Multi-Color Quantum Dot Tracking Using a High-Speed Hyperspectral Line-Scanning Microscope

Many cellular signaling processes are initiated by dimerization or oligomerization of membrane proteins. However, since the spatial scale of these interactions is below the diffraction limit of the light microscope, the dynamics of these interactions have been difficult to study on living cells. We have developed a novel high-speed hyperspectral microscope (HSM) to perform single particle tracking of up to 8 spectrally distinct species of quantum dots (QDs) at 27 frames per second. The distinct emission spectra of the QDs allows localization with ∼10 nm precision even when the probes are clustered at spatial scales below the diffraction limit. The capabilities of the HSM are demonstrated here by application of multi-color single particle tracking to observe membrane protein behavior, including: 1) dynamic formation and dissociation of Epidermal Growth Factor Receptor dimers; 2) resolving antigen induced aggregation of the high affinity IgE receptor, FcεR1; 3) four color QD tracking while simultaneously visualizing GFP-actin; and 4) high-density tracking for fast diffusion mapping.


Introduction
Cellular signal transduction is regulated through the interactions and output of membrane receptors. However, the study of molecular-level protein-protein interactions in living cells continues to be a challenging goal in cell biology. While the nanometerscale distribution of receptors can be observed using electron microscopy (EM), this technique is limited to fixed samples [1][2][3][4][5][6]. Super-resolution microscopy techniques [6][7][8][9] can also reveal nanometer-scale protein distributions, yet these measurements are inherently slow, making the observation of dynamic events difficult. Therefore, new live cell microscopy techniques are needed in order to determine the role of receptor dynamics in the regulation of cell signaling.
One method for studying the dynamics of membrane components is single particle tracking (SPT) [10], in which a fluorescent or light scattering probe is attached to the component of interest and its position is recorded over time. The probe can be localized with precision much better than the diffraction limit (typically tens of nanometers) by fitting the spatial distribution of light intensity recorded to the point spread function (PSF) of the microscope [11]. Connecting particle coordinates across frames generates particle trajectories that can be analyzed to deduce properties of the probe target or the membrane. Although SPT has been used extensively for characterizing membrane environments [12], it has seen limited use for the study of protein interactions since the localization precision of the probe is greatly reduced when the particle spacing approaches that found in protein dimers or oligomers [13]. Simultaneous tracking of two spectrally distinct probes can be easily achieved using filter-based separation of the emission. Two-color SPT allows the spatially overlapping particles to be separated by emission spectrum, and in the case of isolated, two-color pairs, the localization precision is restored. This concept has been used for the study of interacting membrane proteins in several systems [14][15][16][17][18]. Despite its advantages, two-color SPT has several limitations: 1) successful tracking requires each color of probe be present at a density of less than about 1 mm 22 ; 2) samecolor dimers are as likely as two-color dimers; and 3) higher order oligomers cannot be easily resolved.
Here we describe a high-speed hyperspectral microscope (HSM) that is capable of resolving and tracking eight or more spectrally distinct fluorescent probes. A hyperspectral imaging system collects a region of the spectrum divided into many contiguous channels for each spatial element of the image. Our microscope design is based around the concept of a spectrally-resolved, linescanning confocal microscope [19], illustrated in Figure 1. A laser focused to a line provides high excitation intensity in a region in and around the focal plane of the objective. The line geometry and the use of a confocal slit on the emission path rejects much of the out-of-focus light that is emitted from either fluorophores or sample auto-fluorescence, thereby providing optical sectioning [20]. To create a two-dimensional (2D) hyperspectral image, the line is scanned across the sample plane in discrete steps (along the x-axis) and the camera captures a frame at each step. Each single camera capture consists of 64 pixels of 1D spatial information and 128 spectral channels. Therefore, a 2D hyperspectral image consists of a 3D (x,y,l) data cube collected in these steps. The rate at which complete hyperspectral images can be captured is limited by the readout rate of the electron multiplying charge coupled device (EMCCD) camera and is 27 frames per second (fps) when scans are configured for 32 steps per image (each image is 326646128). In this configuration the camera is collecting approximately 900 images per second where each camera image corresponds to a single position of the excitation line. In the application of multi-color SPT (mcSPT), each hyperspectral image is analyzed to find the positions of the fluorescent probes and these positions are used to construct particle trajectories.
We describe the design and performance of our instrument and the analysis process required for mcSPT. We then demonstrate the capabilities of the instrument through several proof-of-principle experiments by applying mcSPT to biological systems of interest in our laboratory. Specifically, we show that: 1) The increased labeling density afforded by mcSPT makes detection of rare and transient receptor interactions more efficient, as demonstrated in observations of Epidermal Growth Factor Receptor (EGFR) dimerization [14]; 2) The multiplexing capability of the HSM can be used for determination of antigen-induced FceRI aggregate size based on spectral signature and therefore could be used for cluster size correlation with diffusion coefficient [17,21,22]; 3) mcSPT can be used while simultaneously visualizing other membrane associated structures, which we demonstrate by performing mcSPT while imaging GFP-actin; and 4) High density tracking facilitated by mcSPT can be exploited for fast diffusion mapping of the cell membrane. Because of the advantages in speed, sensitivity and spectral imaging afforded by our HSM, it is likely that the instrument will find use in a wide variety of biological imaging studies and we briefly discuss a few of these other potential applications.

