Dynamic partitioning of a glycosyl-phosphatidylinositol-anchored protein in glycosphingolipid-rich microdomains imaged by single-quantum dot tracking.

Single-molecule imaging by objective-type TIRF Custom-built dual-color

TIRF microscopes were used to perform two-color imaging of qdot-labeled Av-GPI a n d C T × B A l e × a 4 8 8 l a b e l e d-GM1 or Cav1-EGFP (schematics can be viewed at http://fpinaud.bol.ucla.edu/index_files/Traffic.htm ). In both cases, the ventral plasma membrane of HeLa cells was imaged. The diffusive behavior of Av-GPI was similar in the ventral or dorsal membrane (data not shown). However, we preferred ventral imaging by TIRF to limit the probability of imaging membrane regions containing microvilli or other membrane corrugations thus guaranteeing that the diffusion of the Av-GPI was essentially confined in two dimensions ( 110 ). A detailed description of the setups is available in Supporting Information . Unless stated, imaging was done at room temperature (~27 °C). Typically, 2 to 3 cells per field-of-view were imaged continuously for 90-120 s, using an integration time of 100 ms per frame. Faster acquisitions were also performed (7 ms/frame using only a subset of the EMCCD camera, data not shown). However, because of software limitations that prevented from storing movies with large numbers of frames, such faster frame rate would have resulted in trajectories with shorter duration (and limited field of views), which would have undermined our ability to detect the existence of long-term changes in diffusion regime in the cell membrane. In addition, because CT×B-Alexa 488 and qdots have different brightness, acquisition times had to be optimized so that both signals could be simultaneously detected on each side of the same EMCCD camera. At 100 ms/frame, there was a good balance between signal intensity and acquisition speed. For single-dye tracking (SDT) of Av-GPI labeled with Alexa-488 biocytin (Invitrogen), imaging was performed continuously with an integration time of 60 ms and at a final dye concentration of 75 pM. For all conditions, cells grown at ~70 % confluency on fibronectin coated coverslips were first starved for 3-4 hr in serum-free DMEM at 37 °C to free Av-GPI from biotin present in the serum supplement. Cells were then imaged in HBSS + 1 % BSA for a maximum of 30-45 min after which they were replaced. Biotinylated CdSe/CdS/ZnS qdots emitting at 620 nm and coated with 50 % biotinylated peptides and 50 % of 600 Da polyethylene glycol (PEG) peptides were synthesized and prepared as previously described ( 43 , 111 ). Alternatively, 620 nm-emitting CdSe/ZnS in toluene (Evident Technologies, Troy, NY) were used and coated with peptides as above. The final diameter of the biotinylated qdots was 13.0 ± 1.1 nm ( 112 ). Quasi-monovalent biotinylated qdots were obtained by reducing the amount of biotinylated peptides to 1-2 % during the coating procedure and using 49-48 % of a third, lysine terminated peptide (peptide # 7 in Pinaud et al. ( 43 ), Fig. S3 ). Qdots were added to the cells at a final concentration of 2-3 pM directly in the imaging media to label only few Av-GPI (~ 10 per cell). A low-level labeling facilitated tracking for long duration, by limiting the probability that two or more qdot-labeled Av-GPI crossed paths or blinked simultaneously within the same area. In addition, low concentrations of qdots limit potential toxic effects ( 113 ) and interference with normal cellular metabolism. For dual-color imaging experiments, a final concentration of 0.05-0.1 μg/ml of Alexa 488 cholera toxin B (CT×B) was added simultaneously with qdots.

