A quantitative approach for analyzing the spatiotemporal distribution of 3D intracellular events in fluorescence microscopy
 Cited 0
 Views 883
 Annotations
Abstract
Analysis of the spatial distribution of endomembrane trafficking is fundamental to understand the mechanisms controlling cellular dynamics, cell homeostasy, and cell interaction with its external environment in normal and pathological situations. We present a semiparametric framework to quantitatively analyze and visualize the spatiotemporal distribution of intracellular events from different conditions. From the spatial coordinates of intracellular features such as segmented subcellular structures or vesicle trajectories, QuantEv automatically estimates weighted densities that are easy to interpret and performs a comprehensive statistical analysis from distribution distances. We apply this approach to study the spatiotemporal distribution of moving Rab6 fluorescently labeled membranes with respect to their direction of movement in crossbow and diskshaped cells. We also investigate the position of the generating hub of Rab11positive membranes and the effect of actin disruption on Rab11 trafficking in coordination with cell shape.
https://doi.org/10.7554/eLife.32311.001eLife digest
Proteins are the workhorses of the body, performing a range of roles that are essential for life. Often, this requires these molecules to move from one location to another inside a cell. Scientists are interested in following an individual protein in a living cell ‘in real time’, as this helps understand what this protein does.
Scientists can track the whereabouts of a protein by ‘tagging’ it with a fluorescent molecule that emits light which can be picked up by a powerful microscope. This process is repeated many times on different samples. Finally, researchers have to analyze all the resulting images, and conduct statistical analysis to draw robust conclusions about the overall trajectories of the proteins. This process often relies on experts assessing the images, and it is therefore timeconsuming and not easily scalable or applied to other experiments.
To help with this, Pécot et al. have developed QuantEV, a free algorithm that can analyze proteins’ paths within a cell, and then return statistical graphs and 3D visualizations. The program also gives access to the statistical procedure that was used, which means that different experiments can be compared.
Pécot et al. used the method to follow the Rab6 protein in cells of different shapes, and found that the conformation of the cell influences where Rab6 is located. For example, in crossbowshaped cells, Rab6 is found more often toward the three tips of the crossbow, while its distribution is uniform in cells that look like disks. Another experiment examined where the protein Rab11 is normally placed, and how this changes when the cell’s skeleton is artificially disrupted. Both studies help to gain an insight into the behavior of the cellular structures in which Rab6 and Rab11 are embedded.
Following proteins in the cell is an increasingly popular method, and there is therefore a growing amount of data to process. QuantEV should make it easier for biologists to analyze their results, which could help them to have a better grasp on how cells work in various circumstances.
https://doi.org/10.7554/eLife.32311.002Introduction
Modern light microscopy associated with fluorescence molecule tagging allows studying the spatial distribution of intracellular events. Unfortunately, fluorescent images are complex to analyze and additional software is needed to evaluate statistical differences between different conditions (Meijering et al., 2016; Tinevez et al., 2017). Automatic methods have the obvious advantage of being quicker and reproducible. However, most computational methods are based on the complex combination of heterogeneous features such as statistical, geometrical, morphological and frequency properties (Peng, 2008), which makes it difficult to draw definitive biological conclusions. Additionally, most experimental designs, especially at singlecell level, pool together data coming from replicated experiments of a given condition (Schauer et al., 2010; Merouane et al., 2015; Biot et al., 2016), neglecting the biological variability between individual cells.
Micropatterning is now a wellestablished strategy to reduce morphological variability by imposing constraints on adhesion sites, which has been shown to influence the cytoskeleton geometry and transport carrier localization (Théry et al., 2005; Schauer et al., 2010). This technique opened the way to pairwise comparisons of conditions with a twosample kernel densitybased test by pooling together all data from each condition (Duong et al., 2012). Unfortunately, it does not consider the sampletosample variability because all replicated experiments from a given condition are simply merged together. Additionally, the visualization of the kernel density maps enables to average several experiments but fails to identify specific locations of interest in the cell (e.g. docking areas). Finally, assessing the dynamical behavior of labeled membrane structures, a fundamental task for trafficking analysis, remains out of scope in this framework.
In this paper, we describe a method that we call QuantEv dedicated to the analysis of the spatial distribution of intracellular events represented by any static or dynamical descriptor (e.g. detected points, segmented regions, trajectories...) provided that the descriptors are associated with spatial coordinates. QuantEv offers a unifying frame to decipher complex trafficking experiments at the scale of the whole cell. It is typically able to detect subtle global molecular mechanisms when trajectory clustering fails. An overview of the approach is presented in Figure 1. Our approach first computes 3D histograms of descriptors in a cylindrical coordinate system (parameterized by radius $r$, angle $\theta $ and depth $z$) with computational cell shape normalization, enabling comparisons between cells of different shape. Densities are obtained via adaptive kernel density estimation (Silverman, 1986; Taylor, 2008). Visualization through histograms and densities allows giving a clear biological interpretation of the experiments. We use the Earth Mover’s Distance (Rubner et al., 2000) and the Circular Earth Mover’s Distance (Rabin et al., 2011) to measure the dissimilarity between densities associated with different experimental conditions. A statistical analysis of these distances reliably takes into account the biological variability over replicated experiments. By computing weighted densities for each point in the cell as the reference center, QuantEv identifies the point that gives the most uniform angular distribution. This point may coincide with a biological structure of interest that would act as the events emitter or attractor.
In the section Results, we describe the application of QuantEv to detect significant differences between molecular trafficking and phenotypes observed in cells with various shapes. The first application is concerned with the distribution of membranes labeled by GFPRab6 as a hallmark of vesicular carriers in unconstrained, crossbow and diskshaped cells. Rab6 proteins are transiently anchored to moving transport carriers from the Golgi apparatus located at the cell center to Endoplasmic Reticulum entry sites or to plasma membrane (White et al., 1999; Chavrier and Goud, 1999; Echard et al., 2000; Opdam et al., 2000; Grigoriev et al., 2007; Bardin et al., 2015), both assumed to be located at the cell periphery. Cell shape imposes constraints on the cytoskeleton and consequently influences the spatial distribution of Rab6 transport carriers, as confirmed with kernel density maps (Schauer et al., 2010). We apply QuantEv to visualize and quantify this influence and to localize regions in the cell associated with Rab6 trafficking stages. In addition, Rab6 positive membranes were reported to move from and toward the Golgi in apparent close proportions (Grigoriev et al., 2007, Grigoriev et al., 2011), and yet these membrane associated proteins are believed to traffic in majority from the Golgi located at the cell center to the cell periphery (White et al., 1999; Chavrier and Goud, 1999; Echard et al., 2000; Opdam et al., 2000; Grigoriev et al., 2007) where they should dissociate from membranes and recycle back to the cytosol. To investigate these apparently antagonist statements, we apply QuantEv on Rab6 trajectories to characterize the dynamical behaviors of these transport carriers.
The second application focuses on the dynamics of mCherryRab11positive membranes. Rab11 is known to be primarily localized to the Endosomal Recycling Compartment (ERC), and it organizes spatially and temporally recycling from this compartment (Ullrich et al., 1996; Gidon et al., 2012; Baetz and Goldenring, 2013; Boulanger et al., 2014). Here, we confirm by using QuantEv the hypothesis that the labeled transport intermediates are uniformly distributed around the ERC at the plasma membrane plane. Furthermore, we also investigate the progressive effect of actin disruption induced by Latrunculin A injection on the ERC localization with respect to time. We finally apply QuantEv to analyze the joined influence of actin disruption and cell shape on the radial distribution of Rab11 vesicles trafficking.
Results
Visualizing and quantifying the influence of cell shape on the spatial distribution of Rab6 positive membranes
We applied the QuantEv approach to visualize the spatial distribution of Rab6positive membranes in unconstrained, crossbow and diskshaped cells (see Appendix 1—figure 1) and quantify their differences. To test the generic performance of QuantEv, these image sequences were acquired with two different 3D imaging modalities, a multipoint confocal microscopy and a wide field video microscopy. We compared the results obtained with QuantEv to those obtained with the more conventional kernel density (KD) maps (Schauer et al., 2010; Merouane et al., 2015). The KD approach concludes that the distribution of Rab6positive membranes are clearly different between cells of different shapes (see Figure 2a–c, p value = 0 when considering unconstrained cells versus crossbowshaped cells, unconstrained cells versus diskshaped cells, and crossbowshaped cells versus diskshaped cells). Unfortunately, it also leads to a significant difference when image sequences with the same cell shape are compared (see Figure 2d). This demonstrates that the KD approach is too sensitive. Instead, QuantEv shows a uniform range of p values when cells with same shape are compared (see Figure 2d) while it leads to significant differences between radial, angular and indepth distributions of Rab6 proteins from cells with different cell shapes (see Figure 2e–f). The angular distribution of Rab6 proteins is different for the three cell shapes. It ranges from a completely uniform distribution for diskshaped cells, to a less regular distribution for unconstrained cells and to a distribution oriented toward the three tips of the crossbow for crossbowshaped cells. Indepth and radial distributions are similar for crossbow and diskshaped cells. In contrast, they are different from unconstrained cells. Unconstrained cells show diverse sizes with a strong tendency to spread. This explains why the indepth distribution is flatter for the unconstrained cells than for the constrained cells. Interestingly, QuantEv is able to reflect these differences. QuantEv also highlights a distribution maximum for a radius at the twothirds (resp. fivesixth) the distance between the Golgi region border and the cell periphery for both micropatterns (resp. unconstrained cells) (see Figure 2f–i). These maxima correspond to an accumulation of Rab6positive membranes and identify the area where they enter a docking phase before switching to a tethering phase. The localization difference for these maxima between constrained and unconstrained cells is explained by a smaller adhesion area without micropatterns, pushing the docking phase for vesicles closer to the cell periphery. Both radial and angular distributions unraveled by QuantEv represent a measurement of the environment constraints undergone by living cells.