Microscope Design and Performance
The HSM described here can operate at a frame rate of up to 27 fps for an imaging area of 28 mm 2 while each spatial pixel collects 128 spectral channels. Our system is designed to span the spectral range from 500 to 750 nm. The system is built around a standard inverted microscope base (Olympus IX71) in order to facilitate live cell imaging. A single 488 nm wavelength laser (which is used to simultaneously excite all spectral species of QDs) is modified with a laser line generating lens (Edmund Optics) and other optics to provide a uniform (vertical) and diffraction limited (horizontal) line at the microscope focal plane. The laser line is scanned across the sample using a galvanometer controlled scanning mirror (Cambridge Research). Emission light is descanned and passes through a prism spectrometer and is imaged with a high-speed EMCCD camera (Andor iXon 860). The microscope design layout is provided in Figure 2 and a more detailed description of the instrument is given in Text S1. Synchronization of the camera and scanning mirror is achieved by using the camera as the master clock. Specific digital sequences are pre-programmed in a data acquisition card (NI PCIe-6343) and the camera fire signal is used to trigger the next output state. Figure S1 shows the overall instrument control schematic, and Figure S2 provides a few example plots that illustrate the signal timing. When collecting a time series of hyperspectral images (four-dimensional data sets (x,y,l,t)), the camera is set to collect a single kinetic series, and therefore absolute timing is preserved. Data is saved asynchronously as one file for each hyperspectral image.
The localization precision of single fluorophores is proportional to the square root of the number of photons captured [11] and therefore, for throughput efficiency, a prism spectrometer is used as the dispersing element. Prism spectrometers have a higher transmission efficiency over a wide spectral range as compared with diffraction gratings [23], so they are more suitable for this Figure 1. Conceptual diagram of microscope. The excitation beam (optics not shown) is reflected by a dichroic mirror and forms a laser line focused at the sample plane by the objective, concentrating the excitation light to a small volume of diffraction limited width. Here, the white spheres in the sample represent fluorophores that remain mostly in the ground state while the colored spheres denote those which are excited. The emitted light passes through the dichroic mirror and into a spectrometer, which distributes the light onto an EMCCD camera such that each exposure captures information of wavelength and position along the line. The entrance slit on the spectrometer also serves to reject out-of-focus light, providing a semi-confocal ability for imaging at any depth in the sample. A scanning mirror (not shown) advances the line position by one backprojected pixel length on the sample and another exposure is acquired. One hyperspectral ''frame'' is a reconstructed series of these steps (performed in post processing) to form an image containing x, y, and l. A time series of these hyperspectral frames is acquired at 27 fps, providing spatial, spectral, and temporal resolution that enables localized single molecule tracking of multiple emitters within a given diffraction limited volume. For a more detailed description, see Figure 2 and Text S1. doi:10.1371/journal.pone.0064320.g001 application. The spectrometer is an aplanatic design [24], with a dispersion element that consists of a spherical wedge prism and a spherical concave mirror. The design was optimized using optical design software (Optics Software for Layout and Optimization, Lambda Research Corporation) to minimize distortion. Custom optical components were fabricated (Rainbow Optics, Eugene, Oregon USA) to the specifications of the optimized design. The details of the design ( Figure S3 and Text S2) and the spectrometer distortion and dispersion ( Figure S4) of the design are included in the supporting information. Spectral calibration was performed using a spectral calibration lamp (MIDLH Wavelength Calibration Lamp) and the 543 nm and 633 nm lines from HeNe lasers. There is no observable curvature in the spectral lines with respect to the spatial axis ( Figure S5).
The measured 4D (x,y,z,l) microscope point spread function (PSF) shows good agreement with that expected for a diffraction limited system ( Figure S6, S7, and Text S3). Localization of multi-fluorophore beads (TetraSpeck, Invitrogen) broken into spectral bins of 504-534, 562-611, and 641-712 nm show less than 15 nm of chromatic induced shift anywhere in the image ( Figure S8).
Custom graphical user interfaces (GUIs) were developed using MATLAB (Natick, Massachusetts USA) for data acquisition and live viewing of hyperspectral images ( Figure S9). The data acquisition GUI allows for efficient adjustment and optimization of acquisition parameters for experiments. The live viewing GUI provides the experimentalist instantaneous feedback on sample quality and the ability to find a field of view and focus on the sample.

