Visualizing the Role of Lipid Dynamics during Infrared Neural Stimulation with Hyperspectral Stimulated Raman Scattering Microscopy

Infrared neural stimulation, or INS, is a method of using pulsed infrared light to yield label-free neural stimulation with broad experimental and translational utility. Despite its robust demonstration, the mechanistic and biophysical underpinnings of INS have been the subject of debate for more than a decade. The role of lipid membrane thermodynamics appears to play an important role in how fast IR-mediated heating nonspecifically drives action potential generation. Direct observation of lipid membrane dynamics during INS remains to be shown in a live neural model system. To directly test the involvement of lipid dynamics in INS, we used hyperspectral stimulated Raman scattering (hsSRS) microscopy to study biochemical signatures of high-speed vibrational dynamics underlying INS in a live neural cell culture model. Findings suggest that lipid bilayer structural changes are occurring during INS in vitro in NG108-15 neuroglioma cells. Lipid-specific signatures of cell SRS spectra were found to vary with stimulation energy and radiant exposure. Spectroscopic observations were verified against high-speed ratiometric fluorescence imaging of a conventional lipophilic membrane structure reporter, di-4-ANNEPS. Overall, the presented data supports the hypothesis that INS causes changes in the lipid membrane of neural cells by changing lipid membrane packing order – which coincides with likelihood of cell stimulation. Furthermore, this work highlights the potential of hsSRS as a method to study biophysical and biochemical dynamics safely in live cells.


27
Neuromodulation using directed energy, including optical, ultrasonic, and radio frequency, have gained 28 notable interest recently due to their spatial precision, noninvasive implementation, and promising 29 potential for clinical translation in therapeutic interventions. Label-free optical neuromodulation with 30 pulsed infrared (IR) light, or infrared neural stimulation (INS), offers a spatially and temporally precise 31 means of contact-free activation of neural cells without the need for genetic modification or exogenous 32 mediators. Similar to most label-free directed energy methods of neuromodulation, the biophysical 33 mechanisms underlying INS have remained elusive for more than a decade (1). In contrast to the tools 34 derived from molecular biology, such as optogenetics or photochemical uncaging, INS appears to act 35 through an entirely different photothermal-based mechanism (1, 2). The role of lipid membrane dynamics 36 are thought to play an important role in how IR light depolarizes neurons photothermally (3), but remains 37 to be directly experimentally validated in a live neural model system. Conventional methods of directly measuring lipid bilayer geometry, such as x-ray diffraction and small 54 angle neutron scattering, are slow and not biologically compatible (8-10). Optical methods are well 55 suited for high resolution, biologically compatible experiments, but generally lack the spatial resolution 56 necessary to directly resolve lipid bilayer geometry (< 3 nm thick) on millisecond timescales. Fluorescent 57 functional lipid indicators, such as laurdan or di-4-ANNEPS (11), have been shown to be powerful tools 58 in studying lipid membrane biophysics. However, these indicators offer latent readouts of lipid dynamics 59 and are inherently indirect in that they rely on the molecular interaction of reporter molecules with their 60 molecular environment rather than the lipid molecules themselves. Vibrational spectroscopic methods, 61 such as Raman scattering and infrared absorption, can be performed label-free and offer a feature-rich 62 molecular signature useful in studying lipid organization in live cells. Traditionally, vibrational 63 spectroscopic methods have not been biologically compatible on sub-second timescales (12, 13). 64 Stimulated Raman scattering (SRS) microscopy combines label-free vibrational spectroscopic contrast 65 with subcellular spatial resolution and sub-second temporal resolution enabling time resolved vibrational 66 spectral measurements of live neural cells during INS (14,15 To calibrate the vibrational spectral dimension of hyperspectral imaging space, 50 sequential images were 159 acquired of mounted polymer bead monolayers. Between each acquired image, the optical path length 160 delay of the 798 nm laser line was stepped by 10 µm between each image, over a total of 500 µm or 1.6 161 ps of total optical path length delay. The peak SRS signal for the 2950 cm -1 resonance of PMMA was 162 centered in the spectral scanning range to ensure sufficient spectral sampling. Manual segmentation of 163 PMMA and PS beads from spectral stacks were performed and averaged across each spectral frame to 164 provide high-fidelity spectra for both polymers. The known vibrational peaks of PS (2910, 3060 cm -1 ) and 165 PMMA (2950 cm -1 ) were used as spectral fiducials (Figure 1B and C

171
Neural stimulation was performed by placing a bare 400 µm-diameter core low-OH optical fiber (Ocean 172 Optics, FL, USA) in close proximity to samples (~450 µm) at a 30-degree approach angle into the sample 173 plane of the microscope's field of view ( Figure S1). The optical fiber used for stimulation is connected to 174 a pulsed laser diode centered at 1875 nm (Capella Nerve Stimulator, Aculight -Lockheed-Martin, Bothel, 175 WA, USA). During imaging experiments, samples were exposed to a pulse train of 188 pulses distributed 176 evenly over 1500 ms). Pulses were 400 µs in duration and were delivered at a repetition rate of 125 Hz. 177 Radiant exposures on samples were varied by adjusting the peak current delivered to the laser diode, 178 holding all dosing and geometric configurations constant. Radiant exposure calculations for stimulation 179 were approximated based on power measurements performed externally in air and employing Beer's law 180 under the assumption of an absorption-dominated photon distribution -described in Figure S1 and 181 Figure S2.  (Figure S4A).

