Quantitative imaging of membrane contact sites for sterol transfer between endo-lysosomes and mitochondria in living cells

Mitochondria receive cholesterol from late endosomes and lysosomes (LE/LYSs) or from the plasma membrane for production of oxysterols and steroid hormones. This process depends on the endo-lysosomal sterol transfer protein Niemann Pick C2 (NPC2). Using the intrinsically fluorescent cholesterol analog, cholestatrienol, we directly observe sterol transport to mitochondria in fibroblasts upon treating NPC2 deficient human fibroblasts with NPC2 protein. Soft X-ray tomography reveals the ultrastructure of mitochondria and discloses close contact to endosome-like organelles. Using fluorescence microscopy, we localize endo-lysosomes containing NPC2 relative to mitochondria based on the Euclidian distance transform and use statistical inference to show that about 30% of such LE/LYSs are in contact to mitochondria in human fibroblasts. Using Markov Chain Monte Carlo image simulations, we show that interaction between both organelle types, a defining feature of membrane contact sites (MCSs) can give rise to the observed spatial organelle distribution. We devise a protocol to determine the surface fraction of endo-lysosomes in contact with mitochondria and show that this fraction does not depend on functional NPC1 or NPC2 proteins. Finally, we localize MCSs between LE/LYSs containing NPC2 and mitochondria in time-lapse image sequences and show that they either form transiently or remain stable for tens of seconds. Lasting MCSs between endo-lysosomes containing NPC2 and mitochondria move by slow anomalous sub-diffusion, providing location and time for sterol transport between both organelles. Our quantitative imaging strategy will be of high value for characterizing the dynamics and function of MCSs between various organelles in living cells.


Results
NPC2 in LE/LYSs is necessary for transport of sterol from the PM to mitochondria. Current evidence for NPC2's role in cholesterol transport from LE/LYSs to mitochondria relies on biochemical assays, in which formation of steroid hormone precursors, such as pregnenolone from cholesterol is monitored, a reaction, which takes place in mitochondria 9 . We aim for a direct visualization of sterol transport to mitochondria in living cells, but this is challenging, as the cholesterol content of mitochondria is very low. Also, we want to avoid using tagged cholesterol analogues, which can artifactually accumulate in certain organelles, including mitochondria and lipid droplets 11,12 . Instead, we employed the close cholesterol analogue CTL, which contains only two additional double bonds compared to cholesterol giving CTL a slight fluorescence in the ultraviolet (UV; see Fig. S1A for structures). In human fibroblasts lacking functional NPC2 (NPC2−/− cells), CTL accumulates in LE/LYSs after uptake from the PM 13 . When incubating those cells with NPC2 protein purified from bovine milk, we can rescue the lysosomal sterol storage phenotype 13,14 . Co-staining those cells with MitoTracker, we find close apposition of endo-lysosomes containing CTL and mitochondria and frequently faint staining of mitochondria (Fig. 1A). We would not expect more than faint labeling, as the cholesterol content of mitochondria is much lower than that of LE/LYSs, and CTL is a comparably dim but reliable fluorescent cholesterol probe. Importantly, in NPC2-deficient cells not being treated with NPC2 protein, close apposition between CTL containing LE/LYSs and mitochondria was still found, but no CTL in mitochondria could be detected (Fig. 1B). In contrast, CTL was observed in fibroblasts from healthy subjects and from patients lacking functional NPC1 protein ( Fig. S1B and C). Together, these results support earlier findings that NPC2 but not NPC1 is needed for cholesterol transport from endo-lysosomes to mitochondria 9,15 . Mitochondrial ultrastructure is not altered in NPC disease fibroblasts. Mitochondrial morphology is tightly regulated by controlled fusion and fission processes, and this balance is often altered under starvation and certain disease conditions 16 . To assess, whether the NPC mutations change the length and morphology of mitochondria, we quantified these parameters from fluorescence images using the Mitochondria Network Analysis (MiNA) toolset 17 . We found that the length of mitochondria and their extent of branching was higher in fibroblasts from NPC patients compared to cells from healthy control subjects (Fig. S2). This points to a more fused state of mitochondria in NPC disease cells, an indication of oxidative and other metabolic stress conditions. Mitochondria of cells from NPC2 patients were slightly more elongated and branched compared to those of NPC1 patients but treating the former fibroblasts with NPC2 protein did not reverse this phenotype (Fig. S2). To assess the ultrastructure of mitochondria, we employed soft X-ray tomography (SXT). SXT is a label-free imaging technique, achieving isotropic resolution of 30-50 nm in three dimensions throughout an entire cellular volume by absorption of X-rays 18 . Using SXT, we observe extensively branched mitochondria in NPC2−/− fibroblasts, in which intramitochondrial cisternae could be clearly discerned ( Fig. 2A and B). No difference in mitochondrial ultrastructure was found between NPC2-treated and untreated NPC2−/− fibroblasts ( Fig. 2B and S3). We observed repeatedly that some mitochondria form close contacts to endosome-like organelles, which become partly enwrapped by the mitochondrial network with inter-organelle distances of less than 50 nm (Fig. 2C and D and Fig. S3A and B). This could resemble MCSs between both organelle types, which might allow for exchange of cholesterol, metabolites, or ions.