QD Localization and Trajectory Connection
Single particle tracking is carried out by the localization of single molecules and the subsequent building of trajectories from those localizations [25]. The light emitted from a point source (such as a fluorescently tagged protein) is distributed on the detector according to the PSF of the microscope. The 2D (x,y) spatial PSF is well modeled by a 2D Gaussian distribution [26] and the emission spectra profiles of QDs are also well modeled by a Gaussian shape ( Figure 3A). The position of QDs are found by performing a maximum likelihood estimate of spatial position, intensity and spectral peak position using a 3D (x,y,l) Gaussian PSF model and a Poisson noise model ( Figure S10, Text S4, and Table S1). Localizations are performed using an iterative update method [27] that includes the possibility of fitting multiple point emitters within a diffraction limited area [28]. As with previous work [27,28], the analysis is implemented on graphics processing units in a parallel fashion in order to take advantage of the high computational performance of modern graphical processing units (GPUs) ( Figure S11 and Text S5).
To characterize the imaging capability of the HSM, QDs are non-specifically adsorbed onto a glass coverslip and imaged at the fastest scan rate for this geometry. For the field of view collected in these images (64 vertical spatial pixels and 32 line steps; 28 mm 2 ), the pixel dwell time was 1.14 ms and the frame rate was 27 fps. The spectral characteristics of individual QDs for 8 colors (commercially available from Invitrogen) are investigated in Figure S12 and Table S2. The procedure for localizing QDs using hyperspectral data is demonstrated with all 8 colors of QDs in a single acquisition ( Figures 3B-E). See materials and methods for more details. The simultaneous localization of two highly spatially and spectrally overlapping QDs (QD585 and QD605) is highlighted in Figures 3C and 3F. Additionally, a total of four highly spatially overlapping QDs are localized simultaneously. This is highlighted by the colored boxes in Figure 3B and the corresponding fits in Figures 3C, 3D and 3E. The fit precision for each parameter is estimated from the Cramér-Rao Bound (CRB) however, it is important to note that QD blinking during line scanning can contribute to higher than expected error in QD localization in the line-scanning dimension of the data ( Figure  S13). Trajectories are constructed from found QD (x,y,l) positions (Video S1, Text S6, and Figure S14; manuscript in preparation Relich P, Cutler PJ, Huang F, Lidke KA) [29]. The use of spectral information to build trajectories helps to improve the overall accuracy of the trajectory building process ( Figure S14).