198
Live Cell Hyperspectral SRS Imaging 199 Live cell imaging experiments of endogenous vibrational contrast with hsSRS were conducted with  200 adherent cell preparations imaged in a physiologically balanced saline solution. Following placement of 201 the fiber and calibration of the spectral axis against the known vibrational peaks of PS and PMMA beads, 202 baseline hyperspectral image stacks were acquired for live cell samples. All images were acquired in a 203 point-scanning approach with a 5 µs pixel dwell time and a spatial sampling density of ~500 nm/px. To 204 improve signal to noise ratio of higher fidelity images, square fields of view between 320 and 512 pixels 205 in size were acquired and 6 to 10 images were averaged together for each spectral position. For 206 hyperspectral image stacks acquisitions, 50 images were acquired at evenly spaced intervals (10 μm) over 207 500 μm of optical path length delay -corresponding to a spectral range spanning approximately 2800 to 208 3100 cm -1 . observed to be defocused at the sample plane due to the thermal gradient induced by the stimulating 219 infrared laser (Figure 2A) was observable in each imaging timeseries as an exponential decrease, and 220 subsequent return to baseline (Figure 2B&C), of nonlinear signal on imaging photodetectors. The shift in 221 focal length as a function of laser power was calibrated using microbead (PMMA and PS) preparations 222 and accounted for prior to each IR-stimulation trial on cells. The defocusing phenomenon allowed for 223 precise temporal synchronization of time series across each spectral channel. After repeating and 224 temporally aligning simultaneous imaging and stimulation time courses on lives cells for each SRS 225 spectral position (n = 50), the temporal evolution of live cell endogenous vibrational spectra could be 226 observed as a function of irradiation time and deposited energy. For spectral evaluation, the final ten 227 sampling time points within the of IR exposure were averaged and reported -which was found to help 228 reduce high frequency spectral noise to draw conclusions from. Spectra from stimulation experiments 229 were pooled from n = 24 cells across ten different individual experiments of IR exposure. Each cell 230 spectrum was normalized with respect to its integrated spectral intensity, and standard deviation of the 231 spectra across all cells in each stimulation condition were calculated. The 'no stimulation' conditions are 232 obtained from initial SRS signal from cells prior to each round of IR exposure and pooled from all 233 stimulation conditions being compared. The shape of SRS spectra acquired at high frame rates ( Figure  234 3B) were not found to noticeably differ from higher fidelity spectra ( Figure 1D).

236
To verify cell viability during IR exposure, NG108 cells were subject to the hsSRS and stimulation 237 protocol described above while simultaneously monitoring for cell damage via positive fluorescence 238 staining of cell nuclei with propidium iodide. Imaging protocols were kept identical as previously 239 described while supplementing the cell imaging medium with 1 µM propidium iodide (Thermo-Fisher, 240 Natick, MA, USA). Cell morphology was additionally monitored throughout the experiment by 241 comparing high fidelity images (< 1 μm/px sampling density) of the cells before and after imaging at their 242 peak SRS resonance contrast at 2930 cm -1 . 243 244 245 Imaging protocols were adapted from previously published work (25 of the multidimensional image stack and contrast local histogram equalization was performed to reduce 278 cell signal intensity variations between cells. Post-hoc flat field correction of imaging field heterogeneity 279 of images was implemented by scaling pixel intensities relative to the average intensity projection 280 gaussian blurred with a kernel equal to 0.25-0.5x the largest dimension of a particular image. Prominent 281 peak locations in the image are identified. The filtered average intensity projection is subsequently 282

di-4-ANNEPS Ratiometric Fluorescence Imaging
segmented via Otsu segmentation. The resulting mask and previously identified peak locations are fed 283 into an seeded watershed segmentation algorithm which reliably separates and segments induvial cells as 284 their own ROIs with minimal cell-to-cell overlap (28). Edge maps of cells were acquired by subtracting 285 the watershed-segmented mask from itself following an erosion operation, which reliably identifies 286 borders in a cell-specific manner. The resultant regions of interest are applied to the raw stacks to extract 287 the mean amplitude, standard deviation of signal or amplitude measurements, and their respective 288 centroid locations in image-space for each spatial and temporal point. This process is automated as a 289 macro procedure in FIJI and is freely available with raw data examples as supplementary information. 290 Depicted images provided in the manuscript are derived either from single frames at specific 291 wavenumbers of interest or maximum intensity projections of spectral image stacks. For visualization 292 purposes in publication, intensity scaling for all images were adjusted linearly. 