Spectral Imaging to Measure Heterogeneity in Membrane Lipid Packing

Physicochemical properties of the plasma membrane have been shown to play an important role in cellular functionality. Among those properties, the molecular order of the lipids, or the lipid packing, is of high importance. Changes in lipid packing are believed to compartmentalize cellular signaling by initiating coalescence and conformational changes of proteins. A common way to infer membrane lipid packing is by using membrane-embedded polarity-sensitive dyes, whose emission spectrum is dependent on the molecular order of the immediate membrane environment. Here, we report on an improved determination of such spectral shifts in the emission spectrum of the polarity-sensitive dyes. This improvement is based on the use of spectral imaging on a scanning confocal fluorescence microscope in combination with an improved analysis, which considers the whole emission spectrum instead of just single wavelength ranges. Using this approach and the polarity-sensitive dyes C-Laurdan or Di-4-ANEPPDHQ, we were able to image—with high accuracy—minute differences in the lipid packing of model and cellular membranes.


Introduction
The bioactivity of the cellular plasma membrane is correlated to membrane topology andl ocal variations in lipid-protein compositions. [1][2] Consequently, cellular signalling is often accompanied by molecular re-organization at the membrane, and the role of physicochemical parameters in this is challenged. In particular, changes in the laterald ensity and order of the lipids (denoted lipid packing) are regarded to play ap ivotal role, since thesem ay introduce coalescence and conformationalc hangeso fp roteins,a nd thus compartmentalization of cellular signaling. [3] Membrane-embedded fluorescent molecules such as Laurdan [4] and Di-4-ANEPPDHQ [5][6] are employed to reporto nt he relative levels of membrane lipid packing. [7] The fluorescencee mission of these reporterss hifts depending on the order of the immediate membrane environment. In the case of Laurdan, the emission spectrumi sc haracteristicf or water dipolar relaxation processes (i.e. the water content)i n the vicinity of the probe,w hich is ap roperty sensitive to the membrane lateral packing. [8] Parasassi et al [4,[9][10] determined al arge red shifti nf luorescence emission of Laurdan when comparing its emission spectrum between as ingle component gel-like and al iquid-crystalline lipid model membrane. The most extensivelyu sed measure for relative levels of lipid packing is the generalizedp olarization (GP) parameter,w hose value is arelative index of lipid packing based on red-or blue-shifted emissiono ft he aforementioned probes. [9][10][11] Initially,G Pw as thoroughly studied employing the whole set of information derived from both the excitation and emission spectra. [9] Thereafter,t os implify the calculation, GP generally has been calculated by using the fluorescencei ntensitiesd etected for two specific wavelengths.T hese wavelengthsa re usuallyc hosen as the wavelengths l Ld and l Lo of maximum emissiono ft he probe in ar eference liquid-disordered (Ld) and liquid-ordered (Lo) membrane environment, respectively.F or Laurdan, these are, for example, l Ld = 490 nm (red-shifted) and l Lo = 440 nm (blue-shifted). [12] C-Laurdan, as used in this study,i sabrighter and more sensitive derivativeo fL aurdan, which is based on the same principle and works in the same wavelength region. [13] For Laurdano rC -Laurdan, the GP values are calculated by using the fluorescences ignal intensities I R and I B at the red-and blue-shifted emission wavelengths l Ld and l Lo , respectively [Eq. (1)]: The GP values for othermembrane-sensitivedyes can be calculated in as imilarw ay,w ith l Ld and l Lo defined at different wavelengths. [7,14] Usually,t he spectra are precisely recorded on af luorescences pectrophotometer. [4] Such measurements, however,o nly give values averaged over the whole sample. In many cases,s uch as when observing living cells, it is desirable to determine the spatial heterogeneity of the GP values, that is, to image the sample and calculate the GP at every image Physicochemical properties of the plasma membrane have been shown to play an importantr ole in cellular functionality. Amongt hose properties, the molecular order of the lipids, or the lipid packing,i so fh igh importance.C hanges in lipid packing are believed to compartmentalize cellular signaling by initiating coalescence and conformational changes of proteins. A common way to infer membrane lipid packing is by using membrane-embedded polarity-sensitive dyes, whose emission spectrum is dependent on the molecular order of the immediate membrane environment. Here, we report on an improved determination of such spectrals hifts in the emissions pectrum of the polarity-sensitive dyes. Thisi mprovementi sb ased on the use of spectral imaging on as canning confocal fluorescence microscope in combination with an improved analysis, which considers the whole emission spectrumi nsteado fj ust single wavelength ranges. Using this approach and the polarity-sensitive dyes C-Laurdan or Di-4-ANEPPDHQ, we were able to image-with high accuracy-minute differences in the lipid packingo fm odel and cellular membranes.
pixel. [15] Usually,t his is achieved by recording the detected fluorescencee mission in two discrete wavelengthr anges centered around l Ld and l Lo ,f or example, using two detectors and band-pass filters. [6,12,[15][16] Here, we describe an accuratem ethod for observing spatial heterogeneity in lipid packing or GP values by recording the whole emission spectrum for each image pixel. In such as pectral (or lambda) imaging mode, ac onfocal scanning microscope is equipped with ad iffraction grating, prism or acoustic-opticale lement, which spectrally disperses the collected fluorescencee mission into multiple detection channels on, for example, ana rray of gallium arsenide phosphide (GaAsP) detectors( in our case 32-channel). Thus, it is possible to simultaneously record images over aw ide range of distinct wavelengths, and to generate emission spectra for each image pixel with < 10 nm spectral accuracy( Figure1). Using ac ustom Fiji/ ImageJp lug-in, we demonstrate an improved determination of GP values by modelingt he spectrald ata at each image pixel using aG aussian or Gamma Variate distribution. This allows us to precisely observe differences in lipid packing over space of phase-separated giant unilamellar vesicles (GUVs) or cell-derived giant plasma membrane vesicles (GPMVs), [17] and cellular plasma membranes following cholesterol depletion. Spectral or lambdai maging modesa re realized on most recent confocal microscopes. Consequently, the use of state-of-the-art confocal systemst ogether with the presented advanced calculation of GP values using the provided Fiji/ImageJ plug-in (curve fitting of the full spectra) is straightforwarda nd has the potential to gain accurate insights into membrane heterogeneity and bioactivity.O ur approachi st herefore more general comparedt o ar ecently published approach, which is based on phasor analysis of spectral images recorded for Laurdano natwo-photon microscope. [18] 2. Results and Discussion