mcSPT of Membrane Proteins
Observation of Rare, Transient Events. Protein-protein interactions can be difficult or impossible to observe with single color SPT since the spacing of proteins at distances less than the diffraction limit make it difficult to discern the difference between merely close proximity and real interactions. Two-color tracking restores the ability to calculate separation and deduce dimerization state. Previously, we used two-color SPT to visualize and quantify EGFR homodimerization under various physiological conditions [14]. These studies required low labeling density such that capturing a sufficient number of dimerization events between two labeled receptors required a large amount of data. Multi-color SPT improves upon two-color SPT by increasing the overall labeling density and thereby increases the potential observation rate of rare events while reducing the probability of same color interactions. Using our HSM, we imaged 8 colors of QD-labeled Epidermal Growth Factor (QD-EGF) bound to EGFR expressed on live A431 cells. Video S2 shows the analysis results of a data set collected at 27 fps. Highlighted trajectories in Figure 4A and 4B show several interacting pairs of QD-EGF-EGFR complexes. Consistent with our previous work [14], we observed long-lived ( Figure 4C) and transient ( Figure 4D-F) receptor dimerization events, marked by periods of sustained correlated motion. Note that even in a single time series, multiple interactions are readily identified, demonstrating that the higher labeling density allows for more efficient detection of dimers than with two-color imaging [14].
Oligomer Dynamics. Previous studies of FceRI have shown that receptor mobility is a function of antigen dose and valency, and consequently aggregate size [21]. However, in these experiments, the size of aggregates could not be directly determined since a mixture of dark and QD-labeled IgE (QD-IgE) was used to maintain SPT density. Using the HSM to perform mcSPT of QD-IgE-FceRI on live cells captures not only the dynamics of IgE-FceRI complexes, but also reveals the number of proteins in the aggregate based on spectral signature. Figure 5 and Video S3 show the results of mcSPT of QD-IgE-FceRI approximately 7 minutes after the addition of a crosslinking agent (DNP 3 ). Several long-lived complexes are observed that demonstrate correlated motion. Figure 5A highlights the observation of a trimer and dimer. Spectral separation is further beneficial for accurate quantification of diffusion of QD-IgE-FceRI oligomeric complexes due to their high degree of spatial and spectral overlap ( Figures 5B and 5C). Note that the localization of the QDs in these complexes would not be possible using single color SPT represented by the gray scale projection in Figures 5B and 5C. The red ellipse in Figure 5C (upper left subfigure) is a QD655-IgE-FceRI that comes in close proximity but does not interact with the observed dimer. The spectral features of the QD655-IgE-FceRI are observed in the raw spectrum ( Figure 5C; lower left subfigure; solid black line). In a two color SPT experiment, the proximity of a third QD to an observed dimer would disrupt one of the trajectories in the dimer. The diffusion coefficients of the individual trimer and dimer seen in this example are 0.013 and 0.024 mm 2 /s, respectively. These results demonstrate the ability of mcSPT to investigate the dynamics and composition of protein complexes.
Interactions with Membrane Components. The dynamics of membrane proteins may be influenced by membrane-associated structures such as actin [17], clathrin [30] or other microdomains [31] and interaction with such structures could be correlated with oligomerization state. Our HSM allows the simultaneous observations of several colors of QDs at the same time as other fluorescent probes that can be excited efficiently with 488 nm light, such as GFP. Here we demonstrate mcSPT using 4 colors of QD-IgE-FceRI (605, 625, 655, and 705 nm) while simultaneously imaging GFP-actin. Imaging was performed on rat basophilic leukemia (RBL-2H3) cells transiently transfected with GFP-actin. Figure 6 (Video S4) provides observations that highlight several advantages of this technique. Comparison of the actin structure in the top and bottom images demonstrates how dynamic the actin structure is over the time course of this acquisition. Actin corralling of QD-IgE-FceRI is also observed (red triangle and red circle in the bottom image). These results are consistent with previous reports of actin corralling first observed by tracking QD655-IgE-FceRI with respect to GFP-actin [17]. However, these previous studies were limited to tracking in total internal reflection fluorescence (TIRF) at the adherent surface since wide field imaging of actin structures at the apical surface was insufficient and point scanning confocal was too slow for tracking. Using the HSM, we can perform fast confocal imaging while simultaneously tracking single quantum dots. Moreover, the increased number of spectrally distinct QDs allow for tracking of many receptors simultaneously such that their interactions within the actin domains can be observed. The mobility of the dimer complex (0.004 mm 2 /s; purple blue diamond and red square, top and bottom) is less than that of highlighted monomers (0.044 mm 2 /s). A transition of the mobility of a monomer from relatively immobile (red triangle, top) to relatively mobile (red triangle, bottom) is observed. Crossing trajectories are appropriately connected due to the spectral information (red triangle and yellow star).

Fast Membrane Diffusion Mapping
Labeling with multiple colors enables SPT at densities of several particles per square micron. By binning squared displacement both temporally and spatially, high-density mcSPT can be used to probe local receptor mobility [32]. The uncertainty in the diffusion coefficient scales as approximately D*(A*r*n) 21/2 where D is the diffusion coefficient, A is area, r is particle density and n is the number of time frames (Text S7); therefore, higher density allows better spatial and/or temporal resolution in the diffusion map. In a similar manner, high density mcSPT also allows global changes in protein diffusion to be observed quickly. To demonstrate this, a high density mcSPT experiment was devised in which the relative labeling density for 8 colors of QD-IgE-FceRI was refined to provide similar high density labeling for each color of QD. The labeling density achieved here is ,2.4 particles per mm 2 which to the authors' knowledge is higher density than any previously reported SPT method. Crosslinker (DNP-BSA) was added ,20 seconds into the acquisition and a corresponding change in the estimated instantaneous diffusion coefficient is observed ( Figure 7A). Due to the increased labeling density, this instantaneous change in mobility observed from a single acquisi- tion of a single cell correlates well with that observed in a previous study that required the combined analysis of multiple (,10) SPT experiments of single color of QD-IgE [17,21]. Similarly, squared displacements can be binned both temporally and spatially to build diffusion maps in relation to QD-IgE-FceRI. In Figures 7B  and 7C, squared displacements are binned in 280 frame temporal bins (,10 seconds) and 1 pixel by 1 pixel spatial bins (0.116 mm by 0.116 mm). Changes in diffusion maps subsequent to the addition of crosslinker are apparent upon comparison of Figures 7B and  7C (Video S5).