Inwards and outwards Rab6positive membranes show two distinctive dynamical behaviors
Rab6positive membranes are trafficking from the Golgi located at the cell center to the cell periphery (White et al., 1999; Chavrier and Goud, 1999; Echard et al., 2000; Opdam et al., 2000; Grigoriev et al., 2007) and at the same time move from and toward the Golgi in comparable proportions (Grigoriev et al., 2007; Grigoriev et al., 2011). To reconcile these two antagonist statements, we applied QuantEv as follows. Rab6 trajectories were classified into two categories (Figure 3a–c): (i) vesicles moving toward the cell periphery; (ii) vesicles moving toward the Golgi. As shown in Figure 3d–f, the proportion of Rab6positive membranes moving toward the cell periphery and toward the Golgi are close (0.531 versus 0.469 for unconstrained cells, 0.497 versus 0.503 for crossbowshaped cells, 0.521 versus 0.479 for diskshaped cells). However, the radial distributions shown in Figure 3d–f display two distinctive modes for vesicles moving toward the cell periphery and those moving toward the Golgi (p value = 0.0002 for unconstrained cells, p value = 0.021 for crossbowshaped cells, p value = 0.0008 for diskshaped cells). Between the Golgi and the distribution maxima shown in Figure 2f, Rab6 vesicles are predominantly moving toward the cell periphery. Between these maxima and the cell periphery, they are in majority moving toward the Golgi, indicating that during their dockingtethering phase, the vesicles are predominantly moving toward the cell center. These two distinctive dynamical behaviors are consistent with the aforementioned antagonist statements. To go further in the analysis, we looked at the confinement ratio (Figure 4a), the total path length and the lifetime of Rab6 trajectories, conventional dynamical measures used for particle tracking analysis. The combination of these measures with spatial localization is of high interest (Applegate et al., 2011; Tinevez et al., 2017) and QuantEv provides a good framework to quantitatively analyze and visualize the distribution of these dynamical measures with respect to their intracellular localization. We focus on the radial distribution of Rab6 trajectories from unconstrained and constrained cells as the differences between trajectories moving toward cell periphery and trajectories moving toward Golgi lie in these distributions (see Figure 3d–f). Rab6positive membranes moving toward the cell periphery have a much more direct path than the ones moving toward the Golgi, except near the cell periphery (see Figure 4b, e). Consistently, Rab6 positive membranes moving toward the Golgi have longer total path length and lifetime than the ones moving toward the cell periphery, especially when approaching the cell periphery (see Figure 4c–e). In summary, this analysis clearly demonstrated that Rab6positive membranes move predominantly and quite directly from the Golgi to the cell periphery until they enter a docking phase. Then, they mostly go back toward the cell center by following long and indirect trajectories.
The endosomal recycling compartment organizes Rab11 angular distribution
Rab11positive recycling membranes originate their journey from the socalled endosomal recycling compartment (ERC). We formulate the assumption that Rab11 positive membranes are uniformly distributed at the membrane plane around the ERC position within the cell, whatever the cell shape is. To test this hypothesis, we used images acquired at the membrane with TIRF microscopy showing Rab11 proteins (see Appendix 1—figure 1 c–d). Most labeled membranes of the ERC are not located near the cell surface. However, for each TIRF sequence, one highly inclined wide field image was also acquired, enabling to visually define its location (red disks in Figure 5a). To test our assumption, the QuantEv uniformity analysis is applied by considering intensity on segmented regions. The results are shown in Figure 5a (blue disks). To have a line of comparison, we also plot the cell centers as green disks in Figure 5a. Interestingly, the blue disk is close to the red disk for all image sequences except one (second line, middle image in Figure 5a). The blue disk is also closer to the red disk than the green disk in seven out of eight image sequences (see Figure 5a–b). Although the point that gives the most uniform angular distribution does not strictly coincide with the manually identified ERC, it is sufficiently close to indicate that the Rab11positive membranes are quite uniformly distributed around the ERC position at the membrane plane whatever the cell shape is. This indicates that the ERC corresponds to the organizing hub of the Rab11 carrier vesicles.
Joint actin disruption and cell shape influence on Rab11 radial distribution
Applying the QuantEv uniformity analysis at each time step of a sequence allows studying the location stability of the particle emitter or attractor. To test if the estimated ERC location is stationary over time, we computed the Euclidean distance between the reference point estimated at time t = 0 and the points estimated for the next frames. In untreated cells, this distance remains stable (see Figure 6 green line). We analyzed cells treated with Latrunculin A, which inhibits actin polymerization (see Appendix 1—figure 1 e–f). We show that the ERC location is moving away as the drug is affecting the cell (see Figure 6 blue and orange lines), enlightening the role of cytoskeleton in stabilizing the cellular localization of the ERC. We then acquired image sequences of unconstrained, crossbowshaped and diskshaped cells at 10 and 15 min after Latrunculin A addition, and we extracted Rab11 trajectories. The confinement ratio of Rab11 tracks is decreasing with time (see Figure 7), which is consistent with actin cytoskeleton being involved in Rab11 vesicle trafficking, as already reported (Schafer et al., 2014). The radial distribution of Rab11 vesicles is constantly shifting from the cell periphery to the cell center for unconstrained, crossbow and diskshaped cells (see Figure 8a). However, before and at drug injection time, we observe significant differences in radial distributions between the three tested conditions (p value = 0.0023, see Figure 8b). After Latrunculin A treatment, we progressively observe no difference between the radial distributions, as the actin organization is drastically perturbed. Together, these quantifications allow us to conclude that exocytosis/recycling vesicle trafficking is dependent on both cell shape and actin organization.
Discussion
This article presents a computational framework taking into account cell variability to quantify the distribution of fluorescently labeled proteins. Using dynamical descriptors, detailed insight into dynamical processes is also unraveled and the uniformity analysis allows to localize an organizing region for the observed biological objects.
Additionally to the input image, the user has to define three other inputs that depend on the biological application. First, the user has to decide which coordinate system to use. If the imaged cells are flat as in this study (see Appendix 1—figure 1a–c), a cylindrical coordinate system is well suited while a spherical coordinate system will fit better rounded cells. If the user is not familiar with cylindrical or spherical coordinate systems, a classical Cartesian system is also available, even though less suited to intracellular spatial distribution. Finally, QuantEv also allows to analyze the spatial distribution with respect to a reference point or to membrane borders (Heride et al., 2010). Once the reference coordinate system is chosen, the user has to define a reference point, typically the particle emitter or attractor, and a reference direction in order to fairly compare cells. For example, in this study, the direction between the Golgi and the cell center were used to define a reference direction for unconstrained and diskshaped cells while the crossbow principal axis was used for crossbowshaped cells.
As intensity is proportional to the amount of proteins in fluorescence microscopy, using intensity observed in segmented areas is potentially more informative than binary segmentation masks. However, because of phenomena such as photobleaching, phototoxicity, shading, uneven illumination etc., appropriate normalization procedures within and between images need to be applied. If the user is able to correct for these phenomena, it is preferable to use intensity as weights in QuantEv analysis. Otherwise, intensity weights should be avoided.
Given its genericity, QuantEv can easily be applied to any intracellular event and gives useful insights about their spatial distribution across conditions. From these observations, the user can then apply more sophisticated analyses such as mechanistic models of dynamics (Ponti et al., 2004; Jaqaman et al., 2008) or generative models (Li et al., 2012; Johnson et al., 2015a, Johnson et al., 2015b). QuantEv analysis conclusions can also be the starting point of a new modeling.
We demonstrate with the help of QuantEv that the distributions of Rab6positive membranes from unconstrained, crossbow and diskshaped cells are statistically different. QuantEv also enables to identify the locations where Rab6positive membranes enter their docking phase. By considering the directions of the moving Rab6positive membranes, QuantEv allows demonstrating that these membranes first move predominantly and directly toward the cell periphery before reaching their docking phase. They then go back to the cell center in an undirected and long fashion. This intriguing result showing statistically bidirectional movements of Rab6 was reported before. The Rab6positive vesicles generated at the Golgi membranes are predestined to the cell periphery, in order to deliver their exocytic cargo (Grigoriev et al., 2007; Grigoriev et al., 2011), which should favor a centrifuge directionality. Our data reconciles this two apparently opposed observations and show for the first time, that a majority of Rab6 vesicles reverses their movement only toward close dockingfusion sites and only during this ultimate phase of dockingfusion.
QuantEv also demonstrates that Rab11 positive membranes are uniformly distributed around the ERC at the plasma membrane plane. This shows that the ERC represents an organizing hub for the Rab11 carrier vesicles. By applying the QuantEv uniformity analysis along time, we exhibit how the ERC location is affected by actin disruption caused by Latrunculin A injection. The radial distribution analysis of Rab11positive membranes combined with Latrunculin A injection reveals a dual regulation by cell shape and actin organization on Rab11 trafficking at the plasma membrane, and more generally on the exocytosis/recycling vesicle distribution.
In conclusion, QuantEv has the potential to become a very popular analysis method for dynamics and intracellular event analysis as (i) it is publicly available; (ii) it is fully automated and semiparametric; (iii) it provides results that are easy to biologically interpret; (iv) it performs a statistical analysis that takes into account the biological variability over the replicated experiments of a same condition and is efficient with small and large amounts of data. On a biological prospective beyond the two particular models presented here, QuantEv will be of great interest for studies where quantitative and statistical analysis of intracellular membrane or particle behaviors are required, depending on physical and external constraints. For instance, in singlecell experiments performed in microfluidics devices, QuantEv will efficiently provide automation and diversity of statistical analyses in ‘one shot’, for a relatively small amount of data. Applying QuantEv in multicellular systems, in which cellcell constraints necessarily affect molecular distribution and particle movements will also be of great interest. Finally, in vivo imaging of singlecell intracellular processes in a very confined and constrained environment will benefit from the generic aspect of the QuantEv sensing and measuring of particle spatial distributions, dynamical measures with respect to intracellular localization and cell to cell variability. An Icy plugin and a tutorial are available at http://icy.bioimageanalysis.org/plugin/QuantEv. A QuantEv analysis module is available on TrackMate and a QuantEv track processor is available in Icy.