A population of endo-lysosomes containing NPC2 are in contact with mitochondria. Since MCSs could provide a pathway for sterol transfer from LE/LYSs to mitochondria, we determined the spatial www.nature.com/scientificreports/ relationship between both organelles on a whole-cell level. For that, NPC2−/− cells were incubated with 100 nM Alexa546-NPC2 for 72 h, and mitochondria were stained with a green MitoTracker dye before imaging on a spinning disk confocal microscope. We observed extensive clustering of LE/LYSs containing fluorescent NPC2 around mitochondria ( Fig. 3A-C). To quantify this observation, we developed an image analysis protocol, in which mitochondria positions were first determined by thresholding and skeletonizing followed by inversion to set selected pixels to zero (black; Fig. 3D). To that image, we apply the Euclidian distance transform (EDT; Fig. 3E), which calculates the distance for each pixel of the image to the nearest black pixel of the binary image (Fig. 3D). In a second step, we determine the positions of all endo-lysosomes containing Alexa546-NPC2 in the corresponding image, for which a particle finding algorithm as used in SPT is applied. Finally, we map the found positions onto the EDT image, thereby determining all distances of labeled LE/LYSs to their nearest mito-  www.nature.com/scientificreports/ chondria (Fig. 3E, yellow crosses). The resulting distribution of inter-organelle distances comprises the pooled distance histogram from six cells containing together 3560 endo-lysosomes (see histogram in upper part of Fig. 3F). This distribution peaks at around 0.15 µm with shoulder peaks at around 1.0 and 2.2 µm, respectively ( Fig. 3F and S4A). To infer the underlying subpopulations, we first fitted the sum of three Gaussian functions to the experimental data of distances between endo-lysosomes and mitochondria (Fig. S4A). This gave a reasonable fit but suffers from the fact, that we do not know a priori, which data point belongs to which Gaussian component. To overcome this problem, we applied a Gaussian mixture model, which iteratively calculates the probability for each data point to be generated by one of the three Gaussian components of the model 19 . This analysis revealed three populations of LE/LYSs containing NPC2; the two larger ones comprise together more than 95% of all endo-lysosome/mitochondria distances, in which the average distances are 1) 0.142 µm (cluster 0) and 2) 0.887 µm (cluster 2, Fig. 3F). The third population is very small and broadly distributed with an average distance of 3.106 µm to the next mitochondrion (cluster 1, Fig. 3F). The closest pool (cluster 0) comprises 72% of all endo-lysosomes containing fluorescent NPC2. Given that the average diameter of NPC2-containing LE/LYSs in this 2D projection is 0.297 µm (Fig. S4B), some of those LE/LYSs being closer than their radius of 0.148 µm are likely in direct contact with mitochondria. Figure 2. Soft X-ray tomography (SXT) reveals the ultrastructure of mitochondria and shows contacts to endosomes. NPC2−/− fibroblasts were treated with 200 nM NPC2 for 48 h, labeled with TopFluor-cholesterol, chased for 2 h and fixed with PFA. The cells were kept in PBS until plunge freezing in liquid ethane followed by imaging of tomograms at the X-ray microscope, spatial registration and 3D reconstruction of tomograms (see Materials and Methods for details). A-A' , right red box and zoom highlight a branched mitochondrion, while lower left box and zoom point to a contact area between a mitochondrion and an endosome-like vesicle. B, sections though another mitochondrion reveals its ultrastructure with clearly discernable cristae. C-D'' , 3D rendering of the reconstructed tomogram shown in D. Different colored arrows in D, D' , D'' point to endosomelike in close contact with mitochondria, the corresponding color arrows point to same structures in the rendering in C. Bar, 1 μm. www.nature.com/scientificreports/ The observed proximity of a portion of endo-lysosomes to mitochondria suggests some form of interaction between both organelle types. To test this notion, we set out image-based Markov Chain Monte Carlo (MCMC) simulations which can model physical interactions within the cell geometry 20 . Endosomes and lysosomes have been shown to be organized in characteristic spatial patterns, which play an