293 All hsSRS spectra are smoothed with a 3-point sliding Gaussian window and normalized with 294 respect to their integrated spectral area. Since the intent of the study is to compare the relative spectral 295 shapes of each sample, an integrated spectral normalization was chosen to facilitate this interpretation. 296 Error associated with each plot is presented as standard deviation of all averaged spectra obtained for a 297 given experimental trial. Each individual bead was taken as one sample, and different trials were taken as 298 independent observations for statistical analysis purposes. For peak ratio comparisons, vibrational 299 resonance intensities were calculated utilizing a cubic spline interpolation of the measured spectral data 300 and it's respective standard deviation. Comparisons of peak ratios were assessed using a student's 2-sided 301 t-test, where errors associated with ratiometric comparisons were calculated based on the propagation of 302 error of the interpolated standard deviations (statistical significance was denoted by * for p < 0.05, ** 303 for p < 0.01). All quantitative work was performed in MATLAB (Mathworks, Natick, MA, USA) using 304 native functions. All bar graphs were created using the superbar package. 305

306
Processing of ratiometric fluorescence data is derived in part from previous work (29). Raw 307 image stacks of green (lipid membrane gel phase -ordered) and red (lipid membrane liquid phase -308 disordered) spectral emission channels are acquired simultaneously at a 33.4 Hz framerate. Conventional 309 general polarization (GPconv) was calculated using the following equation (29) The raw image intensity differences between the green (ordered, O(t)) and red (disordered, D(t)) imaging 315 channels were divided by the sum of both channels for each timepoint in the image stack for each 316 experiment. Decreases in GPconv value generally suggest decreases in membrane packing order. Average 317 GP values as a function of time were calculated and each cell's GP value was taken as an average GP of 318 all pixels contained in each cell's ROI. Cell segmentation similar to those segmented for SRS images 319 utilizing a seeded watershed method was performed. However, since di-4-ANNEPS labels the 320 extracellular membrane preferentially, a Huang threshold mask of raw disordered spectral fluorescence 321 intensity images were obtained to determine cell boundaries and a binary fill operation was employed to 322 identify areas in the image that contained cells. The lack of lipid-stained fluorescence in cell nuclei was 323 used to identify center points of cells. The raw disordered fluorescence channel image was smoothed with 324 a 2-pixel Gaussian filter and local minima in the images were used to approximately localize cell center 325 points. These cell center points, as well as the cell position mask, and a distance map calculated from the 326 cell position mask were fed into a seeded watershed algorithm in FIJI to yield segmentation maps of 327 individual cells in a given experiment (26, 28). The regions of interest derived from the segmentation 328 were subsequently applied to each imaging experiment, where time series of both raw fluorescence 329 channels were obtained per cell and the resultant data was exported for processing and analysis in 330 MATLAB (Mathworks, Natick, MA, USA). Statistical comparison of GP values across stimulation 331 conditions was performed using a 2-sided student's t-test and the magnitudes and standard error of means 332 across the GP values were calculated across all individual cells in a particular experimental condition 333 (statistical significance denoted as * for p < 0.05). 334 For image visualization, adapted from previous work (29), 8-bit depth raw fluorescence intensity 335 images from the disordered fluorescence channel were multiplied by each color channel of an color red-336 green-blue (RGB) format image representing the calculated GP images with the desired false-colored 337 look-up table of preference. The resulting images yield an image where pixel brightness represents 338 intensity and color represents calculated conventional general polarization -which are used purely for 339 visualization purposes. All rescaling of intensities in images are linear and performed for clarity of 340 cellular morphologies and biophysical properties in print ( Figure 4A). 341 Due to large variations in total fluorescence measured in any given experiment due to thermal 342 lensing during IR stimulation, the conventional method of calculating GP was found to be unreliable. 343 Since we expect a decrease in overall fluorescence due to the decrease in effective collection efficiency 344 during thermal lensing induced defocusing, the magnitude of changes in the denominator of the GPconv 345 equation are much larger than that of the changes in the numerator of the equation. To account for these 346 effects, we developed an intensity invariant version of GPconv to better reflect these dynamics 347 mathematically over short experimental periods of time undergoing substantial changes in photon 348 collection: 349 350 Where O0 represents initial ordered fluorescence levels, D0 represents initial disordered fluorescence 353 levels 354 Ooff(t) represent the net change in ordered fluorescence relative to O0 as a function of time, and Doff(t) 358 represents -positive or negative -rather than it's magnitude. This consideration makes GPmod a convenient and 367 applicable tool for our experimental approach. 368