Materials and methods
Sample preparation
In the first dataset, we use cell lines stably expressing fluorescently tagged proteins in order to minimize the celltocell variability in fluorescence signal. HeLa cells stably expressing fluorescently tagged GFPRAB6A were previously generated in the Lab at Institut Curie (Teber et al., 2005). They were maintained in DMEM supplemented with 10% fetal bovine serum. Cells were then spread onto fibronectin Cytoo chips (Cytoo Cell Architect) 4 to 5 hr before imaging. Cell adhesion on micropatterns both constrains the cells in terms of lateral movement and averages their size and shape (diskshaped and crossbowshaped, Cytoo Cell Architect, 1100$\mu $ m${}^{2}$). As a control of patterning effect, the same cell line was grown under the same culture conditions, and spread on regular glass coverslips, 4 to 5 hr before imaging.
For a second set of experiments, wildtype RPE1 cells (hTERT RPE1 obtained from ATCC collection) were grown in Dulbecco’s Modified Eagle Medium, Nutrient Mixture F12 (DMEM/F12) supplemented with 10% (vol/vol) FCS in sixwell plates. RPE1 cells were transiently transfected with plasmids coding for Rab11aGFP, and LangerinmCherry using the following protocol: 2 $\mu $g of each DNAs, completed to 100 $\mu $L with DMEM/F12 (FCS free) were incubated for 5 min at room temperature. 6 $\mu $L of XtremeGENE 9 DNA Transfection Reagent (Roche) completed to 100 $\mu $L with DMEM/F12 (FCS free) were added to the mix and incubated for further 15 min at room temperature. The transfection mix was then added to RPE1 cells grown 1 day before and incubated further at 37${}^{o}$C overnight. Cells were then spread on regular coverslips or onto fibronectin Cytoo chips (Cytoo Cell Architect) for 4 hr at 37${}^{o}$C with F12 (with 10% (vol/vol) FCS, 10 mM Hepes, 100 units/ml of penicillin and 100 ug/ml of Strep) before imaging. When specified, 2 mM Latrunculin A (Sigma) was dissolved to 0.02 mM in F12 DMEM. 300 $\mu $L of culture medium with Latrunculin A (600 nM) was added to establish a final Latrunculin A concentration of 3 $\mu $M.
All cell lines were routinely tested for mycoplasma, using PCR or the MycoAlert Mycoplasma Detection Assay.
Data acquisition
For Rab6positive membranes in unconstrained cells, videos were recorded with an epifluorescence video automated system composed of a Ti Eclipse inverted microscope equipped with a 100x objective Plan NA (1.4) and a piezo stage for 3D acquisitions (Nikon, S.A, France). The fluorescence was collected using a 512 × 512 EMCCD (Evolve, Photometric, USA) and driven through the Metamorph software (Molecular Devices). 18 series of 120 Z image stacks of 10 frames were recorded at a rate of about 1 stack/s. The volume rendering of one image from this dataset is shown in Appendix 1—figure 1a.
For Rab6positive membranes on micropatterns, the 488 nm laser of a spinningdisk confocal microscope (Ti Eclipse, Nikon, S.A, France equipped with spinning disk system, a 100x/1.4 oil objective and CoolSnap HQ2 CCD, from Roper Scientific S.A.R.L, France) was used to acquire 3D 380 $\times $ 380 $\times $ 8 stacks at a rate of one stack per second. 18 image sequences with crossbowshaped cells and 22 image sequences with diskshaped cells were acquired. The system was driven by the Metamorph software (Molecular Devices). The volume rendering of two images from this dataset are shown in Appendix 1—figure 1b–c.
For the Rab11 dataset, livecell imaging was performed using simultaneous dual color Total Internal Reflection Fluorescence (TIRF) microscopy. All imagings were performed in full conditioned medium at 37${}^{o}$C and 5% CO2 unless otherwise indicated. Simultaneous dual color TIRF microscopy sequences were acquired on a Nikon TE2000 inverted microscope equipped with a 100x TIRF objective (NA = 1.49), an azimuthal TIRF module (Ilas2, Roper Scientifc), an image splitter (DV, Roper Scientific) installed in front of an EMCCD camera (Evolve, Photometrics) that can be bypassed or not, depending on the experimental conditions, as indicated in the text, and a temperature controller (LIS). GFP and mCherry were excited with a 488 nm and a 561 nm laser, respectively (100 mW). The system was driven by the Metamorph software (Molecular Devices). Four selected image projections from this data set are shown in Appendix 1—figure 1d–g.
Data availability
We use two datasets in this study that are publicly available on the iMANAGE database at https://cid.curie.fr/iManage/standard/login.html with username public and password Welcome!1 in the project entitled QuantEvData.
Event detection and localization
Before applying QuantEv, the intracellular events have to be identified and localized. The Rab6 proteins are extracted from each image sequence by using the CCRAFT method (Pécot et al., 2015) with default parameters, except the p value that ranges from 0.0025 to 0.35 depending on the noise level, available on Icy (de Chaumont et al., 2012). The Rab11positive membranes on micropatterns are segmented at each time point with the ATLAS algorithm (Basset et al., 2015) with default parameters, except the p value that ranges from 0.05 to 0.45 depending on the noise level. In both cases, a variance stabilization transform (Boulanger et al., 2010) is performed to take into account the PoissonGaussian nature of the noise in the CCD sensors. As unconstrained cells are more mobile than cells on micropatterns, the image sequences showing Rab11positive membranes in unconstrained cells are not in focus. To correct this phenomenon, a deconvolution method (Lefkimmiatis et al., 2012) is first applied to the image sequences. The Rab11 positive membranes are then segmented at each time point with the Bernsen local thresholding method (Bernsen, 1986) (radius equal to 15 pixels). Finally, Rab6 and Rab11 trajectories are estimated with the multiple hypothesis tracking method developed by Chenouard et al. (2013) with default parameters, available on Icy (de Chaumont et al., 2012), the combinatorial optimization tracking method developed by Sbalzarini and Koumoutsakos (2005) with default parameters, available on ImageJ (Schneider et al., 2012) and the hybrid approach TrackMate (Tinevez et al., 2017) that first connects detected points into short tracks and then links the resulting tracks together, with default parameters, available on Fiji (Schindelin et al., 2012). To identify the trajectories estimated with different methods, we use the gated distance (Chenouard et al., 2014) defined between two trajectories ${\theta}_{1}$ and ${\theta}_{2}$ as:
where $\u03f5$ is the gate. For each image sequence, the gated distance is computed between the trajectories estimated with the three different methods with $\u03f5=5$ pixels. Only the trajectories for which the gated distance is inferior to 2 pixels for at least two methods are used for the analysis.
Weighted density estimation for spatial localization
The localization of events needs to be defined on a common coordinate system to compare the experiments. We propose to use the cylindrical coordinate system where only a reference point such as the event emitter or attractor and a reference direction have to be specified by the user. To fairly compare experiments with different cell shapes, we define appropriate distances to obtain normalized densities, that is independent from the cell shape. We illustrate the importance of shape normalization in Appendix 2.
More formally, let us define $\Omega $ the 3D cell support and $\partial \Omega $ the 3D cell surface. Let us consider a set of $N$ sample points associated with intracellular events $S=\left\{\right({r}_{i},{\theta}_{i},{z}_{i},{w}_{i},{d}_{{\theta}_{i}},{d}_{{z}_{i}}),i\in [1,N\left]\right\}$, where $({r}_{i},{\theta}_{i},{z}_{i})$ denote the spatial cylindrical coordinates. The weight ${w}_{i}$ enables to take into account features associated to events such as intensity, track length, confinement ratio... ${w}_{i}$ can typically be a function of fluorescence intensity, proportional to the number of molecules observed at a given location. The distance ${d}_{{\theta}_{i}}$ is equal to the Euclidean distance between the coordinate system origin $O\in \Omega $ projected on plane ${z}_{i}$ (${O}_{{z}_{i}}$) and the point ${P}_{{\theta}_{i},{z}_{i}}\in \partial \Omega $ with angle ${\theta}_{i}$ at plane ${z}_{i}$, such that ${d}_{{\theta}_{i}}=\left\right{P}_{{\theta}_{i},{z}_{i}}{O}_{{z}_{i}}{}_{2}^{2}$. The distance ${d}_{{z}_{i}}$ is equal to the Euclidean distance between the coordinate system origin $O$ and the point ${P}_{{r}_{i},{\theta}_{i}}\in \partial \Omega $ with radius ${r}_{i}$ and angle ${\theta}_{i}$ such that ${d}_{{z}_{i}}=\left\right{P}_{{r}_{i},{\theta}_{i}}O\left\right$. These two distances allow estimating normalized densities that are independent from cell shapes. We propose to estimate three densities defined as follows:
where ${G}_{\hat{\sigma}}(\cdot )$ is a Gaussian kernel with bandwidth $\hat{\sigma}$, $H}_{\hat{\kappa}$ is a von Mises kernel with concentration $\hat{\kappa}$ such that $H}_{\hat{\kappa}}\left(\theta \right)=\frac{{e}^{\hat{\kappa}cos\theta}}{2\pi {I}_{0}\left(\hat{\kappa}\right)$ and ${I}_{0}(\cdot )$ is the Bessel function of order 0. The bandwidths $\hat{\sigma}}_{r$ and $\hat{\sigma}}_{z$ are estimated with the Silverman’s rule of thumb (Silverman, 1986) and $\hat{\kappa}$ is estimated using the robust rule of thumb proposed by Taylor (2008). The normalization constants are defined as follows:
Density estimation for dynamical features
In case the distribution of dynamical features such as confinement ratio or lifetime with respect to the track localization is to be studied, the weighted densities are defined differently than in the previous section. In this case, histograms are first computed as the averaged dynamic features for each bin. A density estimation is then estimated from the histograms.
Let us consider a set of $T$ tracks associated with spatial coordinates $T=\left\{\right({r}_{i},{\theta}_{i},{z}_{i},{m}_{i}),i\in [1,T\left]\right\}$, where $({r}_{i},{\theta}_{i},{z}_{i})$ denote the spatial cylindrical coordinates of the median point of the trajectory $i$ and ${m}_{i}$ is a dynamic feature associated to track $i$. The histograms corresponding to the averaged dynamic features for each bin are defined as:
where ${b}_{r}\in [1,{B}_{r}]$ is a radius bin and ${B}_{r}$ is the total number of radius bins, ${b}_{\theta}\in [1,{B}_{\theta}]$ is a polar bin and ${B}_{\theta}$ is the total number of polar bins, ${b}_{z}\in [1,{B}_{z}]$ is an indepth bin and ${B}_{z}$ is the total number of indepth bins, and ${1}_{{b}_{r}}\left[{r}_{i}\right]$ is equal to $1$ if ${r}_{i}$ is defined in bin ${b}_{r}$ and equal to 0$0$ otherwise. Densities are then estimated from the histograms as follows:
where ${G}_{\hat{\sigma}}(\cdot )$ is a Gaussian kernel with bandwidth $\hat{\sigma}$, $H}_{\hat{\kappa}$ is a von Mises kernel with concentration $\hat{\kappa}$. The bandwidths $\hat{\sigma}}_{r$ and $\hat{\sigma}}_{z$ are estimated with the Silverman’s rule of thumb (Silverman, 1986) and $\hat{\kappa}$ is estimated using the robust rule of thumb proposed by Taylor (2008).
Statistical procedure
Quantitative comparison between different conditions is mandatory to analyze biological data. In most computational biology studies, data from different experiments corresponding to the same condition are pooled together (Schauer et al., 2010; Merouane et al., 2015). This usual procedure enables to add statistical power when comparing two conditions. Therefore, it is especially useful when few data are available. Unfortunately, pooling data together presents two main drawbacks. First, if large amounts of data are available, the opposite problem arises and the statistical tests may become significant for every comparison (Olivier and Walter, 2015). One solution is to down sample the data, but the amount of down sampling becomes another issue. Second, pooling data together for one condition partially hides the variability between the replicated experiments for this condition. As an example, let us consider a study aimed at analyzing the effects of a drug on a sample of normal individuals. To evaluate the drug efficiency, a comparison between normal individuals and individuals that were administered the drug is conducted. Let us assume that the drug is effective on half the individuals. Consequently, normal individuals are compared to a mix of normal individuals and individuals with the drug effects. This comparison should not be statistically significant as the drug is not efficient on all individuals. However, the effects on the individuals for which the drug is efficient might hide the fact that it is not efficient on all individuals if all the data are pooled together. In what follows, we propose to compute a distance between all experiments instead of a distance between conditions. The idea is demonstrated in Appendix 3 and validated on synthetic image sequences (see Appendix 1—figure 1c–d and Appendix 3).