Spectral GP Imaging of Model Membranes
We first exemplified GP spectralimaging on model membranes such as giant unilamellar vesicles (GUVs) composed of different mixtureso fl ipids. We have chosen C-Laurdan [13,19] as the environment-sensitive probe due to its common use and its in-creasedf luorescenceb rightness and increased sensitivity in GP determination. [12][13]20] The equatorial planeso ft he vesicles dopedw ith C-Laurdan were imaged using ac onfocal microscope. [19] The fluorescencee mission of C-Laurdan was excited using a4 05 nm laser and detected between 415 and 700 nm by a3 2-channel GaAsP detector at % 8.9 nm wavelength intervals. Figures 2A and 2B show 20 (of the 32) image slices recorded for the equatorial plane of representative single-component dioleoyl-sn-glycero-3-phosphatidylcholine (DOPC) and two-com-ponentSM:Chol [saturated sphingomyelin (SM) and cholesterol (1:1) mixture] GUVs, respectively.Eachs lice exposes the spatially resolved fluorescencesignal detected at 20 (of the 32) different wavelength ranges between 415 and 584 nm, anda ll 32 image slices allow generating emission spectra for each image pixel ( Figure 2C). It is worth noting that we miss as mallf raction of the C-Laurdan spectrum (between 395-415 nm) by using ac onventional microscope equippedw ith a4 05 nm laser as the most bluish laser and ad etector with ac ut-off at 415 nm. However,t his missed fraction does not influence our GP analysis,s ince the first point of the obtained spectrum at 415 nm is very close to zero( Figure 2C)a nd the GP parameter gives relative values only (as pointed out further on), that is, even if biased,o ne only has to be consistent within one experimental study.
The DOPC and SM:Chol GUVs mimicked anL da nd Lo lipid environment, respectively.T his was observed by the blue-shifted fluorescence emission in the case of the Lo environment (SM:Chol). The observed spectra are very similar to the reported spectra of C-Laurdan in these membranes. [12][13] From these spectra,w ec alculated the GP values for each image pixel [Eq. (1)] with as ufficient amount of fluorescences ignal using the directs ampling option of our customF iji plug-in (see Materials and Methods section and Figures S1 and S2 in the Supporting Information, SI). The GP values were very homogeneously distributedo ver the GUVs ( Figures 2D and 2E)w ith af airly disordered (GP %À0.55) and ordered (GP % 0.5) environment for the DOPC and SM:Chol GUVs,r espectively (Figure 2F). The GP values are close to valuesp reviouslyr eported for C-Laurdan in liposomes of the same compositions, [12,21] and correlated as expectedw ith changing the molecular order or lipid packing of the GUVs by varying the lipid content (DOPC, Figure 1. Spectral GP imaging:Fluorescence emission from the sample is generated by illumination with laser light. The collected fluorescencelight is dispersed (using aprism or other device) and guided onto several parallel detectors,such as asensitive 32-channel GaAsP detector.Each channelofthe detector recordsasignal at different wavelengths at (in our case)a pproximately 8.9 nm wavelengthi ntervals. This signal is then used to generate the emissionspectrum of the fluorophore for each image pixel (only 7o ft he 32 channels are shown), giving precise valueso fthe intensities I B and I R (bluea nd red arrows).  Figure S3).

