMD-based targeting strategy. Such light source would also avoid the undesirable speckle artifacts associated 376 with targeting techniques that require coherent sources 6 . Additionally, in our system, the initial GFP channel 377 structural imaging was performed with the same widefield microscope using an extra blue LED excitation. Even 378 though this can also be performed using an additional two-photon microscope 5,6 , the integrated widefield 379 microscope design described significantly reduces the cost and the complexity of the optical system and the 380 software control. Since both GFP and SomArchon fluorescence were capture by the same camera, both images 381 were automatically co-registered, further alleviating the issue of image registration and long-term system stability 382 across multiple microscope modules. Overall, such simplified designs provide a cost-effective, easy to implement 383 solution for large scale voltage imaging analysis of neural networks in behaving animals. 384 385 Single cell level imaging in living animals is always subject to fine movement due to metabolic, physiologic, and 386 vascular changes, and a solution to such fine motion interference is through offline correction via image 387 registration 24,25 . Targeting illumination only to cell membranes using holographic projections however is sensitive 388 to translational movement due to the restricted area of illumination, which introduces additional challenge on the 389 preparation of animal subjects, and the design of the behavioral tasks. The DMD-based widefield targeting 390 strategy described here alleviate some of these concerns, where the illumination window size can be easily 391 adjusted to accommodate fine biological motion. For example, increasing the region of targeted illumination to 392 capture the morphological details of a neuron allows for fine translational movement that can be effectively 393 corrected offline after image acquisition. Single photon widefield imaging is also more amenable to axial motions 394 due to the lack of optical sectioning 18 , which allows for continuous recording of signal during subtle fluctuations 395 in axial positions. While such advantage is retained to some degree over the illuminated regions with DMD-based 396 targeted illumination, it is less so when using holographic projections 6 . Techniques with confined excitation 397 volume limited to narrow z-axis profiles, such as two-photon microscopy 26 , are also more sensitive to image 398 motion expected from behaving animals 27 . 399 To estimate SomArchon fluorescence quality, we calculated spike SBR. The baseline used in this SBR calculation 400 contains both biological subthreshold membrane voltage fluctuations and SomArchon intrinsic fluorescent shot 401 noise. Neurons in intact neural circuits, especially in the awake brain, receive heterogenous synaptic inputs and 402 exhibit distinct membrane biophysical properties, which lead to variation in subthreshold membrane voltage 403 fluctuations that are difficult to estimate. Thus, by itself, spike SBR estimation for each neuron cannot fully capture 404 the quality of SomArchon signal contrasts, and represents an underestimation of SomArchon performance as 405 illustrated with our spiking neuron models. However, spike SBRs of the same neuron when compared under the 406 two illumination conditions can provide a quantitative measure of the fluorescence signal quality, providing direct 407 experimental evidence that targeted illumination significantly improve the quality of SomArchon voltage imaging. 408 Since spike SBR is a key consideration for spike detection, the fact that we detected more spikes under targeted 409 illumination condition further demonstrates the improvement of SomArchon voltage signal quality. analysis, decision to publish, or preparation of the manuscript. 428

Simulated data theory for widefield fluorescence imaging 431
We consider the problem of widefield fluorescent imaging in a mouse brain in the context of imaging through 432 scattering media within the forward scattering limit. In our model (Fig. S1), an incoherent source located at plane 433 = 0 is embedded at depth = inside a scattering medium, whose scattering properties are characterized by 434 the scattering phase function (̂), mean scattering length , and anisotropic factor ≈ 1. The image of the 435 scattered light field from the source is relayed by a unit Eq. (6) is the main results that we use for simulating widefield neuronal imaging, the interpretation of which is 459 that the propagation of mutual coherence can be simply considered as free space propagation with an additional 460 attenuation factor − ( , κ ) due to scattering. Note that this result not only holds for imaging of fluorescent 461 signals in the detection path, but can also be applied to delivering illumination patterns onto a scattering sample 462 in the excitation path (i.e., targeted illumination). 463 Biological tissues such as the brain are typically characterized by strong forward scattering where ≈ 1, where 464 the distribution of scattering angles follows the Henyey-Greenstein phase function 28 : 465 Assuming a circular microscope aperture of radius , substituting Eq. (7) into Eq. (6) and using the Stokseth 466 approximation of free space 3D optical transfer function (OTF) 29 , we arrive at the analytical solution of the 3D 467 SOTF: 468 where Δκ ⊥ = 2 /λ, = / 0 is the numerical aperture of the system, and 469 is the in-focus free space OTF. With Eq. (8), we can calculate the detected image or projected pattern simply by 470 filtering the original object/pattern in frequency space using the corresponding SOTF. 471