Distance between densities
We propose to compute the earth mover’s distance (also known as the KantorovichRubinstein or the first order Wasserstein distance) between every replicate of every condition to apply a statistical test. This transportbased distance demonstrated its efficiency for other studies on cell phenotypes (Wang et al., 2013; Basu et al., 2014). The discrete Earth Mover’s Distance (EMD) between two unidimensional distributions is simply defined as the sum of the absolute differences between their cumulated distribution functions (Rubner et al., 2000):
where ${F}^{1}$ and ${F}^{2}$ are the cumulated distribution functions of ${f}^{1}$ and ${f}^{2}$. Although the EMD depends on the number of bins $K$, EMD proportions are kept intact when the number of bins is high enough as shown in Appendix 4. For the angular distribution, the Circular Earth Mover’s Distance (CEMD) (Rabin et al., 2011) is defined as:
with
Difference between conditions
The EMD and CEMD enable to compute a distance between two single experiments for the radial, angular and indepth densities. The distances between the replicates of one condition and the replicates of the other condition(s) give an idea about the difference between the conditions. However, a baseline distance is also needed to state if the difference is random or significant. Therefore, two distances are defined for each experiment and each density:
the intracondition distance: average distance between the density and all the other densities for the same condition;
the intercondition distance: average distance between the density and all the other densities from the other condition(s).
Considering more than two groups does not change the intracondition distance and only expands the intercondition distance to more than one group. We define as the condition difference the difference between the intercondition distance and the intracondition distance. If the condition difference is high, the conditions are different.
Statistical test
A statistical test is applied on the difference distance to state if the observed conditions are significantly different. A nonparametric statistical test is better suited as there is no underlying model for the condition difference. In addition, a negative condition difference implies that the current experiment is closer to the replicated experiments of the other condition than the replicated experiments of the same condition. Consequently, the condition difference has to be positive if the conditions are different. For those two reasons, we propose to use the onesided nonparametric Wilcoxon signedrank test on the condition differences for all experiments to state if two conditions are statistically different.
Analysis of uniform distribution of events
In case we focus on the intracellular events assumed to be uniformly distributed around a given biological object, for example the events emitter, QuantEv allows us to estimate a location for this trafficorganizing component. This source location is then defined as the reference point with the most uniform angular distribution. It is established that the maximum entropy corresponds to the most uniform distribution. Consequently, the reference point ${O}^{\ast}$ is defined as the location that maximizes the entropy:
The most straightforward way to find this point is to estimate the entropy map that gives, for each point in $\Omega $, the entropy value computed with the current point used as the reference center. We also propose to use the bisection method to speed up the computation (about sixty times faster than the entropy map computation, see Appendix 5—table 1). A uniformity analysis conducted on simulations is presented in Appendix 6. The entropy criterion can be extended to detect multiple organizing components if needed.
Code availability
The jar file of the QuantEv Icy plugin is available at http://icy.bioimageanalysis.org/plugin/QuantEv. The jar file of the QuantEv track processor is available at http://icy.bioimageanalysis.org/plugin/ QuantEv _(track_processor). The source codes can be extracted from the jar files. The QuantEv analysis module for TrackMate is available on GitHub (Pécot, 2018; copy archived at https://github.com/elifesciencespublications/QuantEvForTrackMate). These codes are released under the GNU Affero General Public License v3.0.
Appendix 1
Datasets
An example for each condition of the two datasets used in this study is shown in Appendix 1—figure 1.
The four different scenarios of simulated images that were generated to evaluate QuantEv performance are shown in Appendix 1—figure 2.
Appendix 2
Sensitivity to cell shape
The cell shape influences the spatial distribution of intracellular events. The distances ${d}_{\theta}$ and ${d}_{z}$ were introduced to compute a distribution that is invariant from the cell shape (see Section Weighted density estimation). To quantify the cell shape influence and to evaluate the pertinence of the normalization with distances, we generate image sequences with vesicles trafficking on a squareshaped region. In these simulations, vesicles are uniformly distributed over 16 different paths and are moving from the cell center to the cell periphery (see Appendix 1—figure 2a). As the cell is squareshaped, the vesicles moving to the cell corners travel a longer distance than the other vesicles so the number of vesicles on these paths is higher. Consequently, the spatial distribution of vesicles is not uniform over the radius and angle ranges (see purple histograms in Appendix 2—figure 1). Nevertheless, the vesicles are generated over the paths with an equal probability in the simulations, meaning that the distribution over the different paths is uniform. By weighting the distribution of spatial coordinates with the distance between the cell center and the cell periphery, the shape dependence is accurately corrected as shown in the green histograms of Appendix 2—figure 1. Although other registration approaches are more powerful (Zhao and Murphy, 2007; Peng and Murphy, 2011), the results shown in Appendix 2—figure 1 demonstrate that these sophisticated methods are not needed in our case.
Appendix 3
Statistical analysis
To evaluate the effect of pooling data together on the statistical analysis, 20 image sequences with uniform distribution over the paths (Appendix 1—figure 2b) and 10 image sequences with isotropic distribution over the paths (Appendix 1—figure 2c) are generated. Four groups are then defined from these simulations:
group #1: 10 image sequences with uniform distribution;
group #2: 10 other image sequences with uniform distribution;
group #3: 10 image sequences with isotropic distribution (six paths with a probability equal to $0.1$ and 10 paths with a probability equal to $0.04$);
group #4: five image sequences with uniform distribution and five image sequences with isotropic distribution.
The analysis of variance (ANOVA), usual method for biological studies, is compared to the QuantEv statistical approach. For the ANOVA analysis, the vesicle mass centers are extracted from the simulations and the pair $(r,\theta )$ is used to compare two groups. The intensity observed in the segmented vesicles is used for the QuantEv approach. For both methods, several amounts of data are considered: from 1% to 100% data for the ANOVA analysis; from 2 vs. 2 to 10 vs. 10 image sequences for QuantEv.
With the ANOVA analysis on pooled data, the pvalues are low with a small amount of data when comparing groups #1 and #3 (see Appendix 3—figure 1a). But they also start to be low when comparing groups #1 and #2 for an amount of data that reaches about 50% (see Appendix 3—figure 1a). These results indicate that there is a gradient of pvalues consistent with actual differences between the spatial distributions. However, the values lead to a significant difference between all groups (see Appendix 3—figure 1a). It demonstrates that it is difficult to deal with pooled data when the amount of data is high. When comparing groups #1 and #4, there should not be any statistical difference, as group #4 is constituted of particles with different distributions. But the ANOVA analysis on the pooled data is not able to grasp this variability between replicated experiments of a same condition and the pvalues are low with a small amount of data (about 5%, see Appendix 3—figure 1a).
The QuantEv statistical approach does not lead to any statistical difference for radius for the three comparisons (see Appendix 3—figure 1b), a result that is consistent with the data. By using QuantEv, it turns out that angular distributions are statistically different when comparing groups #1 and #3 while they are not for the two other comparisons (see Appendix 3—figure 1b). These experiments demonstrate that the QuantEv statistical approach is not disturbed by large amounts of data because it considers the distributions over the sequences. They also demonstrate that QuantEv takes into account the variability between replicated experiments of a same condition as the comparison involving groups #1 and #4 does not conclude to any statistical difference.
Appendix 4
Binning influence on earth mover’s distance
The QuantEv statistical procedure relies on the circular and regular earth mover’s distances, computed from densities defined from binned data. To evaluate the effect of binning on these distances, we computed the Earth Mover’s Distances (EMD) between $cos\left(x\right)$ and $cos\left(2x\right)$ (purple line in Appendix 4—figure 1a) and between $sin\left(x\right)$ and $cos\left(2x\right)$ (green line in Appendix 4—figure 1a) with a number of bins ranging from $1$ to $500$. As expected from (Bernsen, 1986), the EMD increases with the number of bins. However, the proportion between distances remains quite constant when the number of bins is higher than $25$ (see Appendix 4—figure 1b). Consequently, if the EMD between two distributions $f$ and $g$ is higher than the EMD between two distribution $f$ and $h$ for a given number of bins ${n}_{1}$, the EMD between $f$ and $g$ is also higher than the EMD between $f$ and $h$ for a number of bins ${n}_{2}$. As the statistical test used in the QuantEv framework (onesided Wilcoxon signedrank test) is based on the ranks of the difference between interEMD and intraEMD, the result of the test is not affected by the number of bins.
Appendix 5
Processing time
QuantEv computation time to process images with different size and object coverage is shown in Appendix 5—table 1.
Appendix 6
Uniform distribution of events
To evaluate the QuantEv uniformity analysis, we simulate 10 image sequences with particles uniformly distributed over the different paths on a network for which the origin is not centered in the image (Appendix 1—figure 2d). Appendix 6—figure 1a shows the entropy map obtained for one simulation. Appendix 6—figure 1b shows the different reference points estimated over the ten simulations as green disks. These results are not perfect, as the reference centers are not estimated to be located at the exact particle emitter location. However, if the particles are distributed with equal probability on all paths, this does not imply that the actual number of generated particles is the same on all paths so the estimation cannot be perfect. The estimated reference points for these simulations are close to the particle emitter location, which demonstrates the potential of this approach.
References