Other Imaging Modalities
Our fast, line scanning HSM has the potential to be used for many other imaging modalities that are not explicitly demonstrated. Some of these modalities are variants of combinations of the demonstrated approaches. mcSPT could be performed to image several types of proteins labeled with various permutations of QD colors. (e.g. two colors on one membrane receptor and two on another). mcSPT could also be performed simultaneously with more than one additional marker, such as both GFP and yellow fluorescent protein (YFP), which could be separated from each other using multivariate curve resolution (MCR) [22].
Since the instrument is a line scanning system with optical sectioning capability, the microscope can be used to obtain hyperspectral z-slices of a sample. Using the fast scanning ability and a piezoelectric stage, our HSM can make three dimensional observations of samples with hyperspectral information, providing 5D data sets with spatial (x,y,z), spectral (l), and temporal (t) dimensions. A demonstration of live cell, 5D imaging of three cellular markers is shown in the ( Figure S15 and Video S6).
Given the ability of the instrument to image single molecules with high temporal and spectral resolution, the HSM could be used with a variety of probes that report on local environment via spectra [33][34][35][36], for single-molecule, multi-color Foerster reso- nance energy transfer (FRET) studies [37][38][39] or for studying the fluorophores themselves, particularly QDs [40][41][42].
Finally, the instrument could also be used for more typical hyperspectral imaging tasks, such as imaging samples labeled with many spectrally overlapping fluorophores [43], and imaging spectral shifts of native fluorescence in organisms [22].

Discussion
Here we have detailed the development of a novel hyperspectral microscope that provides an unprecedented combination of speed, sensitivity and a spectral detection. Our HSM design provides acquisition rates of 27 fps over a 28 mm 2 field of view with 128 spectral channels. These features allow the determination of stoichiometry and dynamics of small oligomers, which cannot be measured using any other currently available techniques. We have demonstrated these capabilities of our HSM by multi-color single QD tracking and diffusion mapping of membrane protein dynamics. While EM or super-resolution techniques do provide a nanometer-scale image of protein distributions and clustering, they cannot provide information on protein diffusion or clustering dynamics. One variant of super-resolution that captures dynamics uses SPT of sequentially photo-activated fluorescent proteins [32], but it still requires long acquisition times to build up a diffusion map and cannot track a single protein or aggregate for more than a few seconds and cannot be used to identify interactions.
The maximum speed is limited by the EMCCD readout rate and is achieved by imaging only 64 of 128 pixels in the vertical shift direction. In this configuration, the maximum frame rate scales directly with the image area. A higher frame rate or larger field of view could be achieved by rotating the camera such that spectral binning in the vertical shift direction would increase frame rate. This would most benefit fast, 3D imaging of conventional fluorophores. We estimate that the imaging area could be increased by a factor of two while maintaining similar signal to noise levels with a state-of-the-art scientific complementary metal oxide (sCMOS) camera [44,45]. However, pixel-to-pixel noise and gain variation of sCMOS cameras may complicate particle localization and MCR analysis [46]. Several other spectral imaging methods could be envisioned for performing mcSPT. In principle, a widefield microscope could be used with a series of dichroic and emission filters to setup an arbitrarily large number of spectral channels. This approach has been used for two color [14] and four color imaging [47]. However, for image registration better than 20 nm, which is required for studying protein-protein interactions, careful mapping between all channels must be periodically performed using multispectral reference beads. In contrast, our spectrometer system provides less than 15 nm shift in spatial localization due to chromatic aberration from 500 nm to 750 nm ( Figure S8). Another recently described approach for wide field, fluorescence spectral imaging uses an image mapping spectrometer [48]. This apparatus would seem not to suffer from channel registration problems and although has only been demonstrated at ,7 fps, would in principle be capable of video rate imaging. Of course, these wide field approaches do not provide a mechanism for rejecting out of focus light and could therefore suffer as compared to a line-scanning system when imaging in an environment with high cellular auto-fluorescence or other background.
Spectral imaging capabilities can be found on many commercial point scanning microscopes as well as custom built point-scanning hyperspectral systems [23,43,49]. However the finite lifetime of fluorophores limits photon yield even at arbitrarily high laser power and leads to a practical limit for SPT in point scanning systems. For a region of interest that is 1286128 spatial pixels, with a desired localization precision of ,10 nm, the maximum point scanning frame rate is ,2 fps and the maximum line scanning frame rate is ,250 fps (Text S8). In terms of achieving a maximum area scanned for a given specific frame rate of 30 fps (33 ms per frame) and with the same 10 nm localization precision, these limits are ,512 um 2 /33 ms for organic fluorophores (t = 2 ns) and as low as 51 um 2 /33 ms for QDs (t = 20 ns) for a point scanning system. The latter value is fairly close to the area covered by this instrument when acquiring at 30 fps, and does not include consideration of spectral data acquisition. In addition, any non-linear photobleaching, photo-destruction or long-lived dark states would further reduce these values in practice. In comparison, a line-scanning system has the advantage of parallel observations along the spatial dimension and therefore gives an additional factor equal to the number of pixels that are simultaneously observed.
QDs have a broad excitation spectrum, narrow emission spectrum and high photo-stability. All of these properties are taken advantage of for mcSPT using our HSM. However, QDs do have practical limiting properties [40] such as fluorescence intermittency, the potential for non-fluorescent QDs [50] and a difficultly in producing monovalent probes [51]. For example, the presence of non-fluorescent QDs leads to the possibility of underestimating the oligomer size of complexes in mcSPT experiments. Consideration must also be taken when using biotin-protein:streptavidin-QD conjugation schemes since unlabeled protein could also lead to underestimation of oligomer size. Consequently, the oligomers identified in Figures 5 and 6 should be considered as minimal dimers and trimers. Contribution from non-biotynilated protein can be minimized by purification on streptavidin resin prior to conjugation with the QDs. The dark fraction of QDs for each spectral species can be measured and used in a statistical analysis of the properties under study as a function of oligomer size. mcSPT could also be performed with other probes that may have overlapping spectra, such as nonblinking QDs [52,53], 'Cornell dots' [54] or other systems that use combinations of organic fluorophores and could have the benefit of reduced or non-existent fluorescence intermittency.
In summary, we have developed a high-speed hyperspectral microscope and the required analysis software to perform simultaneous SPT of up to 8 spectrally distinct QDs at video rate. We have provided a description of the instrument, software, and several proof-of-principle applications of the HSM. The applications presented in this manuscript (mcSPT, mcSPT with GFP background, high density mcSPT, and 3D scanning) highlight how the HSM can provide unique insight into dynamic cellular events.