Curve Fitting:I ncreasing the Accuracy of GP Imaging
For the calculation of the GP values [Eq. (1)] in Figure 2, we used ad irect sampling approach, that is, the intensity values I R and I B were approximated from the image slices registering fluorescencee missionf or the wavelength ranges most closely matching l Ld and l Lo ,r espectively (see Materials and Methods). To be comparable to previouss tudies, we verified that the use of wavelength intervals (as done in previous studies), instead of two discrete wavelengths (as done here), yieldeds imilar results for the GP values ( Figure S4). We now anticipated increasing the accuracyo fd etermining I R , I B and thust he GP values by employing information from the whole emission spectrum, that is, from all data points. For this, we fit aG aussiano r aG amma Variate distribution to the emission spectrum captured for each pixel, as shown in Figures 3A-C for C-Laurdan in phase-separated GUVs. From the fit, the values of I R and I B can be determined with arbitrary wavelengtha ccuracy.I naddition, extracting values from the fits, rather than just directly sampling the raw data, has ad e-noising effect. TheG Pi mages determined for the curve fitting are visibly de-noised compared to the direct sampling analysis ( Figure 3D). This becomese ven more obvious when comparing the distribution of GP values over images of GUVs from the direct sampling (nonfitting) with that of the curve-fitting analysis ( Figure 3E). While the GP histograms generated from each of the approaches are peaking at similar values, the GP histograms from the curve fitting are more discrete. This is an atural consequenceo fh aving used the whole spectrum to infer the fluorescence value rather than sampling the raw data, and is an improvemento ver the existing GP analysism ethods, whichc an often require significant smoothing prior to GP calculation. [22]  In addition, we experienced that the Gamma Variate fitted the C-Laurdan spectra much better than the Gaussian distribution, especiallyw hen the spectrum at that pixel was skewed, as is often the case in the orderedv esicle phase ( Figure 3B). This is also revealed in the average valueso ft he coefficient of determination (R 2 = 0.86 for Gaussian and 0.94 for Gamma Variate fitting, p < 0.05, Student'st -test, n = 5), which we have used for checking the accuracy of the fits, and whose value is one for ap erfect fit and decreases for increasingly inaccurate fitting. As ac onsequence, the distribution of GP values following the GammaV ariate fitting is slightly more discrete than that resulting from the Gaussian fits ( Figure 3E), especially for the ordered phase. The Gaussian fitting was, however,m uch quicker (% tenfold), and is thus ag ood compromise between accuracya nd analysis speed.