1
plusTipTracker: Quantitative image analysis software for the measurement of microtubule dynamicsJournal of Structural Biology 176:168–184.https://doi.org/10.1016/j.jsb.2011.07.009
 2

3
Phenotypic characterisation of RAB6A knockout mouse embryonic fibroblastsBiology of the Cell 107:427–439.https://doi.org/10.1111/boc.201400083

4
Adaptive spot detection with optimal scale selection in fluorescence microscopy imagesIEEE Transactions on Image Processing 24:4512–4527.https://doi.org/10.1109/TIP.2015.2450996
 5

6
Dynamic thresholding of greylevel imagesEighth International Conference on Pattern Recognition. Proceedings. pp. 1251–1255.
 7

8
Patchbased nonlocal functional for denoising fluorescence microscopy image sequencesIEEE Transactions on Medical Imaging 29:442–454.https://doi.org/10.1109/TMI.2009.2033991
 9

10
The role of ARF and Rab GTPases in membrane transportCurrent Opinion in Cell Biology 11:466–475.https://doi.org/10.1016/S09550674(99)800672

11
Multiple hypothesis tracking for cluttered biological image sequencesIEEE Transactions on Pattern Analysis and Machine Intelligence 35:2736–3750.https://doi.org/10.1109/TPAMI.2013.97

12
Objective comparison of particle tracking methodsNature Methods 11:281–289.https://doi.org/10.1038/nmeth.2808
 13
 14

15
Alternative splicing of the human Rab6A gene generates two close but functionally different isoformsMolecular Biology of the Cell 11:3819–3833.https://doi.org/10.1091/mbc.11.11.3819
 16

17
Rab6 regulates transport and targeting of exocytotic carriersDevelopmental Cell 13:305–314.https://doi.org/10.1016/j.devcel.2007.06.010
 18

19
Distance between homologous chromosomes results from chromosome positioning constraintsJournal of Cell Science 123:4063–4075.https://doi.org/10.1242/jcs.066498
 20

21
Joint modeling of cell and nuclear shape variationMolecular Biology of the Cell 26:4046–4056.https://doi.org/10.1091/mbc.e15060370
 22

23
Hessianbased norm regularization for image restoration with biomedical applicationsIEEE Transactions on Image Processing 21:983–995.https://doi.org/10.1109/TIP.2011.2168232
 24

25
Imagining the future of bioimage analysisNature Biotechnology 34:1250–1255.https://doi.org/10.1038/nbt.3722
 26
 27

28
The small GTPase Rab6B, a novel Rab6 subfamily member, is celltype specifically expressed and localised to the golgi apparatusJournal of Cell Science 113:2725–2735.

29
Bioimage informatics: a new area of engineering biologyBioinformatics 24:1827–1836.https://doi.org/10.1093/bioinformatics/btn346
 30
 31

32
Background fluorescence estimation and vesicle segmentation in live cell imaging with conditional random fieldsIEEE Transactions on Image Processing 24:667–680.https://doi.org/10.1109/TIP.2014.2380178
 33

34
Transportation distances on the circleJournal of Mathematical Imaging and Vision 41:147–167.https://doi.org/10.1007/s1085101102840

35
The earth mover’s distance as a metric for image retrievalInternational Journal of Computer Vision 40:99–121.https://doi.org/10.1023/A:1026543900054

36
Feature point tracking and trajectory analysis for video imaging in cell biologyJournal of Structural Biology 151:182–195.https://doi.org/10.1016/j.jsb.2005.06.002
 37
 38

39
Fiji: an opensource platform for biologicalimage analysisNature Methods 9:676–682.https://doi.org/10.1038/nmeth.2019

40
NIH Image to ImageJ: 25 years of image analysisNature Methods 9:671–675.https://doi.org/10.1038/nmeth.2089
 41

42
Automatic bandwidth selection for circular density estimationComputational Statistics & Data Analysis 52:3493–3500.https://doi.org/10.1016/j.csda.2007.11.003