Simulation of widefield illumination versus targeted illumination conditions 472
Using the theoretical model developed above, we compared the background fluorescence signals generated using 473 widefield and targeted illumination. We estimated the reduction of background fluorescent signals from non-474 targeted SomArchon expressing neurons, or in other words, signal cross-contamination, with the use of targeted 475 illumination compared to standard widefield illumination. For simplicity, here we only modeled a pair of neurons 476 that are separated by a distance laterally and axially ( Fig. S2 and Fig. S3). Each neuron was assumed to be a 477 15 µm diameter uniformly fluorescent sphere. For widefield illumination, the entire FOV was illuminated equally. 478 For targeted illumination, only a 15 µm circular ROI was projected onto the sample centered at the location of the 479 neuron of interest. Although both neurons were imaged onto the camera, only the targeted one contained the 480 signal, and the contribution from the other non-targeted neuron within the ROI of the targeted neuron (the red 481 circle in Fig. S2 d-i and Fig. S3 c-h) was considered as background (or crosstalk). 482 In the simulation, we assumed the imaging system has unit magnification and = 0.4. The excitation and 483 emission wavelength are = 637 nm , = 670 nm respectively, with corresponding tissue anisotropic 484 factor 637nm = 0.89, 670nm = 0.90, and mean scattering length ,637nm = 110 μm, ,670nm = 119 μm 30 . 485 Two different scenarios for optical voltage imaging were considered, namely in vitro imaging in 2D neuronal cell 486 culture and in vivo imaging in a mouse brain. 487 For in vitro imaging in cultured neurons, since it typically consists of a monolayer of cells, we therefore assumed 488 the two neurons are at the same depth = 0 (Fig. S2 a,b) with no tissue scattering. By varying lateral distance 489 , we plotted the amount of crosstalk induced by the non-targeted neurons in Fig. S2  is close to 0. Therefore, we expect very little benefit of using targeted illumination for reducing crosstalk in in vitro 494 imaging. However, targeted illumination still pertains certain advantages over widefield illumination in terms of 495 photobleaching and SBR because of the reduction of stray light and non-specific background signals. 496 For in vivo imaging in a mouse brain, we assumed that the targeted neuron was located at depth = 100 µm 497 (see Fig. S3 The two coupled differential equations were numerically solved using the Euler method with 1 ms step size, 513 modeling a 1KHz sampling rate. For the parameters, we chose: a = 0.02, b = 0.2, c = -55 mV, d = 2 as often used 31 . 514 The input to the neuron was composed of a fixed input current to each neuron (2.8 mv) and gaussian current noise 515 (standard deviation = 2 mV). 516 To estimate the spike SBR (see details below), we divided the spike amplitude (here defined as -50 mV to 30 mV 517 = 80 mV) by the baseline voltage fluctuations, including both the biological (dynamic) noise variance and the added 518 measurement gaussian white noise variance. SNR was defined as the spike amplitude divided by the added 519 gaussian white noise variance only. 520 and 800 µM L-Glutamine. 11 days later, 5-fluoro-2-deoxyuridine (Millipore cat. # 343333) was added at a 531 concentration of 4 µM to prevent glial cell overgrowth. 50% of the cell culture medium was exchanged every 3 532 days. Neurons were transduced with 0.25 µL of AAV9-syn-SomArchon per well in 0.25 mL of feeding media, 3-4 533 days after plating. Cells were imaged 14-16 days after plating, in an imaging buffer containing 145 mM NaCl, 2.5 534 mM KCl, 10 mM glucose, 10 mM HEPES, 2 mM CaCl2, and 1 mM MgCl2, pH 7.4. 535