Spectral GP Imaging of Cell-Derived Model Membranes
Much attention in the field of biophysical membrane research has been given to the investigation of cell-derived vesicles such as GPMVs. [23] GUVs with only am ixture of af ew lipids are limitedi nc omplexity when compared to ac ellular plasma membrane with its high diversity in lipids and proteins, [24] which is not the case for GPMVs, since they are directly generated from cellular plasma membrane. [23] Despite the increased molecular complexity,G PMVs can also show the coexistence of more ordered andd isordered phases, [17,23] although with am uch lower differencei nm olecular order or GP values between the phases compared to GUVs. [24] Figure 4d epicts how this minute difference in lipid packing is accurately exposed by spectralG Pi maging, taking GPMVs derived from RBL (rat basophilic leukaemia)c ells as an example. Figure 4s hows the results of our GP analysis foran on-phase-separated and www.chemphyschem.org ap hase-separated GPMV.A se xpected: [8,25] 1) the shift in wavelengths and GP values betweeno rdered and disordered phase is much less pronounced than in GUVs (ordered:G P % 0.2 versus % 0.4 in GUVs, and disordered:G P %À0.15 versus %À0.5 in GUVs, compareF igures 2a nd 3), [17] and 2) the lipid packing in non-phase-separated GPMVs (GP %À0.05) is in-between that of the Lo and Ld environments of the phase-separated GPMVs ( Figure 4C,D). These minor differences in lipid packingo rG Pv alues are more accurately explored using spectral GP imaging in combination with our curve-fitting approach,a sh ighlighted in Figure 4E.P lease note that (as for the results from the GUV experiments)t he Gamma Variate fitting gives more reliable results than the Gaussiana nalysis; especially the GP value of the Lo environmenti sm ore closer to the value of the conventionald irect samplinga nalysis.

Spectral GP Imaging of the Live-Cell PlasmaM embrane
Next, we applied spectralG Pi maging to live-cell plasma membranes. Unfortunately,C -Laurdan is strongly internalized in living cells, [7] despite the cell labeling on ice, which makes the determination of GP maps less accurate ( Figure S5). Fortunately,s pectral imaging allows the straightforward implementation of other polarity-sensitive probes with different emission spectra as it obviates probe-specific filter sets. Here, we used the membrane-dyeD i-4-ANEPPDHQ, [5,14] whichs hows significantly less internalization in living cells. [7] Its fluorescencee mission was excited at 488 nm and recorded between 495 and 691 nm. For this dye, we picked l Ld = 565 nm and l Lo = 610 nm, following previous work [7] (see Materials andM ethods). Spectral imagingu sing this probe in GUVs and GPMVs yielded aG P value of À0.2 (with an emission maximum of 610 nm) in liquid-disordered DOPC GUVs and 0.5 (with an emission maximum of 565 nm) in liquid-ordered SM:Chol GUVs ( Figure S6). However,t he selection of l Ld and l Lo for Di-4-ANEPPDHQ is rather arbitrary,w ith another choicer esultingi no ther GP values [14] (see Materials and Methods). Also, the GP values for C-Laurdan and Di-4-ANEPPDHQ are different due to different dependencies of their respective fluorescencee mission on lipid packing (note that the spectrals hift between the Lo and Ld conditions is much less pronouncedt han for C-Laurdan). In an utshell, the GP values are relative values, and one only has to be consistent within one experimental study.
Generally,t he sensitivity of Di-4-ANEPPDHQ on changes in lipid packing is lower than that of C-Laurdan, [7] which makes an accurate determinationo ft he intensity values I R and I B even more essential. Figure 5s hows the spectralG Pi mage analysis of Di-4-ANEPPDHQ labeled live RLB cells, with an average GP value of %À0. 16. As expected, [26] lipid packing is further reduced (GP %À0.26) after cholesterol depletion of the cellular plasma membrane using treatment with methyl-beta cyclodextrin (MbCD). Again, the GP histograms following the curve-fitting analysis are much more discrete, allowing for am uch better distinction of lipid packing between the two conditions ( Figure 5D). In contrast to C-Laurdan, for Di-4-ANEPPDHQ the Gaussianf itting produces the most discrete distribution of GP values, but again with slightly different peak values compared to the Gamma Variate fitting and direct sampling approach. As mentioned before,t he GP values are relative values, ando ne has to be consistent within one experimental study,h ere also with respect to the analysis procedure.