43
Rab6 interacts with the mint3 adaptor proteinBiological Chemistry 386:671–677.https://doi.org/10.1515/BC.2005.078

44
The extracellular matrix guides the orientation of the cell division axisNature Cell Biology 7:947–953.https://doi.org/10.1038/ncb1307
 45

46
Rab11 regulates recycling through the pericentriolar recycling endosomeThe Journal of Cell Biology 135:913–924.https://doi.org/10.1083/jcb.135.4.913

47
A linear optimal transportation framework for quantifying and visualizing variations in sets of imagesInternational Journal of Computer Vision 101:254–269.https://doi.org/10.1007/s112630120566z

48
Rab6 coordinates a novel Golgi to ER retrograde transport pathway in live cellsThe Journal of Cell Biology 147:743–760.https://doi.org/10.1083/jcb.147.4.743
 49
Decision letter

Anna AkhmanovaSenior and Reviewing Editor; Utrecht University, Netherlands
In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.
Thank you for submitting your article "A quantitative approach for the spatiotemporal distribution of 3D intracellular events in fluorescence microscopy" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by Anna Akhmanova as the Senior/Reviewing Editor. The reviewers have opted to remain anonymous.
The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.
Summary:
The paper presents a software tool (QuantEv) for quantitative analysis and visualization of the spatial distribution of intracellular events imaged by fluorescence microscopy and represented by static or dynamic descriptors associated with spatial coordinates. The potential practical value of the tool is demonstrated by studying the distribution of moving Rab6 fluorescently labeled membranes with respect to their direction of movement in differently shaped cells, as well as the position of the generating hub of Rab11 positive membranes, and the effects of actin disruption on Rab11 trafficking in relation to cell shape. The paper is well written, and the results are interesting.
Essential revisions:
1) The manuscript ignores extensive prior relevant work on similar problems. The analysis methods described are simple, and not particularly innovative. Evidence of generalizability is lacking. By not considering and incorporating prior approaches, the potential capabilities and applicability of the software have been significantly limited. It is especially important in describing new software to compare its performance to existing approaches. These comparisons should include not just one similar method (kernel density estimation) that is not widely used but other methods (e.g., http://doi.org/10.1038/nmeth.1486). For example, although developed for a different application, the software tool plusTipTracker (http://dx.doi.org/10.1016/j.jsb.2011.07.009) can also perform various dynamics analyses related to cell location, and should be discussed. The same goes for TrackMate (http://doi.org/10.1016/j.ymeth.2016.09.016). More generally, some discussion is needed of what, exactly, can and cannot be done with existing tools, to make the novelties and benefits of the proposed tool more explicit.
2) It seems that the statistical test the authors are proposing is focusing on comparing two groups using the Wilcoxon signedrank test. While this is fine as such, often in Biology more than two groups need to be compared with each other; like wt, mutant, and rescue for example. Also, the possibility to compare more than two groups is important to avoid problems with repetitive testing between groups. The problem is the additive per comparison error. This issue should be addressed, and potential solutions should be included in the next version.
3) The authors suggest using intensity as a weight for the analysis. In the cases used in the synthetic test images test and likely in the experimental data here the intensities are comparable. However, the potential user needs to be instructed that appropriate normalization procedures need to be applied in case that intensities are used as weight. Likewise, the segmentation needs to start from a similar selection. The authors should at least discuss this necessity and provide what the prerequisites are for the input.
4) The only datasets used in paper are artificial, in that cells were constrained to specific geometries. This reduces the inherent complexity with unknown other effects. Most investigators would not choose to use artificial geometric constraints, and no analysis is presented for images of cells that show natural variation in shape, either in vitro or in vivo. Such variation might overwhelm the straightforward approaches the authors describe and this possibility should be investigated. Application of the methods to an image dataset for unconstrained cells should be included.
5) The claim that the proposed framework is "generic and nonparametric" seems too strong. In the paper only a few very specific applications are investigated. And many of the underlying components of the framework are not exactly nonparametric. For example, the kernels involved in the weighted density estimation have parameters, and the distance measures depend on the number of bins. This claim should be toned down.
6) The authors claim that their framework is more sensitive than the Kernel density maps. This asks the question of the discriminatory power of the method. The potential to differentiate distribution patterns depends on the resolution of the input; this should be discussed.
7) The meaning of the results of the various analyses is often unclear and very limited in terms of providing understanding of mechanisms for any of the systems studied. Since the paper is written for a general audience, this should be improved.
8) A number of specific comments on the text must be addressed:
Introduction, first paragraph “Automatic methods have the obvious advantage of being quicker and reproducible. However, most computational methods are based on the complex combination of heterogeneous features such as statistical, geometrical, morphological and frequency properties (Peng, 2008), whichmakes difficult to draw de1nitive biological conclusions”: This statement ignores extensive work on generative or mechanistic models, which produce interpretable parameters. Such work includes mechanistic models of dynamics of endocytic vesicles (e.g., http://doi.org/10.1038/nmeth.1237) and cytoskeletal dynamics (http://doi.org/10.1126/science.1100533), and generative models of vesicle distribution (e.g., http://doi.org/10.1371/journal.pcbi.1004614).
Introduction, first paragraph “Additionally, most experimental designs, especially at single cell level, pool together data coming from replicated experiments of a given condition (Schauer et al., 2010; Merouane et al., 2015; Biot et al., 2016), neglecting the biological variability between individual cells.”: Again, this ignores work on generative models that specifically analyzes and captures variation between cells. Past examples include microtubule networks (e.g., http://doi.org/10.1371/journal.pone.0050292), and cell and nuclear shape (e.g., http://dx.doi.org/10.1091/mbc.E15060370). Traditional featurebased methods also frequently analyze heterogeneity within populations (e.g., http://doi.org/10.1371/journal.pone.0102678).
Introduction, third paragraph and subsection “Weighted density estimation”, first paragraph – The use of circular and/or cylindrical coordinate systems for description of object positions within a cell is well established (e.g., http://doi.org/10.1002/cyto.a.20487 and http://doi.org/10.1002/cyto.a.21066) and in these cases rotation angle was more powerfully defined relative to the major axis of each cell rather than being defined by the confinement fields. Alternative approaches to the problem, such as morphing, were not discussed.
Introduction, third paragraph and subsection “Distance between densities”, first paragraph – There is no discussion of more recent metrics related to Earth Mover's Distance that have been described and used to compare subcellular patterns (http://doi.org/10.1007/s112630120566z).
"The KD approach concludes… Instead, QuantEv selectively identifies…" How do we know which method comes closest to the truth? Can this be verified? Without some control experiment or simulation, how can we conclude that QuantEv is to be preferred over other methods?
Subsection “Visualizing and quantifying the influence of micropatterns on the spatial distribution of Rab6 positive membranes” – There is no clear peak at the twothirds position in Figure 2D and no evidence of significance or reproducibility is presented.
"Rab6 trajectories were classified into two categories…" How do we know these trajectories are trustworthy? What kind of control experiment was performed to confirm this? This is especially important since it seems the trajectories were not obtained with the best methods available these days (for example according to http://doi.org/10.1038/nmeth.2808 there seem to be better tracking methods than the method mentioned in the subsection “Event detection and localization”).
"we extracted Rab11 trajectories…" Same as previous comment.
"On the image sequences considered in the previous section (see Figure 4A), this distance remains stable (see Figure 5A). We analyzed cells treated with Latrunculin A… the ERC location is moving away as the drug is affecting the cell (see Figure 5B)". But the time scales are very different in these two cases (seconds versus minutes). Control experiments would be needed to confirm that the ERC location in nontreated cells remains stable over the same time scale as in the treated cells.
The claim that the presented software tool "is efficient with small and large amounts of data" has not been demonstrated in the paper. Neither dataset sizes nor processing times are mentioned.
The claim that "QuantEv is quite flexible since the user can specify any distance…" contradicts the statement that "it is fully automated and nonparametric".
"a reference point… and a reference direction have to be specified by the user…" Same as previous comment.
[Editors' note: further revisions were requested prior to acceptance, as described below.]
Thank you for resubmitting your work entitled "A quantitative approach for the spatiotemporal distribution of 3D intracellular events in fluorescence microscopy" for further consideration at eLife. Your revised article has been favorably evaluated by Anna Akhmanova (Senior/Reviewing Editor) and three reviewers.
The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:
There were significant reservations about using constrained cells, as the use of such cells greatly simplifies the analysis. The authors have now done additional work to add results comparing unconstrained cells and two different types of constrained cells. The results show that QuantEV is able to distinguish among the three groups. However, no perturbation studies (e.g., Latrunculin B) were done with unconstrained cells. Thus the main concern remains about the suitability of QuantEV for use in future studies, the majority of which are expected to be done with unconstrained cells. This is an important point: the method may be able to distinguish changes within constrained cells upon various treatments, but may not be able to distinguish perturbations on the background of significant variation within unconstrained cells. There is no information provided on the variance of the profiles in Figure 2F within the unconstrained population. This is a major concern in the context of the very broad claims made in the manuscript (especially in the Discussion) about the power and generality of QuantEV. To support these broad claims, the authors must provide conclusive evidence that QuantEV can distinguish physiologically relevant changes upon perturbations in unconstrained cells. Since the necessary datasets are undoubtedly available to the authors, no collection of new experimental data is expected to be necessary to address this point.