Animal surgical procedures 536
All procedures involving animals were approved by the Boston University Institutional Animal Care and Use 537 Committee (IACUC). C57BL/6 adult female mice (3-6 months old on the day of recording) were used in this study. 538 Mice were surgically implanted with an imaging chamber and a head-plate as described previously 3 . AAV-syn-539 SomArchon was injected either through an infusion cannula attached to the window after the surgery, or injected 540 during the surgery. 541

Custom widefield optical imaging setup 542
We customized a dual color epi-fluorescence fluorescence microscope, which used a 470 nm LED (Thorlabs,  543 M470L3) for GFP fluorescence excitation, and a 637 nm fiber-coupled laser (Ushio America Inc., Necsel Red-HP-544 FC-63x) for SomArchon fluorescence excitation. The two illumination channels were combined using a dichromatic 545 mirror (Thorlabs, DMLP550R) and subsequently directed onto the sample. The generated fluorescent signal was 546 epi-collected by a microscope objective (Nikon, 40×/0.8NA CFI APO NIR) and imaged onto a camera (Hamamatsu,  547 ORCA-Lightning C14120-20P) with a 175 mm tube lens. A combination of excitation filter, dichromatic mirror, and 548 emission filter (Semrock, LF405/488/532/635-A-000) was used to separate fluorescent signals from the excitation 549 light. 550 To pattern the illumination in the SomArchon imaging channel, the output of the 637 nm multimode laser was 551 collimated (Thorlabs, F950SMA-A), expanded (Thorlabs, BE02M-A), and directed onto a DMD (Vialux, V-7000 VIS) 552 at approximately 24° to its surface normal. The DMD was further imaged onto the sample with a 175 mm lens and 553 the objective, so that only sample regions corresponding to the 'on' pixels of DMD were illuminated. The axial 554 position of the DMD was adjusted so that it is conjugate to the camera, and an additional affine transform was 555 estimated to register the pixels between the DMD and the camera. The DMD was controlled using custom Matlab 556 script based on Vialux ALP-4.2 API. 557 During each imaging session, a GFP fluorescence image was first taken for illumination target identification, where 558 a small rectangular ROI was manually selected for each individual neuron to be imaged. A binary illumination mask 559 was then generated based on all the selected ROIs and uploaded to the DMD for illumination targeting. 560 SomArchon voltage imaging was performed at 500 Hz, with 2 × 2 pixel binning, resulting in an imaging area of 561 1152 × 576 pixels on the sCMOS camera sensor, corresponding to a 360 × 180 µm 2 FOV at the sample. To estimate 562 sCMOS camera dark level and intrinsic noise, videos were collected with the camera set to the same acquisition 563 parameters as during regular imaging experiments, but without light exposure (500 Hz, 2 × 2 pixel binning, 1152 564 × 576 pixels imaging area). The sensor dark level was estimated to be 767.7, with an intrinsic noise of 12.6 565 (standard deviation) per pixel. 566

Data analysis 567
All imaging data were acquired by HCImage software (Hamamatsu), and further processed using MATLAB 568 (Mathworks) offline. 569