Data Analysis

Post-acquisition, each frame of a recorded TIRF movie was split into two images (green and red channels) using Metamorph software (Molecular Device, Sunnyvale, CA). Images of both channels were overlaid after correcting for offsets and spherical and chromatic aberrations as determined from the emission of 40 nm TransfluoSpheres. A software was developed in Labview (National Instruments, Austin, TX) to track and analyze trajectories of qdot-labeled Av-GPI (See Supporting information for details). Briefly, regions of interest in images were selected and a semi-automatic algorithm was used to fit the point-spread-function (PSF) of the tracked qdots with a 2D Gaussian in each frame ( 52 , 53 ) ( Fig. 2 ). The quality of the fit was verified visually for each frame. When the signal in a frame was lost because of blinking, no fitting was performed until re-appearance of the PSF. When a PSF did not re-appear within 10 s, tracking was aborted. Tracking was also aborted when two qdot-labeled Av-GPI crossed paths. Because of blinking, binding to cell membrane during imaging, or internalization of qdot-labeled Av-GPI, the mean duration of trajectories varied between 50-75 s. After tracking, a trajectory file containing all information was saved for each Av-GPI molecule. Subsequent analysis (described in detail in Supporting Information ) included (i) computing the mean-square displacement (MSD) curve for a trajectory, (ii) representing the fluorescence intensity along the diffusion path, (iii) representing the instantaneous diffusion coefficient ( 38 ), (iv) calculating the probability distribution of square displacement P ( r → 2 , τ ) (PDSD) for each molecule or for all molecules by global analysis ( 55 ) and (v) analyzing PDSD for different time lags in terms of diffusion models ( Table S2 ). The presence of single or multiple diffusive regimes within each trajectory was deduced from PDSD analysis on the first 10 % of time lags t, and by fitting the PDSD with one, two or three exponentials. The quality of each fit was evaluated using the normalized residual curve. Normalized residuals deviating systematically by more than 10 % (from zero) resulted in the rejection of the fitted curve. The fitted exponents obtained were then used to plot r i 2 ( t ) curves for different time lags t (see Supporting Information ). If the PDSD curves were well fitted with a single exponential, the corresponding single r 2 (t) curve was plotted ( Fig. 2D , Av-GPI with a single diffusion mode). If the PDSD was best fitted with more than one exponential, then as many r 2 (t) curves were plotted ( Fig. 2D , Av-GPI with multimodal diffusion). For each r 2 (t) curve, the diffusion mode and corresponding diffusion coefficient were determined as follows. The diffusion model that best described a r 2 (t) curve was selected by fitting the r 2 (t) curve with the most likely among diffusion models picked from Table S2 . For instance, for a concave r 2 (t) curve, we skipped the directed motion model. The respective quality of each fit was assessed visually although a quantitative method such as normalized residuals analysis could be implemented if needed. Additionally, we used Occam's razor principle, keeping the model with the least parameters if two or more models accounted well for the r 2 (t) curve. The fitting parameters for the best fit diffusion model were used as is (see Supporting Information ). This approach does not only permit a classification of each mode of diffusion into pure Brownian, restricted or directed motion as does the approach of Wilson et al. ( 114 ) but also allows the computation of a diffusion coefficient for each mode. Since the PDSD curves are calculated for the first 10 % of time lags, there is ample statistics to obtain reliable r i 2 ( t ) curves. Monte-Carlo simulations of trajectories of freely diffusing particles, confined or switching between free and confined diffusion, confirmed that PDSD analysis could detect multimodal diffusion and correctly recover the diffusion coefficient for each mode of diffusion ( Fig. S5 ). In few cases, no convincing PDSD fits or no convincing r 2 (t) fits could be obtained. In these cases, the trajectories were rejected. In addition to the classification into different mode of diffusion, we identified a category of immobile Av-GPI (if D < D min ) using a diffusion coefficient threshold D min = 4.8 10 -5 μm 2 /s. This cut off value D min corresponds to the 95 th percentile of the distribution of diffusion coefficients for qdots imaged on fibronectin-coated coverslips, and represents the level of drift in the setups ( Fig 3 , grey histogram). The apparent