369
Thermal Lensing during IR Stimulation 370 Following confirmation of our instrument's ability to obtain hsSRS image stacks from live NG108 cells 371 (Figure 1D and E), initial experiments with IR stimulation during nonlinear microscopy (i.e. any 372 coherent Raman modality, multiphoton fluorescence, or higher harmonic generation) resulted in a 373 substantial loss in measured signal during IR exposure (Figure 2B and C) (22). This was apparent in 374 both short periods of heating from a millisecond pulse of IR light (unpublished data), or pulse trains of 375 multiple microsecond pulses of light. The shape of the disappearance and reappearance of the nonlinear 376 signal appears to follow the shape of the expected heating and cooling dynamics that is typically observed 377 during IR mediated heating (30)-suggesting that a temperature related phenomenon may be responsible 378 for the loss in signal. Considering the goal of this work is to image the high-speed chemical dynamics in 379 live cells during IR exposure, the loss of signal during this critical time period posed a challenge. To 380 better understand the role of this signal loss with immersion medium temperature, a vegetable oil sample 381 was imaged with SRS (2885 cm -1 ) through warmed immersion medium at a range of physiologically 382 relevant temperatures. Temperature of the immersion medium was monitored by a thermocouple placed 383 adjacent to the microscope's field of view at the coverglass-immersion medium interface ( Figure S3). 384 Warmed deionized water (approximately 50 °C) was added between the objective and sample with the 385 edge of vegetable oil sample placed in focus. Images were acquired continuously as the immersion 386 medium slowly cooled to room temperature (22 °C). Contrary to the signal decrease observed during 387 rapid IR heating (Figure 2B and C), this experiment showed that changes in immersion medium 388 temperature revealed a positive correlation with temperature and SRS signal of vegetable oil. This data 389 suggested that changes in immersion medium temperature on its own was not sufficient to explain the 390 decrease in nonlinear optical signal during IR heating. 391 The refractive index of the objective immersion medium, water (H2O), is negatively correlated 392 with temperature (31). This concept suggests that the spatial thermal gradients generated by the IR 393 absorption from IR stimulation would defocus the ultrafast laser driving nonlinear contrast and thus 394 reduce observed nonlinear optical signal. To test this hypothesis, the immersion medium for the objective 395 lens was replaced with heavy water (D2O), which has a five-fold lower absorption coefficient at 1875nm 396 than deionized water with nearly identical refractive indices (2). If the thermal gradient causes the 397 decrease in nonlinear signal observed in the sample, then reducing the immersion medium's IR absorption 398 properties should reduce the magnitude of the nonlinear signal decrease during stimulation. The results 399 shown in Figure 2B validates this hypothesis ( Figure 2B) suggesting that the thermal gradient from IR 400 stimulation was defocusing the ultrafast laser source resulting in a decrease in nonlinear signal ( Figure  401 2A). 402 Since water's index of refraction is negatively correlated with temperature, the thermal gradient 403 generated during IR stimulation in front of the stimulation fiber and within the microscope's field of view 404 behaves like a negative lens during imaging. Imaging out of focus samples during IR stimulation would 405 bring samples into focus (Figure 2A). This hypothesis was found to be true for both nonlinear imaging 406 and IR transillumination imaging. By moving the microscope's focal plane above the sample by a few 407 microns prior to IR exposure, the samples (polymer microbeads in this case) would come into focus 408 ( Figure 2C).  Figure 2D shows a representative image of mixed microbead monolayers, highlighting PMMA in 417 cyan using the band at 2950 cm -1 (terminal methyl C-H resonance) and PS in orange using the band at 418 3050 cm -1 (aromatic C-H stretch resonance). The mixed bead sample was exposed to ~12 J/cm 2 IR 419 stimulation and the resultant spectra for both bead types are shown in Figure 2E&F. Relevant spectral 420 band assignments for polymer microbead samples are summarized in Table 1. Infrared-exposed PMMA 421 beads exhibit several distinct spectral changes upon heating -decreases in the 2880 and 2910 cm -1 422 resonances of skeletal C-H stretching, as well as relative increases in resonances at 3000 cm -1 and 423 decreases at 3050 cm -1 . Shifts in PS hsSRS spectra during IR exposure show relative increased vibrational 424 activity around 2850 cm -1 , implying the possibility of relaxed steric hinderance of skeletal sp 3 CH2 425 symmetric stretching modes, while broadening the 3050 cm -1 peak attributable to aromatic sp 2 C-H 426 stretching and suggesting reduced steric hinderance around aromatic side chains. These observations 427 show that utilizing a time resolved approach to obtaining hsSRS spectra of samples heated by pulsed IR 428 light is feasible in highly Raman active idealized chemical samples. 