https://doi.org/10.7554/eLife.32311.040Author response
Essential revisions:
1) The manuscript ignores extensive prior relevant work on similar problems. The analysis methods described are simple, and not particularly innovative. Evidence of generalizability is lacking. By not considering and incorporating prior approaches, the potential capabilities and applicability of the software have been significantly limited. It is especially important in describing new software to compare its performance to existing approaches. These comparisons should include not just one similar method (kernel density estimation) that is not widely used but other methods (e.g., http://doi.org/10.1038/nmeth.1486). For example, although developed for a different application, the software tool plusTipTracker (http://dx.doi.org/10.1016/j.jsb.2011.07.009) can also perform various dynamics analyses related to cell location, and should be discussed. The same goes for TrackMate (http://doi.org/10.1016/j.ymeth.2016.09.016). More generally, some discussion is needed of what, exactly, can and cannot be done with existing tools, to make the novelties and benefits of the proposed tool more explicit.
We agree that plusTipTracker and TrackMate showed analyses that associate dynamic features with localization, and we failed to report these studies in the first version of the manuscript. However, plusTipTracker and TrackMate do not provide a framework to quantitatively analyze the dynamic features with respect to intracellular localization, which QuantEv does. Actually, neither precise way to look at localization nor statistical tool are available with these tracking methods. Nonetheless, studying dynamic features such as confinement ratio or lifetime with respect to their intracellular localization is not currently available for biologists while it potentially represents a routine analysis for tracking experiments. Consequently, we proposed to the developers of TrackMate to add a QuantEv analysis module in their plugin and to the Icy developers to add a QuantEv track processor. They both accepted, demonstrating a need for this type of analysis. We are currently collaborating with JeanYves Tinevez, TrackMate developer, to implement a QuantEv analysis module (https://github.com/tpecot/QuantEvForTrackMate), the QuantEv track processor will follow. We added a new section explaining how we compute histograms and densities of dynamical features with respect to localization (subsection “Density estimation for dynamical features”). We also changed the second paragraph of the Results to analyze the radial distribution of confinement ratio, total path length and lifetime instead of using these features as weights in the densities.
2) It seems that the statistical test the authors are proposing is focusing on comparing two groups using the Wilcoxon signedrank test. While this is fine as such, often in Biology more than two groups need to be compared with each other; like wt, mutant, and rescue for example. Also, the possibility to compare more than two groups is important to avoid problems with repetitive testing between groups. The problem is the additive per comparison error. This issue should be addressed, and potential solutions should be included in the next version.
We thank the reviewers for this remark, we actually did not think about comparing more than two groups together. The statistical test is applied on the difference between the average interdistance and the average intradistance for each image sequence in the study. Considering more than two groups does not change the average intradistance and only expands the average interdistance to more than one group. Accordingly, we changed the text in the section entitled “Difference between conditions” and modified the plugin to compare more than two groups.
3) The authors suggest using intensity as a weight for the analysis. In the cases used in the synthetic test images test and likely in the experimental data here the intensities are comparable. However, the potential user needs to be instructed that appropriate normalization procedures need to be applied in case that intensities are used as weight. Likewise, the segmentation needs to start from a similar selection. The authors should at least discuss this necessity and provide what the prerequisites are for the input.
We acknowledge that intensity normalization is an important step in the analysis of fluorescent images and needs to be discussed. As intensity is proportional to the amount of proteins in fluorescence microscopy, it potentially provides useful information. However, several phenomena such as photobleaching, phototoxicity, shading or uneven illumination potentially alter this proportionality. If the user is able to correct for these phenomena, it is preferable to use intensity as weights for the analysis. Otherwise, it is safer not to use it. We added a paragraph in the Discussion (third paragraph) to address this point.
4) The only datasets used in paper are artificial, in that cells were constrained to specific geometries. This reduces the inherent complexity with unknown other effects. Most investigators would not choose to use artificial geometric constraints, and no analysis is presented for images of cells that show natural variation in shape, either in vitro or in vivo. Such variation might overwhelm the straightforward approaches the authors describe and this possibility should be investigated. Application of the methods to an image dataset for unconstrained cells should be included.
Thank you for this valuable suggestion. In the revised manuscript, we added a set of image sequences with Rab6 positive membranes in unconstrained cells (Appendix 1 Figure 1A) and compared them with crossbow and diskshaped cells (subsections “Visualizing and quantifying the influence of cell shape on the spatial distribution of Rab6 positive membranes” and “Inwards and outwards Rab6 positive membranes show two distinctive dynamical behaviors”, Figures 24). Additionally, these sequences were acquired with a different modality, demonstrating that QuantEv allows us to compare image sequences with different cell shapes and acquired with different modalities. This study demonstrates that QuantEv is suited to compare images coming from different databases and laboratories, acquired recently or several years ago.
5) The claim that the proposed framework is "generic and nonparametric" seems too strong. In the paper only a few very specific applications are investigated. And many of the underlying components of the framework are not exactly nonparametric. For example, the kernels involved in the weighted density estimation have parameters, and the distance measures depend on the number of bins. This claim should be toned down.
As described in the section “Weighted density estimation”, the bandwidths of the Gaussian kernels are estimated with the Silverman’s rule of thumb and the concentrations of the von Mises kernels are estimated with the robust rule of thumb proposed by Taylor et al. These usual parameters in nonparametric density estimation are consequently automatically estimated and the user does not have to set them. Meanwhile, as the statistical test used in the QuantEv framework is based on the ranks of a difference between earth mover’s distances or circular earth mover’s distances, the result of the test is not affected by the choice of the number of bins. In the Appendix, we evaluated the influence of binning on the earth mover’s distance (see Appendix 4) to clarify this point. Actually, the user just needs to provide the coordinate system, the coordinate system center and a reference direction. For all these reasons, QuantEv can be considered as “semiparametric framework” for traffic phenotype analysis. We added a full paragraph in the Discussion about this point to clearly state what are the needed inputs by the user (second paragraph).
6) The authors claim that their framework is more sensitive than the Kernel density maps. This asks the question of the discriminatory power of the method. The potential to differentiate distribution patterns depends on the resolution of the input; this should be discussed.
QuantEv is able to significantly detect differences if the phenotypes are positively different. The Kernel Density maps approach is too sensitive as it wrongly detects differences for image sequences showing Rab6 proteins trafficking in cells with the same shape (see Figure 2D). The usual ANOVA test also fails as shown in simulations in Appendix 3. In this experiment, the ANOVA test leads to statistical significance when comparing image sequences with different distributions but also when comparing image sequences with same or mixed distributions when the amount of data becomes large (see Appendix 3). The same experiment demonstrates that QuantEv accurately identifies differences for sequences with different distributions with a small number of image sequences, but does not lead to significant differences for sequences with same or mixed distributions, even with a large number of image sequences, demonstrating a good discriminating power.
7) The meaning of the results of the various analyses is often unclear and very limited in terms of providing understanding of mechanisms for any of the systems studied. Since the paper is written for a general audience, this should be improved.
Maybe we misunderstood the comment. Our intention was not to propose a method for deciphering mechanisms related to Rab6 and Rab11. Actually, these proteins are known to be involved in dedicated molecular complexes and interact with other molecules (e.g. Rab11 interact with actin via Myosin VB et Rab11FIP2). The underlying mechanisms of Rab proteins can be better elucidated if two or more fluorescent markers are used. Probably, a generative model would be helpful to analyze the mechanisms. Instead we propose here a computational framework to compare traffic phenotypes related to cell shape, cytoskeleton organization and spatial organization of organelles.
8) A number of specific comments on the text must be addressed:
Introduction, first paragraph “Automatic methods have the obvious advantage of being quicker and reproducible. However, most computational methods are based on the complex combination of heterogeneous features such as statistical, geometrical, morphological and frequency properties (Peng, 2008), whichmakes difficult to draw de1nitive biological conclusions”: This statement ignores extensive work on generative or mechanistic models, which produce interpretable parameters. Such work includes mechanistic models of dynamics of endocytic vesicles (e.g., http://doi.org/10.1038/nmeth.1237) and cytoskeletal dynamics (http://doi.org/10.1126/science.1100533), and generative models of vesicle distribution (e.g., http://doi.org/10.1371/journal.pcbi.1004614).
Introduction, first paragraph “Additionally, most experimental designs, especially at single cell level, pool together data coming from replicated experiments of a given condition (Schauer et al., 2010; Merouane et al., 2015; Biot et al., 2016), neglecting the biological variability between individual cells.”: Again, this ignores work on generative models that specifically analyzes and captures variation between cells. Past examples include microtubule networks (e.g., http://doi.org/10.1371/journal.pone.0050292), and cell and nuclear shape (e.g., http://dx.doi.org/10.1091/mbc.E15060370). Traditional featurebased methods also frequently analyze heterogeneity within populations (e.g., http://doi.org/10.1371/journal.pone.0102678).
We improved the stateoftheart and we referenced the aforementioned modeling approaches. These methods are generative or dedicated methods to analyze specific dynamics. QuantEv is more generic and enables to analyze the spatial distribution of intracellular events with no prior on dynamics. It can be understood as a statistical tool to detect significance evidence between phenotypes for a large range of applications. We added a paragraph in the Discussion (third paragraph) to guide the user about when to use QuantEv and when to use generative or mechanistic models.