Imaging of QDs Non-specifically Adhered to Glass
The quantum dots (Invitrogen Corporation QdotH 525, 565,  585, 605, 625, 655, 705, and 800) were each diluted to 50 pM in 16 phosphate buffered saline (PBS) and then mixed together in equal volumes. 300 ul of this mixture was then added to each well in an 8-well chamber. 10 ul of 2 M NaCl (aqueous) was then added to each well and then incubated at 4uC for up to 8 hours to allow the quantum dots to attach to the glass. The solution was then replaced by 16PBS. 250 frames of data (providing x, y, and l information) were acquired at 27 frames/sec over an area of 28 mm 2 (the intended scan rate for our typical live cell imaging) with an excitation laser line intensity of ,1000 W/cm 2 .

Image Acquisition for SPT Experiments on Live Cells
Samples were maintained at 34-36uC using an objective heater (Bioptechs, Butler, PA, USA) for imaging. For an acquisition rate of 27 fps, a region of interest (ROI) of 3.7 mm in x and 7.4 mm in y is acquired. This ROI consists of 32 steps (x dimension; ,0.116 mm steps) of the line across the sample by the scanning mirror and 64 pixels (y dimension; ,0.116 mm pixel size) along the line.

Single Particle Tracking (Preprocessing, Localizations, and Trajectory Building)
Immediately prior to the acquisition of each HSM image, 1000 frames of dark images were acquired for background correction. The mean dark image was subtracted from each frame of the HSM image. The spectrometer was aligned so that several columns of spectral pixels are dark due to a long pass filter.. The mean value of a dark column of pixels was subtracted from the SPT image to correct a baseline offset. The HSM image was then multiplied by a calibrated gain factor [56] to convert arbitrary pixel counts to photons.
The steps in SPT in an HSM image are similar to steps used in traditional SPT algorithms: image segmentation, single particle localization, and trajectory building. Uniform and maximum filtering procedures outlined in previous work by Huang et. al. [28] for two dimensions are expanded to three dimensions (x,y,l) and used to identify centers for sub-volumes in which to localize single molecules. The empirically determined sub-volume size used for single molecule localization in this work was 25 spectral pixels by 7 spatial pixels by 7 spatial pixels. A 3D Gaussian PSF/ spectral model and Poisson noise model ( Figure S10 and Text S4) were used to model the spatial and spectral features of individual QDs. The variance in the parameters estimated by the localization algorithm converges to the CRB ( Figure S10B). Multiple 3D Gaussians were used to fit multiple spatially and spectrally overlapping QDs ( Figure S11 and Text S5). An artifact of using a line-scanning instrument is that blinking during the scanning of a single QD causes errors in QD localization in the scan direction ( Figure S13). Typically, each sub-volume was fit using 1-4 emitter models. Accurate fits were then selected using an empirically derived filtering scheme (Text S4). Trajectories are built from filtered localizations using a cost matrix approach to form a linear assignment problem [29] (Text S6).