429 The dominant Raman scatterers in the 2800-3100 cm -1 spectral region primarily include lipids 430 and proteins -with some marginal nucleic acid contribution (33, 34 live cells. These vesicles serve as a coarse chemical representation of cells to provide an isolated lipid 442 preparation, free of protein or carbohydrate contribution to vibrational spectra. Infrared-exposed MLV 443 spectra (Figure S4A, B) show distinct shifts in lipid molecule resonances relevant to lipid molecular 444 packing order. Relevant spectral band assignments for biological lipid samples are summarized in Table  445 1. The 2850 cm -1 symmetric aliphatic C-H stretch resonance is markedly decreased, along with its Fermi 446 resonance at 2880 cm -1 . Meanwhile, sp 2 C-H stretching resonances associated with unsaturated aliphatic 447 chain motifs at 3010 cm -1 are substantially decreased. Crucially, ratiometric comparison of 2880 and 2850 448 cm -1 shows reduced rotational restriction in alkane chains, or a decrease in aliphatic tail packing order 449 within the hydrophobic region of the membrane (Figure S4C). This is further supported by the observed 450 decrease in the ratio of 2940 to 2830 cm -1 , which relates to increases in the solvent interaction with lipids 451 (Figure S4C). These observations suggest that thermodynamic changes in lipid vibrational signatures 452 during IR stimulation are discernable with hsSRS. 453 To characterize protein vibrational signature changes during IR-induced heating, the edge of a 454 10%w/v BSA solution meniscus was imaged with hsSRS using radiant exposures equivalent to threshold 455 levels of IR exposures in live cells ( Figure S4D). Changes in protein spectra during IR exposure appear 456 to be negligible ( Figure S4E). Furthermore, the contribution of protein vibrational spectra in ratiometric 457 comparisons that reveal significant changes in MLV samples appear to contribute negligibly to IR-458 exposed changes in the BSA sample ( Figure S4F) The resultant area-normalized hsSRS spectra of NG108 cells under baseline (unstimulated), 481 subthreshold, and threshold stimulation conditions are shown in Figure 3B. Relevant spectral band 482 assignments for biological samples in the CH-stretch region are summarized in Table 1. Shoulders 483 appearing at 2850 cm-1 during stimulation are indicative of relatively increased vibrational resonant 484 activity from symmetric aliphatic C-H stretching in lipid tail chains. Decreases in the relative intensity 485 ratio between 2940 and 2885 cm -1 (Figure 3C) suggest a decrease in packing order within the 486 hydrocarbon tails of the lipid molecules due to trans-gauche isomerization of sp 3 hydrocarbon chains. 487 Interestingly, the 2850 cm -1 shoulder appears to increase in spectral intensity relative to the associated 488 Fermi resonance at 2880 cm -1 , possibly suggesting a reduction of intermolecular steric hindrance between 489 aliphatic lipid tails, or more rotational freedom of hydrocarbon chains. These observations were 490 quantified by calculating the intensity ratio between 2850 and 2940 cm -1 (Figure 3D), as well as 2880 and 491 2850 cm -1 (Figure 3E). These metrics respectively offer a quantification of lipid tail chain packing order 492 -which was previously hypothesized to decrease during IR stimulation (3). Figure 3C-E shows these 493 intensity ratios from NG-108 whole cell spectra obtained at baseline, sub-threshold, and threshold levels 494 of INS previously established to elicit calcium transients. Statistically significant differences (p < 0.05, 495 indicated with asterisk) in these ratios suggest decreased hydrocarbon tail chain packing in cellular lipid 496 membranes. Notably, in each comparison, the ratios calculated for subthreshold exposure fall between 497 unstimulated and stimulated conditions. Of particular note, the shoulder around 3030 cm -1 -which is a sp 2 498 CH (methylene) resonance assignable to CH bonds at points of unsaturation in lipid hydrocarbon tails -499 appears at the threshold stimulation but is reduced in the subthreshold and no stimulation cases ( Figure  500 3B). 501 The hsSRS spectral acquisition as described above requires cells to be exposed to 50 different 502 rounds of IR stimulation -possibly damaging the cells and yielding biologically irrelevant observations. 503 Though no morphological changes were observed in the stimulation experiments, cell viability was 504 verified after repeated IR exposure. Exposed NG108 cells were imaged with multiphoton fluorescence to 505 track the uptake of a cell damage indicator -propidium iodide (PI) -simultaneously with SRS tuned to 506 the 2940 cm -1 CH3 resonance. Cells were imaged through 50 rounds of stimulation, using parameters 507 similar to those used during a live cell hsSRS imaging experiment ( 508 509 Figure S5A). Some cell swelling was observed morphologically, but no uptake of PI was observed ( 510 511 Figure S5B)-suggesting that the repetitive nature of hsSRS acquisition did not have any immediate 512 impact on acute cell viability. 513 514 lipid bilayer packing order 515 Ratiometric fluorescence of di-4-ANNEPS emission, a probe of membrane packing order, was employed 516