important role in their function 21,22 . To model such a distribution, the spatial position of particles simulating endo-lysosomes is dictated by three types of interaction in our model; (1) a 'Lennard Jones Potential' between the particles, (2) an attractive potential to the nucleus modelled by a 'Morse Potential' to account for interaction of LE/LYSs with perinuclear microtubules 20 , and (3) an attractive potential between the endo-lysosome particles and mitochondria defining the 'Mitochondria Potential (M-potential)' (Fig. 4A). The Lennard Jones 6-12 potential acts between the synthetic endolysosomes and ensures their mutual attraction for intermediate distances (1/r 6 -dependent term) and repulsion upon contact (1/r 12 -dependent term). This type of potential energy is often employed in molecular and soft matter simulations 23 . The second term, the Morse potential is mostly used to model chemical bonds but is here employed to describe the attraction of endo-lysosomes towards the perinuclear region, as has been observed in many experiments ( Fig. 4A and see Materials and Methods for details) 20,21,24,25 . The effect of adjacent mitochondria, to which simulated endo-lysosomes might bind, is defined in the M-potential which is directly inferred from the EDT map of the mitochondria image for a given cell with some scaling factor (Fig. 4A). The EDT map is taken as a stationary potential energy landscape, in which simulated endo-lysosomes move along the Markov Chain given by the Metropolis criterion (Fig. S5A). In other words, particles resembling endo-lysosomes 'move' to regions with lowest values of the EDT map, corresponding to energy minima, i.e., regions of attraction to mitochondria until they reach their thermodynamic equilibrium distribution in the statistical sense. (Fig. 4A, most right panel and Fig. 4B). Accumulation of particles mimicking LE/LYSs around mitochondria is thereby a consequence of attraction defined by the EDT and simulates the binding process, underlying formation of MCSs between mitochondria and endo-lysosomes. The thermal energy of the simulated endo-lysosomes comes from the Boltzmann factor, and sampling is based on the Metropolis criterion 20,23 . By varying the scaling factor for the EDT map, shown in Fig. 4G, we can vary the steepness of the potential energy landscape and thereby control the extent of vesicle clustering around mitochondria. We simulated first 288 endo-lysosomes interacting with a Lennard Jones potential and being strongly attracted to mitochondria. Under those conditions, all vesicles cluster around mitochondria to an extent, which exceeds the experimentally observed clustering (compare Fig. 4C and D, and Fig. 4E and F). This is seen in the simulated images and reflected in the narrower distribution of distances to mitochondria for simulated endo-lysosomes compared to the experimental ones ( Fig. 4I blue and red bars and curves). By including a second population of LE/LYSs which interacts more weakly with mitochondria and including an interaction term to the nucleus for both populations (the Morse Potential), we can simulate a distance distribution between endo-lysosomes and mitochondria, which closely resembles the experimental one of LE/LYSs containing Alexa546-NPC2 in the same cell (see simulation snapshots in Fig. 4H and analysis in Fig. 4I red and green curve). This result confirms our statistical analysis of the experimental data, which suggests that at least two subpopulations of endo-lysosomes containing NPC2 exist in fibroblasts; one which interacts strongly and one which interacts weakly with mitochondria. Note, that the system energy, which is in arbitrary units due to the definition of the energy landscape and the chosen scaling factors, equilibrates already during the first 100 MCMC steps (Fig. S5B). Also, the initial positional order of simulated particles is rapidly lost, ensuring that the start configuration has no impact on the simulation results (Fig. S6). Together, this ensures, that we indeed sample from an equilibrium distribution (only the last 10 simulation snapshots were converted into images and considered for analysis). In summary, by Monte Carlo simulation of physical encounters of endo-lysosomes with mitochondria, we can reconcile the experimentally observed distribution of endo-lysosome/mitochondria distances.