diffusion coefficients measured from fitting each r 2 (t) curves were reported in histograms. Because the diffusion coefficients are distributed log-normally over several orders of magnitudes, histograms of the decimal logarithms of diffusion coefficients are reported (D-histograms in the following). The bin size of the D-histograms was determined statistically using an approach introduced by Knuth ( 115 ). This method adapts the bin size to the underlying distribution without a priori knowledge of its nature. After evaluating the optimal bin size for each experimental condition, we chose twice the largest bin size and applied this modified optimal bin to all data. With this common bin size, a direct comparison between all distributions is possible and sub-populations in histograms are sufficiently separated. In the text, the diffusion coefficients reported correspond to the peak position(s) of single (or multiple) Gaussian fit(s) of the D-histograms. Standard errors values (SE) were determined using 1,000 bootstrap replica of the distributions ( 116 ), and are given in the form of an upper and lower diffusion coefficient. Because the distributions of the diffusion coefficients are relatively large, we only considered differences in diffusion coefficients larger than 3 times the SE to be significant. To evaluate the number of Av-GPI switching between the fast and slow regime (or vice versa), a cut-off value corresponding to the 95 th percentile of the Gaussian fitted on the slow distribution was used (e.g. 1.3 10 -2 μm 2 /s for Fig. 3A ). Av-GPI trajectories having two very different diffusions regimes distributed above or below this cut-off value were classified as switching trajectories. The sizes of confinement domains are extracted for Av-GPI traces fitted with a restricted diffusion model ( Table S2 ) and are separately determined for fast and slow subpopulations. The confinement/zone geometry is assumed to be circular (corral). As for diffusion coefficient values, corral sizes appeared to be distributed log-normally. We therefore computed histograms of decimal logarithm of sizes. Confinement sizes were determined by Gaussian fitting of the distribution histograms. Standard error of the mean (SE) was determined as before. Differences in confinement size larger than 3 times the SE were considered to be significant. To study the colocalization of Av-GPI with immobile/slow diffusing CT×B-labeled GM1-rich domains or Cav1-EGFP-labeled caveolae, Av-GPI trajectories were overlaid with the mean intensity projection image of the movie (green channel, ΣI mean ). ΣI mean is defined as the image whose pixel value p(k, l) at coordinates ( k, l) is the mean value of all pixels p i (k, l) for N frames of the movie at this location: p ( k , l ) = 1 N ∑ i = 1 N p i ( k , l ) . With this approach, diffusing GM1 or Cav1-EGFP are efficiently filtered out from the final image and GM1-rich microdomains and caveolae are easily identified (data can be viewed at http://fpinaud.bol.ucla.edu/index_files/Traffic.htm ). Further confirmation of colocalization was done by overlaying the fluorescence intensity time trace of a qdot-labeled Av-GPI along its trajectory (in a 3×3 pixel region centered on the qdot location) with the corresponding fluorescence intensity time trace in the green channel. Periods showing overlap of red and green fluorescence after background subtraction were interpreted as interaction of Av-GPI with GM1 domains or caveolae. However, this approach is limited by the bleaching of organic dyes and EGFP, which make the distinction between labeled and non-labeled regions more difficult at longer times. In few cases, we could correlate changes in diffusion (D in and D out ) with the entry or exit from GM1-rich domains or caveolae by performing MSD analysis on sub-trajectories and PDSD analysis ( Fig. S5 ). Parts of the trajectory in which Av-GPI is colocalized with a GM1-rich domain or a caveola (identified by fluorescent signal above background in the CT×B or Cav1-EGFP channel) were selected and separated from the non-colocalized parts. The MSD of each sub-trajectory was separately computed. MSD fitting with a simple Brownian motion model results in a D value for each types of region, D in and D out . However, this analysis has several drawbacks: (i) it assumes that the boundaries of these regions (along the trajectory) can easily be identified and (ii) the MSD analysis does not permit an accurate determination of the type of diffusion, and hence the corresponding diffusion constant. At least qualitatively, MSD analysis reveals that diffusion in GM1-rich domains is much slower than outside them (D in