Introduction, third paragraph and subsection “Weighted density estimation”, first paragraph – The use of circular and/or cylindrical coordinate systems for description of object positions within a cell is well established (e.g., http://doi.org/10.1002/cyto.a.20487 and http://doi.org/10.1002/cyto.a.21066) and in these cases rotation angle was more powerfully defined relative to the major axis of each cell rather than being defined by the confinement fields. Alternative approaches to the problem, such as morphing, were not discussed.
We included the aforementioned references as suggested (Appendix 2). Actually, spherical and cylindrical representations have already been considered in many approaches. The added value of QuantEv is mainly to analyze spatial and temporal information in a common and appropriate reference system. Moreover, our aim was no to create an atlas of traffic based on a set of points. Morphing registration is probably the best approach to align cells in a common reference but it is generally based on shape features and contours. In our study, we evaluate several vesicle trafficking and the number of detected events is different in each cell. We normalized the distances to avoid morphing cells of different shapes.
Introduction, third paragraph and subsection “Distance between densities”, first paragraph – There is no discussion of more recent metrics related to Earth Mover's Distance that have been described and used to compare subcellular patterns (http://doi.org/10.1007/s112630120566z).
Thank you, we added this reference to the first paragraph of the subsection “Distance between densities” with the reference to https://doi.org/10.1073/pnas.1319779111.
"The KD approach concludes… Instead, QuantEv selectively identifies…" How do we know which method comes closest to the truth? Can this be verified? Without some control experiment or simulation, how can we conclude that QuantEv is to be preferred over other methods?
We acknowledge that this point was not clear enough in the previous version of the manuscript. In this new version of the manuscript, we emphasized the fact that the KD approach leads to statistically significant results when comparing image sequences with cells of same shape (subsection “Visualizing and quantifying the influence of cell shape on the spatial distribution of Rab6 positive membranes”, Figure 2D), demonstrating the KD approach is too sensitive, which is not the case with QuantEv.
Subsection “Visualizing and quantifying the influence of micropatterns on the spatial distribution of Rab6 positive membranes” – There is no clear peak at the twothirds position in Figure 2D and no evidence of significance or reproducibility is presented.
We acknowledge that the term “peak” might be misleading. We replaced it with maxima to avoid confusion.
"Rab6 trajectories were classified into two categories…" How do we know these trajectories are trustworthy? What kind of control experiment was performed to confirm this? This is especially important since it seems the trajectories were not obtained with the best methods available these days (for example according to http://doi.org/10.1038/nmeth.2808 there seem to be better tracking methods than the method mentioned in the subsection “Event detection and localization”).
"we extracted Rab11 trajectories…" Same as previous comment.
We agree that only relying on one tracking method is risky. We also processed the image sequences with the methods proposed by Sbalzarini and Koutmoutsakos and TrackMate, so we now have a multiple hypothesis tracking method, a combinatorial optimization tracking method and a hybrid approach. We then computed a gated distance as defined in Chenouard et al. Between all the trajectories estimated with the three methods, we selected the trajectories for which this distance was inferior to 2 pixels in at least two methods for the analysis (subsection “Event detection and localization”). These selected tracks were used to draw conclusions and detect evidence about phenotypes.
"On the image sequences considered in the previous section (see Figure 4A), this distance remains stable (see Figure 5A). We analyzed cells treated with Latrunculin A… the ERC location is moving away as the drug is affecting the cell (see Figure 5B)". But the time scales are very different in these two cases (seconds versus minutes). Control experiments would be needed to confirm that the ERC location in nontreated cells remains stable over the same time scale as in the treated cells.
Thank you for this valuable remark. We agree and included new experiments with diskshaped cells without latrunculin A treatment acquired every 30 seconds for 20 minutes. Figure 6 was changed accordingly.
The claim that the presented software tool "is efficient with small and large amounts of data" has not been demonstrated in the paper. Neither dataset sizes nor processing times are mentioned.
In Appendix 3, we show how large amount of data can be a problem when using a regular ANOVA analysis. This is not the case with QuantEv that accurately identifies simulations with different distributions but does not state any statistical difference between sequences with same or mixed distributions, even with a large number of image sequences. We also added the Appendix 5 to report the processing times for the different analyzes proposed in QuantEv.
The claim that "QuantEv is quite flexible since the user can specify any distance…" contradicts the statement that "it is fully automated and nonparametric".
"a reference point… and a reference direction have to be specified by the user…" Same as previous comment.
We refer the reviewers to our answer to the point 5 raised in the main essential revision.
[Editors' note: further revisions were requested prior to acceptance, as described below.]
The manuscript has been improved but there are some remaining issues that need to be addressed before acceptance, as outlined below:
There were significant reservations about using constrained cells, as the use of such cells greatly simplifies the analysis. The authors have now done additional work to add results comparing unconstrained cells and two different types of constrained cells. The results show that QuantEV is able to distinguish among the three groups. However, no perturbation studies (e.g., Latrunculin B) were done with unconstrained cells. Thus the main concern remains about the suitability of QuantEV for use in future studies, the majority of which are expected to be done with unconstrained cells. This is an important point: the method may be able to distinguish changes within constrained cells upon various treatments, but may not be able to distinguish perturbations on the background of significant variation within unconstrained cells. There is no information provided on the variance of the profiles in Figure 2F within the unconstrained population. This is a major concern in the context of the very broad claims made in the manuscript (especially in the Discussion) about the power and generality of QuantEV. To support these broad claims, the authors must provide conclusive evidence that QuantEV can distinguish physiologically relevant changes upon perturbations in unconstrained cells. Since the necessary datasets are undoubtedly available to the authors, no collection of new experimental data is expected to be necessary to address this point.
To demonstrate that QuantEv is able to distinguish changes induced by perturbation studies, we have run a new series of experiments consisting of the acquisition of Rab11 positive membranes in unconstrained cells after Latrunculin A treatment. As unconstrained cells are more spread out and have a less organized cytoskeleton than cells on micropatterns, they show a weaker Latrunculin A resistance. Consequently, unconstrained RPE1 cells shrink faster than constrained cells, leading to a complete detachment from the slide approximately 20 minutes after Latrunculin A injection. This does not happen for constrained RPE1 cells as they are more stably attached to the slide through the micropatterns. In order to compare constrained and unconstrained cells, we only kept image sequences at 10 and 15 minutes after Latrunculin A treatment, removing from the previous version of the manuscript the image sequences acquired at 20 and 25 minutes after Latrunculin A treatment for constrained cells. It has to be noted that the radial distributions observed at 20 and 25 minutes after Latrunculin A treatment for constrained cells are very similar to the radial distributions observed at 15 minutes after treatment. The Latrunculin A effect on the dynamical behavior of Rab11 positive membranes is similar in unconstrained and constrained cells (see Figure 7). Furthermore, Latrunculin A injection induces a shift of the radial distribution of Rab11 positive membranes from the cell periphery to the cell center in unconstrained cells, a phenomenon also observed in constrained cells (see Figure 8A). Finally, the difference between radial distributions of Rab11 positive membranes in the three different conditions (unconstrained, crossbow and diskshaped cells) at Latrunculin A injection time is statistically significant (see Figure 8B) while the differences of the same radial distributions 10 and 15 minutes after Latrunculin A injection are not (see Figure 8B). This emphasizes that Latrunculin A influence on radial distribution is similar for the three conditions. We modified accordingly the subsection entitled “Joint actin disruption and cell shape influence on Rab11 radial distribution”. We believe that the addition of a perturbation study in unconstrained cells supports the power and generality of QuantEv and we thank the reviewers for this valuable suggestion. There is no information provided on the variance of the profiles in Figure 2F within the unconstrained population. This is a major concern in the context of the very broad claims made in the manuscript (especially in the Discussion) about the power and generality of QuantEV. The variance of the profiles are not shown on the graphs as it is difficult to add other curves on such packed graphs. However, the fact that the statistical test applied to the radial distributions is low (pvalue = 7.3x10^{4}) demonstrates that the variance of the profiles has to be small. Actually, the averaged standard deviation per bin for the radial distribution of Rab6 positive membranes in unconstrained cells is equal to 0.0074. For each analysis in the manuscript, a statistical test is performed to both demonstrate differences between distributions and take into account the variability in each dataset.
https://doi.org/10.7554/eLife.32311.041Article and author information
Author details
Funding
FranceBioImaging (ANR10INBS04)
 Charles Kervrann
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
This work was supported by the FranceBioImaging infrastructure (ANR10INBS04). The Cell and Tissue Imaging Facility at Institut Curie is a member of FranceBioImaging. We thank Sabine Bardin for regularly testing mycoplasma infection and for the sample preparation of experiments with Rab6 proteins. We thank JeanYves Tinevez for his help to implement the QuantEv analysis module for TrackMate and the Icy track processor.
Senior and Reviewing Editor
 Anna Akhmanova, Utrecht University, Netherlands
Publication history
 Received: September 26, 2017
 Accepted: June 8, 2018
 Version of Record published: August 9, 2018 (version 1)
Copyright
© 2018, Pécot et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics

 883
 Page views

 118
 Downloads

 0
 Citations
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.