Ratiometric fluorescence imaging of functional lipid dye during INS verify changes in
to verify cellular lipid dynamics as observed in vibrational spectra (25). Di-4-ANNEPS rotoisomerization 517 is known to be dependent on fatty acid tail chain packing order in lipid membranes. During IR 518 stimulation, if lipid tail chain packing order is decreased, a similar decrease in general polarization (GP) 519 metric should follow. In place of the conventional approach for calculating GP, intensity-invariant 520 adaptation of GP was utilized to circumvent the defocusing effect during IR stimulation (detailed in 521 Methods and Figure S6). Figure 4A depicts an intensity image of di-4-ANNEPS loaded NG-108 cells 522 overlaid with color denoting GP calculation at each pixel. Figure 4B  Our demonstration of characterizing and precompensating for dynamic defocus during INS with hsSRS is 540 a novel approach in biomedical microscopy that is applicable to studying the molecular biophysics of live 541 cell models more generally. 542 543 Photothermal events are notoriously difficult to address with biological microscopy due to the 544 relationship between temperature and refractive index in water. While bulk changes in sample 545 temperature can impact optical aberrations in microscopes, spatial thermal gradients that vary on the order 546 of the microscope's field of view can have significant impacts on the refraction of light into the sample 547 ( Figure 2B). Accounting for defocusing actively on millisecond timescales may be possible with dynamic 548 adaptive optics approaches but is far from trivial to implement. Instead, our approach to adjust for IR-549 induced defocusing of the fluorescence excitation empirically (Figure 2A, C) -though coarse compared 550 to adaptive optics -enables us to gather useful insight to the biophysical phenomena associated with INS 551 (Figure 3). The reliable timing of stimulation can be leveraged to employ a time-resolved spectroscopy 552 approach to hsSRS imaging at high framerates. In doing so, we demonstrate that high-speed vibrational 553 dynamics can be resolved in live cell preparations safely to yield biologically meaningful observations. In 554 studying INS using high numerical aperture microscopy, where IR induced deflections in focal length can 555 equal or exceed the depth of focus of a particular imaging objective, we urge others to interpret their 556 results cautiously. Thermal defocusing can have a disproportionate impact on single-channel 557 intensiometric-based measurements and need to be carefully accounted for (Figure S6). In cases where 558 intensity noticeably changes during exposure, we encourage others to employ ratiometric or multi-559 spectral approaches to allow for defocusing artifacts to be readily accounted for. With fluorescence 560 microscopy, where quantum yield, fluorescence intensity, and spectral profiles are well known to be 561 sensitive to both heating and defocusing (39-41), having simultaneous or time-resolved multispectral 562 reference bases will allow for such artifacts to be accounted for in post-processing. 563 564 There are several spectral changes in the CH-stretch region of the Raman spectrum (2800-3100 cm -1 ) that 565 one might expect to see if the current INS mechanistic model was valid. Trans-gauche isomerization, or 566 rotoisomerization, of sp 3 hydrocarbon chains -primarily associated with lipid hydrophobic tail groups in 567 Raman imaging -can give rise to a number of steric effects that drive lipid membrane deformations (36, 568 37, 42, 43). Specifically, lipid packing order -or the ability for lipid molecules to stack neatly alongside 569 each other within the membrane leaflets -was hypothesized to decrease with elevated temperature during 570 INS. Rotoisomerization in membrane lipids geometrically shortens acyl tail groups, resulting in 571 membrane thinning. While quantifying the absolute deformation of lipid membrane thickness with SRS 572 would require additional calibration experiments, relative indicators of molecular interactions can be 573 quantified with hsSRS. An increased quantity of gauche rotamer within the hydrophobic region of the 574 membrane leads to geometric acyl tail shortening and sterically drives lipid molecules apart from each 575 other. The result is a decrease in membrane packing order. In the CH-stretch region of the Raman 576 spectrum, relative changes in symmetric (2850 cm -1 ) and asymmetric (2880 cm -1 ) aliphatic C-H stretching 577 indicate shifts in molecular packing order due to changes in the rotational freedom of hydrocarbon chains 578 in lipid tails. Raman signal at these resonances is largely attributed to biological lipids ( Figure S4) (33). 