Conclusions
Lipid packing is considered an important organizational and functional aspect of the cellular plasma membrane. [1][2] Accurate measurements of lipid packingo fm odel and cellular membranes allow gaining insights into physicalc hemistry properties of the membranes and relating these to their bioactivity.F luorescent polarity-sensitive probesi nc ombination with generalizedp olarization (GP) analysis are often used to quantify the lipid packing. While the determination of GP values from af luorescences pectrophotometer is straightforward anda ccurate,t he imaging of the spatial variation in lipid packing is more challenging. Here, we approach this challenge by spectralG Pi maging using the possibility of state-of-the-art confocalm icroscopest os imultaneously detectt he spectrally dispersed fluorescencesignal on different detectors. The resulting image stack (where each slice represents the fluorescence signal detected within as mall wavelength interval) enables the instantaneous setup of fluorescence emission spectra for each image pixel. The spectrald ispersion of the signal can either be achieved using ap rism and a3 2-channel GaAsP line-array detector,a si nt his case, or gratings or acoustic-optical elements in conjunction with an array of multiple detectors. Using ac ustom-designed Fiji/ImageJp lug-in,w es howedt hat this spectralG Pi maging approach revealed as traightforwardw ay for observing spatial heterogeneity in lipid packing of model as well as cellular membranes. Specifically,w ed emonstrated that heterogeneity in lipid packing can more accurately be determined by using the whole spectrum to infer fluorescence values at discrete wavelengths for GP calculation rather than using raw data. More importantly,d ue to itsa bility to yield as ingle pixel spectrum, we believe that spectrali maging will in the future enable us to overcome the two-wavelength GP calculation limitation and to use the whole spectrum to understand the dipolar relaxation phenomenoni nt he context of biological membrane ordering.

Giant Unilamellar Vesicle Preparation
GUVs were prepared as previously described. [17] Briefly,1mg mL À1 lipid solution was prepared in chloroform. Then, 5 mLo ft his solution were dried onto two parallel platinum wires mounted in ac ustom-built GUV Te flon chamber. [24] A3 00 mm sucrose solution was added to the chamber and a1 0Hzc urrent was applied to the wires for an hour. [24] GUV preparation of SM:Chol, DPPC:Chol or DOPC:SM:Chol mixtures was performed at 70 8Ca bove the respective lipid transition temperature. Other GUVs (such as pure DOPC) were formed at room temperature.

Cell Maintenance
RBL cells were grown in 60 %R PMI, 30 %M EM and 10 %F CS medium. They were seeded out two days before the experiments so that they would reach 70 %confluence on the day of the experiment.

Cholesterol Modulation
For cholesterol removal, 0.12 gm ethyl-beta cyclodextrin was dissolved in 10 mL MEM (10 mm). Then the cells seeded out on 8-well glass bottom Ibidi slide (#1.5) chambers were incubated with 250 mL/well of this suspension for 30 min at 37 8C. Cells were washed with phosphate-buffered saline (PBS) af ew times before labeling.

Membrane Labeling with C-Laurdan and Di-4-ANEPPDHQ
The GUVs and GPMVs were spiked with C-Laurdan or Di-4-ANEPPDHQ at afinal concentration of 0.4 mm at room temperature. The cells were incubated with a0 .4 mm probe solution in PBS for 5min on ice to reduce the internalization of the probes. Then, the cells were washed af ew times with PBS. 250 mLo fc ell media without phenol red and serum was added onto the cells for imaging. Samples were imaged in 8-well glass bottom Ibidi chambers (#1.5).

Confocal Spectral Imaging
Spectral imaging of the different membrane samples was performed on aZ eiss LSM 780 confocal microscope equipped with a3 2-channel GaAsP detector array.L aser light at 405 and 488 nm was selected for fluorescence excitation of C-Laurdan and Di-4-ANEPPDHQ, respectively.T he lambda detection range was set between 415 and 691 nm for C-Laurdan, and between 495 and 691 nm for Di-4-ANEPPDHQ ( Figure S1). The wavelengths 415 and 691 nm were the ultimate limits of our detector.D espite the fact that wavelength intervals of down to 4nmc ould be chosen for the individual detection channels, we have set these intervals to 8.9 nm, which allowed the simultaneous coverage of the whole spectrum with the 32 detection channels ( Figure S1). The images were saved in .lsm file format and then analyzed by using ac ustom plug-in compatible with Fiji/ImageJ, as described further on.