Some of the LE/LYSs appearing as being in direct contact with mitochondria in 2D images could locate in fact above or below a studied mitochondrion. To account for this possibility, we extended the above analysis to three dimensions using two-color confocal z-stacks of double labeled cells (Fig. 5A). In deconvolved confocal sections, endo-lysosomes co-localized extensively with mitochondria along the optical axis (Fig. 5B). Since the same endo-lysosome is found in several planes along the optical axis, a simple peak-finding algorithm would overcount the number of LE/LYSs. We therefore used a 3D object segmentation by which endo-lysosomes containing Alexa546-NPC2 were identified as individual objects ( Fig. 5C; see Materials and Methods). This allowed us to determine all positions of LE/LYSs based on the three coordinates of their centroids. The mean distance of the centroid to the organelle surface is a measure of the average organelle radius, when approximating each LE/LYSs as a sphere. The distribution of these radii of all 3D endo-lysosomes pooled from five cells was rather broad with a mean radius of 0.398 ± 0.165 µm ( Fig. 5D; N = 4615). The cumulative histogram of this mean radius calculated for each of the five cells reveals some cell-to-cell heterogeneity but also shows that between 20-60% of all LE/LYSs have a radius of at least 0.4 µm corresponding to an approximate diameter of 0.8 µm ( Fig. 5D and E). The size heterogeneity of endo-lysosomes containing Alexa546-NPC2 was also reflected in a rather broad distribution of organelle volumes (Fig. 5F).
To find 3D distances to mitochondria, the centroid position of all LE/LYSs pooled from five different cells was mapped onto a 3D version of the EDT calculated from the corresponding z-stack of the MitoTracker images. Note that the EDT has in three dimensions to account for the asymmetric voxel size, for which we used another program than in 2D (see Materials and Methods for details). Also, we kept the physical dimensions of the mitochondria in the 3D analysis, as obtained after thresholding and binarization, i.e. their factual volume. Accordingly, the smallest possible distance between endo-lysosomes and mitochondria in our 3D analysis is 0.0 µm, exactly when centroid positions of LE/LYSs locate to the binarized volume of mitochondria. Applying a Gaussian mixture model to the distance distribution between LE/LYSs and mitochondria, we found again three populations; the first population are those endo-lysosomes, whose centroid positions falls within the volume determined from the binarized mitochondria volume. This population comprises 28.36% of all endo-lysosomes ( Fig. 5  www.nature.com/scientificreports/ 0'). The second population of LE/LYSs locates on average 0.439 ± 0.266 µm from the next mitochondrion and comprises 50.0% of all endo-lysosomes containing fluorescent NPC2 (Fig. 5G, 'cluster 1'). The third population of endo-lysosomes is broadly distributed and on average 1.912 ± 0.879 µm from the next mitochondrion ( Fig. 5G, 'cluster 2'). While the first population of LE/LYSs likely forms MCSs to mitochondria according to our distance definition, a portion of the second population will eventually do that as well. This can be inferred from a comparison of the mean distance of this population to mitochondria with the mean radius of endo-lysosomes (compare Fig. 5D, E and G). We conclude from our combined experimental and computational analysis, that at least 30% of all LE/LYSs containing Alexa546-NPC2 are close enough to mitochondria to establish MCSs to these organelles. The centroid-based analysis presented above does not account for the heterogeneous size and shape of endolysosomes and provides no information about the actual surface area of LE/LYSs in contact with mitochondria. To overcome this limitation, we developed a complementary approach which enables assessing the distance of the surface of all LE/LYSs from mitochondria in 3D. More precisely, we determine surface pixels of all endolysosomes segmented in 3D and map them to the 3D version of the EDT, which provides directly those pixels belonging to LE/LYSs which are in contact with mitochondria ( Fig. 6A-C). The results of this analysis compare favorably with the centroid based 3D analysis of MCSs, as shown in the histograms of centroid distances versus surface pixel distances ( Fig. 6D and E). Thus, endo-lysosomes which form contacts to mitochondria based on the centroid-distance criterion, also have a certain fraction of their surface in direct contact with mitochondria. We applied the surface-based analysis to control fibroblasts, to NPC1-and NPC2-deficient fibroblasts as well as to NPC2-deficient cells additionally treated with unlabeled NPC2 protein ( Fig. 6F-H). Those cells were labeled with Mitotracker Green and additionally with rhodamine-dextran (Rh-dextran), a fluid phase endocytosis marker, which accumulates in LE/LYSs 14 . By visual inspection of the images, we found examples of close contact between endo-lysosomes and mitochondria in all cell types and conditions, as exemplified in Fig. 6G for NPC1-deficient cells. Using the surface-based 3D distance analysis we found that about 10% of the surface area of all dextran containing endo-lysosomes is in direct contact with mitochondria (i.e. distance to mitochondria equals 0 µm), irrespective of the cell type studied (Fig. 6H). These results lead to two important conclusions; 1) MCSs between endo-lysosomes and mitochondria are not limited to NPC2-containing LE/LYSs and 2) NPC1 and NPC2 are not directly involved in contact site formation. In an additional co-localization analysis, we found that about 45% of all LE/LYSs containing Rh-dextran also contain a green fluorescence labeled NPC2, Alexa488-NPC2, and that both vesicle populations cluster at a spatial scale of 0.1 to 0.4 µm (Fig. S7). Colocalization of Alexa488-NPC2 with Alexa647-tagged transferrin (Alexa647-Tf), a marker for early and recycling endosomes was also observed but was much lower than that with Rh-dextran ( Fig. S7G), confirming our previous analysis 26 . We conclude that NPC2 resides primarily in a subpopulation of endo-lysosomes, and that LE/LYSs, even with different cargo, localize near each other in the cell on a spatial scale of less than 1 µm.