Neuron ROI selection 570
All data analysis was performed offline in Matlab 2019b or 2020a. SomArchon fluorescence images were first 571 motion corrected using a pairwise rigid motion correction algorithm as described previously 24 . For targeted 572 illumination recordings, each ROI was centered on a neuron of interest, with the ROI size slightly greater than the 573 outline of the neurons, so that motion correction can be performed on each targeted ROI that had distinguishable 574 features identifiable by the algorithm. After motion correction, we manually selected ROIs corresponding to 575 individual neurons, based on the average SomArchon fluorescence image during the first recorded trial. ROIs were 576 cross-referenced by comparing SomArchon fluorescence with the stable EGFP fluorescence. The identified 577 neurons were then applied to all subsequent trials in the same recording session. SomArchon fluorescence traces 578 were then extracted for each neuron by averaging all the pixels within the neuron across the entire experiments. 579 For direct comparison of SomArchon fluorescence of the same neurons between widefield and target illumination 580 conditions, the same neuron ROIs were used for both recording conditions. Trace time segments with sharp, 581 drastic changes in fluorescence (e.g. due to motion) were detected as outliers and excluded from further analysis 582 in both the widefield and targeted illumination analysis. Specifically, for the outlier detection we applied the 583 generalized extreme Studentized deviate test on the moving standard deviation values using a sliding window of 584 ±60 ms on spike-removed traces (see Method Section spike detection and spike SBR calculation). In some cases, 585 not all time points during the period of an artefact were marked as outliers. Time points between outliers (< 3 586 data points) were therefore interpolated. To remove further artefacts, we excluded time points that were 6 587 standard deviations outside the trace fluorescence distribution. Time points between and around the detected 588 outliers were also removed (±350 ms) as those periods often coincided with extended animal motion artefacts. 589

Fluorescence decay estimation 590
To estimate SomArchon fluorescence decay, we first removed spikes by applying a median filter (window of 51 591 frames), and then subtracted the camera dark level (measured as 767.7). We calculated fluorescence decay as the 592 ratio of the mean fluorescence intensity during the first 600 ms and that during the last 600 ms for each trial, and 593 then averaged across all trials. In cultured neurons, we detected a drastic fluorescence drop within the first couple 594 seconds of recording, likely mainly due to bleaching of autofluorescence unrelated to SomArchon, thus we 595 excluded the first trial from subsequent analysis for culture neuron analysis. 596

Spike detection and spike SBR calculation 597
To separate spikes from subthreshold voltage fluctuations, we first generated a "Smoothed Trace" (ST) by 598 averaging the fluorescence trace using a moving window of ±100 frames. To estimate baseline fluctuation, we first 599 removed potential spikes by replacing any fluorescence values above ST with the corresponding values of ST. The 600 amplitude of the baseline fluctuation was then estimated as 2 times the standard deviation of the resulting trace, 601 since half of the fluctuations were removed in the spike removal step described above. For spike SBR estimation, 602 we also subtracted the camera intrinsic noise (standard deviation = 12.6 per pixel) from the trace noise to obtain 603 camera-independent estimates. 604 For spike detection, we first removed small subthreshold rapid signal changes by replacing the fluorescence below 605 ST with corresponding values of ST. The derivative of the resulting trace was then used for spike detection, where 606 spikes were identified as the time points above 4.5 times of the standard deviation of the resulting derivative 607 trace. Spike amplitude was calculated as the peak fluorescence for each spike minus the mean of the fluorescence 608 during the three time points before spike onset. Spike SBR was calculated as spike amplitude divided by the 609 amplitude of baseline fluctuations described above. 610

Pearson correlation analysis 611
Pearson cross-correlation was performed using the Matlab functions corrcoef and xcorr, for both Vm-Vm and 612 spike-spike correlation analysis. To calculate spike-spike correlation, spike vectors were smoothed over a ±10 ms 613 time window for each spike before applying correlation analysis. To calculate Vm-Vm correlation, we removed 614 spikes by replacing 3 data point centered at each identified spike times with the adjacent values that largely 615 eliminated the contribution of spikes in Vm-Vm correlation analysis. 616

Statistical analysis 617
Paired student's t-tests were used for comparisons involving the same neurons between the targeted 618 illumination condition and the widefield illumination condition. A Kolmogorov-Smirnov test was used to test the 619 difference of cross-correlation over distance between targeted illumination and widefield illumination 620 conditions. For Kolmogorov-Smirnov test, the data in each of the two conditions were first sorted by distance 621 before comparing. A Friedman's test, 2 factor non-parametric ANOVA, was used to compare the difference 622 between the average correlations of each bin in Fig. 3. 623

Data and software availability statement 624
Codes used for data analysis is available on our lab website and Github repository: 625 https://www.bu.edu/hanlab/resources/ and https://github.com/HanLabBU 626