Quantification of Trajectory Mobility
Diffusion coefficients (D) for groups of trajectories ( Figures 5  and 6) were determined by fitting a line to the first 3 points of the ensemble mean squared displacement. A maximum likelihood estimator based on independent jump analysis (manuscript in preparation Cutler PJ, Relich P, Lidke KA) is used to estimate the instantaneous D and diffusion mapping (Figure 7). For the instantaneous D, individual jumps were grouped temporally using a sliding window of 14 frames (,0.5 s) according the temporal midpoint of the observations. For diffusion mapping, individual jumps were binned spatially according to camera pixel sampling (,116 nm in each dimension) and temporally according to the midpoint of the temporal jump using a sliding window of 280 frames (,10 s).

Cell Treatment for SPT of QD-EGF
Procedures for labeling A431 cells with QD-EGF were adapted from previous work [14]. Cells were serum starved in the presence of 1 mM PD153035 for several hours prior to imaging. PD153035 (1 mM) was included in imaging buffer throughout the remainder of the experiment to prevent receptor endocytosis. Tyrodes solution with 0.1% BSA and 20 mM glucose was used as an imaging buffer. Media was exchanged with 200 ml of imaging buffer prior to imaging. Immediately prior to data acquisition, 100 ml of 12 pM (concentration for each individual color) of QD-EGF (final concentration of 4 pM) equilibrated at 37uC was added to the well to be imaged and data was acquired within 10 minutes.

Cell Treatment for SPT of QD-IgE
Procedures for labeling RBL-2H3 cells with QD-IgE for SPT were adapted from previous work [17].

3D-scanning of Live Cells
RBL-2H3 cells were incubated in 200 ml of 5 nM QD-IgE (655) in HBSS for 30 minutes at 37uC followed by 2 washes with HBSS. The cells were incubated with 200 ml of 5 mg/ml CellMask orange for 3 min at 37uC followed by 3 washes with HBSS. Imaging was performed immediately after labeling. Crosslinker (DNP-BSA; final concentration 10 mg/ml) was added ,60 seconds into the image acquisition. In order to assure rapid mixing of crosslinker and to prevent focal drift, 100 ml of crosslinker stock at 36 final concentration in HBSS preheated to 37uC was added (contains 200 ml HBSS).
Z-stacks covering an ROI of 14.8 mm in x, 14.8 mm in y, and 7.0 mm in z were acquired at a rate of 7.72 seconds per z-stack. Each z-slice for this ROI consists of 128 steps (x dimension; ,0.116 mm steps) of the line across the sample by the scanning mirror and 128 pixels (y dimension; ,0.116 mm pixel size) along the line to make each z-slice, and each z-stack for this ROI contains 29 z-slices with a step size of 0.25 mm. In order to conserve CPU memory for storage and post-processing, the data was binned into 32 spectral bins instead of 128. It is important to note that altering the number of spectral bins does not significantly alter the acquisition time with the current camera configuration. A total of 50 z-stacks were acquired giving a total acquisition time of ,6.5 min. Z-stacks were processed using 3D deconvolution. The experimentally measured PSF was used with the built-in Matlab deconvolution algorithm deconvlucy. An isosurface rendering was then used to represent the data.