Frequency and dynamics of contacts between NPC2-positive LE/LYSs and mitochondria.
To obtain insight into the dynamics of MCSs between endo-lysosomes and mitochondria, we carried out time-lapse experiments, in which we focused on NPC2-deficient cells loaded with red fluorescent Alexa546-NPC2 and co-stained mitochondria with MitoTracker Green. Images were taken every 1 s on a temperature-and focuscontrolled spinning disk confocal microscope. We observed that LE/LYSs can undergo fission while being in contact with mitochondria ( Fig. 7A and Supplemental video 1). Fusion and fission of mitochondria were repeatedly observed (Supplemental video 2 and 3). Some endo-lysosomes almost slide along mitochondria before eventually split into two, suggesting that endosome fission is tightly coupled to mitochondrial dynamics, similar as previously shown for endosome-ER contacts (Supplemental video 4 and 5) 27 . A mitochondrion containing an endo-lysosome at its tip underwent branching, suggesting that both organelles were pulled by motor proteins along the same cytoskeleton track ( Fig. 7B and Supplemental video 6). We also observed frequent transient encounters of spherical mitochondria with LE/LYSs containing NPC2 in a 'kiss-and-run' fashion, especially with punctate-shaped mitochondria ( Fig. 7C and Supplemental video 7 and 8). Such small spherical mitochondria could resemble vesicles, which have been shown to bud from elongated mitochondria under metabolic stress conditions to exchange cargo with endo-lysosomes and peroxisomes 28 . Our time-lapse experiments demonstrate that interactions of endo-lysosomes with mitochondria can be highly dynamic and that fusion and fission of both organelles can be coupled to each other. On the other hand, many LE/LYSs in contact with elongated and branched mitochondria moved very little and kept the contact for tens of seconds (e.g. Supplemental video 9). Such long-term concerted motion is indicating some form of interaction between LE/LYSs and these mitochondria. . Automated quantification of distances between LE/LYSs and mitochondria. NPC2 −/− fibroblasts were incubated with 100 nM Alexa546-NPC2 for 72 h in LPDS medium and labeled for 30 min with MitoTracker Green before imaging at a spinning disk confocal microscope. A-C, single confocal sections showing Alexa546-NPC2 (A), MitoTracker (B) and Overlay (C). The skeletonized version of the MitoTracker image (D) was used to calculate the Euclidian distance transform (EDT), onto which the position of endolysosomes containing Alexa546-NPC2 (from A) was mapped (yellow crosses in E; distance is color-coded in a FIRE LUT and given in µm). The distance distribution from six cells comprising 3560 LE/LYSs containing Alexa546-NPC2 to neighboring mitochondria was analyzed with a Gaussian mixture model to identify subpopulations, resulting in three clusters (F). The x-axis shows the mitochondria-lysosome distance in µm. The y-axis of the main plot shows the three identified cluster populations, while that of the upper histogram shows the number of counts. The largest cluster (# 0) comprises 72.5% of all identified endo-lysosome/distances with a mean value of 142 nm. Bar, 25 µm. www.nature.com/scientificreports/ To quantify the dynamics of MCSs between LE/LYSs and mitochondria in living cells, we devised a method for the reliable and automated quantification of MCSs in time-lapse sequences of living cells (Fig. 8A). In this method, we first applied a color-threshold to merged RGB time-lapse stacks of both channels, thereby identifying overlapping pixels in each frame ( Fig. S8A-C, orange and blue spots). In a second step, these binary regions were classified as objects in a new time-lapse stack allowing for comparing their movement along the underlying mitochondrial structure. Using a time-lapse color stamper, which color codes the position of imaged features from live-cell video sequences, one can visualize the identified MCSs and finds that some of them move back and forth along mitochondria (Fig. S8D, E and F), while others are almost immobile ( Fig. S8G and H). To assess this