579 A decrease in the ratio of 2880 and 2850 cm -1 during INS (Figure 3E) is indicative of a 'loose' packing 580 order between lipid molecules or an increase in trans-gauche isomerization (44,45). The 581 rotoisomerization of lipid tails is well known to both decrease membrane thickness and increase the area 582 of each lipid molecule's solvent interactions (46-48). Changes in the ratio between 2940 and 2885 cm -1 583 offer insight to water interaction with lipid molecules, which should increase with temperature. The data 584 show a decrease in the ratio between 2940 and 2885 cm -1 (Figure 3C), which is in line with the idea that 585 lipid molecules expand within the membrane leaflets to leave room for more potential solvent interactions 586 (e.g. hydrogen bonding) with elevated temperatures. The IR dose dependance of this observation further 587 suggests that the relative degree of isomerization correlates with levels of IR exposure that would evoke 588 neural activity in vitro. The observations of a progressive increase in isomerization with IR exposure 589 support the existing mechanistic model of INS, where transient temperature changes are accompanied by 590 changes in physical bilayer geometry. 591 592 The shoulder appearing around 2990 and 3030 cm -1 during INS in cells (Figure 3B) arise from relative 593 increases in vinyl C-H resonances, which correspond to points of unsaturation in lipid tail acyl chains. 594 Relative increases in vinyl C-H signal can arise from reduced steric hinderance of sp 2 C-H stretching as 595 well as compositional or membrane potential related changes when the lipid bilayer undergoes thermal 596 changes. Curiously, the appearance of the 3030 cm -1 shoulder in threshold stimulated cells was reduced in 597 sub-threshold levels of stimulation. This resonance at 3030 cm -1 may provide a key marker for neural 598 biophysics during INS. 599 600 The vinyl C-H portion (2980-3100 cm -1 ) of the C-H stretch region does contain SRS signal contributions 601 from proteins-particularly from amino acid residues such as tyrosine, phenylalanine, and tryptophan. 602 These amino acids play a key structural role in stabilizing hydrophobic domains of transmembrane 603 proteins in the cell membrane. Control experiments observing the IR-related dependance of BSA SRS 604 spectra in solution ( Figure S4) as well as evidence from others (36, 37, 49) reinforce that thermally-605 mediated changes in protein dynamics are not major contributors in the CH stretch region of the Raman 606 spectrum. As such, we conclude protein signal contributes minimally to the photothermal mediated SRS 607 changes that would be expected during INS. Others have attributed relative decreases in 2930 cm -1 signal 608 to changes in cellular membrane potential, enabling the visualization of neuronal action potentials with 609 SRS microscopy (18, 32). These spectral changes were attributed to the decrease in positively-charge 610 proteins electrostatically accumulating at the extracellular membrane surface when a cell is at its resting 611 membrane potential. A reduction in membrane potential was suspected to reduce membrane-associated 612 proteins in solution at the cell membrane surface. Our results show a considerable reduction in relative 613 2930-2940 cm -1 signal during INS (Figure 3B), thus electrostatic association of soluble proteins with cell 614 surfaces may play some role in our results. Several experimental details suggest that membrane potential 615 changes from electrostatic protein association would not be contributing to our spectra. Defocusing 616 artifacts make it difficult to obtain conclusions about absolute molecular concentrations during INS 617 (Figure 2A-C). Practically, our approach to region of interest selection, non-balanced detection, and 618 imaging medium formulation confounds any comparability of our results with these previous studies. 619 However, Lee et al. did employ a similar time-resolved approach for acquiring SRS spectra as a function 620 of membrane potential -demonstrating the utility of such an approach for certain types of experiments 621 beyond photothermal phenomena. 622 623 The physical changes in the lipid bilayer during rapid heating with IR light are thought to give rise -at 624 least in part -to the cell capacitance increase that drives cellular depolarization during INS (2, 3). Our 625 results (Figure 3) support the idea that the lipid bilayer undergoes some thermally mediated chemo-626 physical change during INS that is observable via vibrational imaging and correlate with the level of 627 delivered stimulus. While these findings are promising, they do not definitively support that bilayer 628 deformation is directly causal to the stimulatory effect of INS. Though beyond the scope of this work, 629 questions remain about how transmembrane ion channels may be independently sensitive to lipid 630 membrane geometry and thermodynamics. Lipid thermodynamics are known to affect the conformational 631 and functional properties of transmembrane ion channels (50-53). It is not clear whether the capacitive 632 effect or the actual physical change in the lipid bilayers themselves give rise to stimulatory phenomenon. 