Obtaining Spectra from the Spectral Images
The spectra for each image pixel were obtained from the intensity values of the 32 different detection channels by using the ImageJ plug-in "Stacks-T functions-Intensity vs. Time Monitor". We usually applied the plugin only on pixels within ar egion of interest of the acquired images. The background signal was determined by applying the same plugin on ad ark region (of the same size) of the image, and subtracted from the signal from the region of interest.

Definition of l Ld and l Lo
For the calculation of the GP values [Eq. (1)],o ne has to define the wavelengths l Ld and l Lo of maximum emission of aprobe in areference liquid-disordered (Ld) and liquid-ordered (Lo) membrane en-vironment. Following the literature, for C-Laurdan we have chosen l Ld = 490 nm and l Lo = 440 nm. [27] For the other probe, Di-4-ANEPPDHQ, we have chosen l Ld = 605 nm and l Lo = 565 nm, as defined in ap revious work. [7] However,t he wavelength selection for Di-4-ANEPPDHQ is rather arbitrary,a nd ac hoice of am ore blueshifted l Lo and amore red-shifted l Ld may increase the GP contrast. Therefore, GP values are relative values;t hat is, the GP values determined via C-Laurdan are different from those determined via Di-4-ANEPPDHQ, or when choosing other wavelengths l Ld and l Lo . One only has to be consistent within one experimental study.
GP Calculation Using the Plug-in AJ ava Fiji plug-in was developed for an efficient analysis of the acquired spectral data (plug-in is available at https://github.com/ dwaithe/GP-plugin). This plugin facilitates calculation of the GP equation, with direct sampling of the original wavelength intervals or through fitting of the entire available spectrum at each pixel using ac urve-fitting algorithm. In any case, the plug-in determined discrete intensity values, I R and I B ,f or the detected signal at l Ld and l Lo ,and calculated GP values according to Equation (1).
As ap reprocessing step for both techniques, the spectral image stacks are integrated via Z-projection and then masked using auser-defined threshold on an 8-bit scale (0-255). This step creates am ask that is used to restrict subsequent analysis to areas positive for fluorescence above the background only.F urthermore, an optional step for the user is to include an average background subtraction correction in the pre-processing. If this option is selected, the background signal intensity is calculated from the dark areas of each image of the whole spectral stack (individual slices), and this value is then subtracted from each slice in the areas selected for the positive signal. This measure is included to correct minor variations in the background signal levels between slices.
Every slice in the image stack represents ac ertain wavelength l, usually the central wavelength in the corresponding 8.9 nm-wide wavelength interval, and values of intensity I l at different wavelengths l,t hat is, the emission spectra are obtained from the preprocessed image stack for each pixel. At this point, correction factors can be included, accounting for ap otential difference in sensitivity between the different spectral detection channels (which was not required in our case, since the microscope supplier guaranteed pre-normalization of the different channels) ( Figure S2). From this I l distribution, the plug-in determines I R and I B for each pixel in the masked image by:1 )selecting I l values for aw avelength range most closely to l Ld and l Lo (direct sampling), or 2) I R and I B modelling using curve-fitting, where the ImageJ "CurveFitter API" is applied to fit either the Gamma Variate or the Gaussian no-offset distribution to the I l distribution. To judge the quality of the fit, the R 2 measure is extracted and only pixels above au ser-defined noise tolerance are included (0.0-1.0, 1.0 being the best fit). Upon successful fitting, inference is performed on the generated distributions to precisely (with 1nmp recision) determine l Ld and l Lo and from that obtain I R and I B .B efore final calculation of the GP values using Equation (1), the values of I R and I B are normalized through division by 255.
Application of the above calculation produces as patial GP map representing the GP value for each pixel of the image. The final stage of the plugin calculates and outputs ah istogram of the GP map and applies ac ustom look-up-table (LUT) to the data. Along with the histogram and GP map, the plug-in also outputs general histogram statistics along with median values for the positive and negative phases of the GP distribution, the selected l Ld and l Lo ChemPhysChem 2015, 16,1387 -1394 www.chemphyschem.org input wavelengths (either from the original input stack or from inference), the calculated mask image, and also, for the curve-fitting, the goodness of fit image, where each pixel is labeled with its R 2 value.