Supporting Information
Video S1 SPT on glass. SPT of 8-colors of QDs nonspecifically adsorbed on glass. (top) RGB representation of the raw data with overlaid trajectories (coloration of trajectories are according to color bar in Figure 3). The white box is approximately representative of the spatial region for the highlighted sub-volumes in Figure  In the HSM setup, the maximum distance from the axis is 1.5 mm, so the designed maximum distortion is 0.4 mm, which is less than 2% of CCD pixel size (24 mm). (B) Lateral shift shows the dispersion, designed to span a total of 3 mm from 500 nm to 800 nm. It shows dispersions in four relative heights (distance from the optical axis), the four curves are right on top of each other, so the spectral lines are straight, which also indicates a minimal distortion. (PNG) Figure S5 Spectral calibration. (A) Image of the calibration lamp on the camera (the vertical axis is the spectral dimension). The calibration lamp is a LightForm multi-ion discharge lamp (MIDL), the slit width is 20 mm, and the camera is an Andor iXon 860 EMCCD with an array size of 1286128, pixel size 24 mm. (B) Spectral curve, generated by the sum projection of (A) along the x dimension; the peak wavelengths are identified from the specification sheet of the calibration lamp. The blue, green, and red lines correspond to the regions in (A) with the same color. All three colored lines almost overlap with each other, which demonstrates a very small distortion in the spectral lines. (C) Dispersion curve, generated from the spectral curve (C), by polynomial fit of specific wavelengths. (PNG) Figure S6 4D PSF theoretical. The 4D PSF is calculated from the theory as described in Text S3. The spectra and wide field point spread function is based on 100 nm yellow bead. The simulation includes aberrations generated from phase retrieving process. A linear stretch (A) and log stretch (B) of intensity for projections on the x-y, x-z, and y-z planes are shown along with the x, y, and z vs. wavelength projections (C, D, and E). (PNG) Figure S7 4D PSF experimental. The sample is a 100 nm TetraSpeck bead imaged at 100 nm steps in z. A linear stretch (A) and log stretch (B) of intensity for projections on the x-y, x-z, and y-z planes are shown along with the x, y, and z vs. wavelength projections (C, D, and E). The figures show three emission peaks of TetraSpeck bead, 505/515 nm (green), 560/580 nm (orange) and 660/680 nm (dark red). Data were acquired at room temperature, using a 606 water objective, NA = 1.2. (PNG) Figure S8 Color dependence of localizations. Scatter plot of 300 localizations of a single 100 nm TetraSpeck bead on glass using three different wavelength ranges. The purpose here is to quantify any chromatic dependence on particle localization. Emitters were 100 nm Tetraspeck beads diluted 2000:1 in 16 PBS and allowed to stick to a glass cover slip. The sample was then scanned 300 times and localizations were obtained for each scan using emitted light from each of three different wavelength ranges.  Figure S9 Hyperspectral viewing GUI. Hyperspectral viewing GUI provides real-time feedback for instrument control (focusing, ROI determination, etc.) and experimental observations (sample quality, QD labeling density, etc.). The GUI can also be used for preliminary investigations of data in post processing. For example, the blue, green, and red boxes in the RGB image (top left of GUI) highlight single 565, 625, and 800 nm QDs respectively. Spectra for pixels within those boxes are shown in the axes below the RGB image. The GUI provides flexibility in the coloration scheme for the RGB image. For the default RGB representation, each spectral channel is assigned an RGB color progressing from blue to green to red. RGB values for each spatial pixel are determined by relative contributions from of RGB values from each spectral channel. (PNG)  Table S2). Column 2 contains histograms of fit spectral emission peaks (h l ) identified by 3D Gaussian fitting (x,y,l) for individual QDs. Red ticks indicate the expected emission peak for each QD specie. Note that the expected emission peak for QD800 is significantly different from the histogram. This is attributed to the spectral range of the camera (i.e. only the tail of the spectrum is observed on the camera). Column 3 contains histograms of the fit standard deviation in the spectral dimension (h sl ) identified by 3D Gaussian fitting for individual QDs. Column 4 contains histograms of fit photons per frame (h sl ) for individual QDs. Table S1 includes further description of the model parameters. Statistics are included in Table S2. (PNG) Figure S13 QD localization and blinking. (A) Localizations for a single QD605 overlaid on the sum projection of raw data. Coloration of localizations corresponds to the localized emission peak (color bar). The estimated intensity (left axis; blue) and spectral emission peak (right axis; green) are shown over time (,3.8 fps) in (B). Note the blue shifting of the QD is observed in (A) and (B). Localizations were binned by intensity in (C) and the standard deviation of the localization parameters were compared to the square root of the CRB in (D), (E), and (F). The standard deviation in the line dimension (E) corresponds well to the theoretical lower bound. The standard deviation in the scanning dimension (D) and the spectral dimension (F) do not correspond well with theory. This is attributed to QD blinking during the line scan and QD blue shifting respectively. (PNG) Figure S14 Spectral information improves trajectory connections. Hyperspectral image data (100 frames) with diffusing QDs were simulated (QD simulation parameters attained from the characterization of QDs in Figure S12 and Table S2) and localized (Text S4, Figure S10, Text S5, and Figure S11). Trajectories were subsequently built either utilizing (solid colored lines) or ignoring spectral information (solid black lines). Localizations are shown in ellipsoids (size of the ellipsoid corresponds to 36 the estimated localization accuracy), and solid colored lines correspond to the connected trajectories. Text S1 Details about general instrument design.