633 It is difficult to decouple chemo-physical and thermal electrodynamic changes in biologically relevant 634 preparations. A preparation of lipid vesicles or cells expressing voltage gated ion channels loaded with a 635 UV photo-switchable lipid analogue (e.g., containing an azobenzene moiety in the tail group) may be a 636 useful experiment. The photo-switching property of such synthetic lipids would allow for optical control 637 of membrane packing order with substantially reduced photothermal effects. 638 639 The current hypothesis for how INS occurs is that rapid heating causes a capacitive inward current that 640 can depolarize neurons and lead to action potential generation (2). This capacitive current is thought to 641 arise from biophysical changes within the extracellular membrane -namely trans-gauche isomerization 642 of lipid acyl tail chains -that change the physical dimensions of the extracellular membrane due to 643 temperature elevations (3). This deformation is accompanied by a movement of membrane-associated 644 charge that -when hot and fast enough -can generate an inward current that depolarizes cells. The model 645 of this phenomenon relies on steady-state chemical assessments of synthetic lipid bilayer geometry (54, 646 55). The changes in bilayer geometry are used to inform a computational electrodynamic model that is 647 compared against previous experimental work (2, 38). While the model of chemo-physical and 648 electrodynamic phenomena convincingly reproduces experimental data, capacitance changes and cellular 649 electrodynamics are ultimately influenced by more than lipid dynamics alone in vitro and in vivo. While the implementation of hsSRS here can resolve high speed spectral dynamics well below a second, 669 it does take several minutes to build observations of events on a spectral basis. In situations where 670 repeated perturbation of cells is not practical, the same approach can be implemented with a drastically 671 reduced number of spectral channels. Alternatively, multispectral approaches leveraging simultaneous 672 acquisition of multiple resonances would be advantageous. To account for the defocusing artifacts 673 described here, at least two spectral channels need to be acquired to accurately draw conclusions -thus 674 single-shot perturbations are not readily applicable with the demonstrated approach here. The fast rates of 675 development in bioimaging with SRS show promise in pushing SRS based imaging methods to their 676 limits. Our work shows that hsSRS can be applied to a range of lipid biophysics experiments as a 677 complement to more conventional fluorescence-based approaches. In contrast to fluorescence-based 678 approaches that rely on indirect readout from reporter molecules interacting with lipids in the cell 679 membrane, vibrational contrast like that of hsSRS enables direct inference to be made specific to lipids at 680 the intra-and intermolecular levels. the signal-to-noise performance in the fingerprint window of the Raman spectrum (400-1700 cm -1 ). 698 Utilizing other features of the Raman spectrum that are more directly attributed to lipid tail chain 699 rotoisomerization (e.g. the skeletal vibrational C-C modes between 1030 and 1150 cm -1 , as well as C=C 700 stretching modes around 1650 cm -1 ) might provide more direct mechanistic insight to INS once possible 701 (44). Some promising newer spectroscopic and computational denoising methods that circumvent these 702 noise issues are gaining popularity, but still require careful validation for high-speed imaging of cellular 703 dynamics (61-64 analysis, data visualization, and wrote the manuscript. All authors contributed to editing manuscript. 772

773
The authors declare no conflicts of interest. 774

775
Any raw or processed data, processing, and analysis code are available upon request from the 776 corresponding authors. 777 pulsed-IR neurostimulation within the microscope's field of view, it is possible to recover some lost 790 nonlinear signal due to defocusing. Asymmetric C-H stretch of -CH 2 2930 cm-1 corresponds to the overtone of the CH 2 scissoring (δ(CH 2 )) enhanced by Fermi resonance with the v s -(CH 2 ) mode.

-3060
Proteins sp 2 C-H stretch of aromatic/vinyl amino acid residues (═C-H) activation threshold radiant exposures [n = 10-24 cells per group]. Peak ratio comparisons indicative (C) 804 asCH2/asCH3 as a measure of trans-to-gauche isomerization of lipid tail groups, (D) symCH2/symCH3 as a 805 measure of increased polar headgroup association with water due to membrane packing order decrease, 806 and (E) asCH2/symCH2 as an indicator of decreasing acyl chain packing order. *indicates p < 0.05 807 represent SEM across all cells within each condition. * indicates p < 0.05. 816 Figure S1: A) Illumination geometry and B) calculation of approximate fiber distance for estimating 817 radiant exposure -where dfiber is the optical fiber diameter, rfiber is the optical fiber radius, θA is the fiber 818 approach angle, dcs+ is the fiber edge's distance off of the surface of the cover slip, and l is the normal 819 distance from the optical fiber face to the cover slip plane.