observation quantitatively, we localized all identified MCSs throughout a 3-min time lapse recording by applying a 2D Gaussian fitting routine and projected their positions from different time points onto one image map, similar as done in single molecule localization microscopy 29 . Overlaying this position map onto a sum projection of the time-lapse sequence of MitoTracker from double-labeled cells, we find that many MCSs stay confined in very distinct clusters (Fig. 8B-D). While mitochondria, of course, contribute to the dynamics of MCSs, as they are part of the segmented structure, their overall dynamics is rather limited since mitochondria can be well-discerned from the sum projection of the time-lapse sequence ( Fig. 8B and C). Analyzing the size distribution of such dynamic clusters reveals that most of them cover an area smaller than 0.1 µm 2 ( Fig. 8E and F). Thus, many MCSs are rather static, and the extent of confinement suggests a strong interaction between the involved endo-lysosomes and mitochondria. Having segmented image regions of overlap between mitochondria and LE/LYSs, we can track their position over time and analyze the trajectories. From all trajectories of at least 10 s length in a given cell, we calculate the mean squared displacement (MSD) and from that the diffusion properties of MCSs. We obtained 388 trajectories of MCSs compared to 1794 trajectories of endo-lysosomes containing Alexa546-NPC2. Similar values were found for three other cells (not shown), suggesting that about 20-25% of all LE/LYSs containing fluorescent NPC2 establish MCSs to mitochondria.
Examples of trajectories and corresponding MSDs are shown in Fig. 9. Many MCSs could be tracked for 30-100 s, showing that the 10-s cut off, we used for tracking is indeed a lower limit. Durations of MCSs between endo-lysosomes and mitochondria for up to 3 min have been observed previously, suggesting that our observation is not unique to NPC2-containing LE/LYSs 30 . By fitting an anomalous diffusion model to the MSDs of such MCSs, we found evidence for slow sub-diffusive motion on short times (right panels in Fig. 9). On longer time scales, some MCSs show a transition to upward curved MSDs (Fig. 9A, C and E, middle panels), while for other trajectories the MSDs reach plateau values for long times ( Fig. 9B and D, middle panels). Upward curved MSDs for long times is characteristic for diffusion coupled to active transport (flow) of organelles along microtubules and other filaments, the latter depending quadratically on time 31 . Such dynamic behavior has been observed for both, LE/LYSs and for mitochondria, since both can be transported actively in cells 21,32 . To determine, whether subdiffusion on short time scales is a general trend for MCSs between LE/LYSs and mitochondria, we analyzed time-averaged MSDs of all 388 identified trajectories for this cell and compared it to the ensemble averaged MSD (Fig. 10A and B). The time-averaged MSDs differ significantly from each other, suggesting that the dynamics of MCSs is rather heterogeneous 33 . The ensemble-averaged MSD could be well-described by an anomalous diffusion model, and the same holds for three additional cells analyzed in the same way (Fig. 10B, S9). The parameters estimated for each cell were averaged (n = 4) giving + /-standard error of the mean: D = 0.0055 ± 0.0005 µm 2 /s, α = 0.4110 ± 0.0505). In all cases, we found a left-skewed distribution of instantaneous velocities with a peak around 0.6 µm/s and maximal values up to 2.3 µm/s (Fig. 10C,D and S9). About half of all tracked MCSs had stepto-step velocities between 0.5 and 1.0 µm/s, as inferred from the cumulative velocity distribution (Fig. 10E). The mean of the ensemble averaged MSD from all four cells could also be well-described by the anomalous diffusion model (Fig. 10F). Together, we conclude that LE/LYSs containing Alexa546-NPC2 can interact with mitochondria for 10 s and longer during which the contact areas undergo slow sub-diffusive motion. Such long-term contact formation could pave the way for cholesterol transfer from LE/LYSs to mitochondria.

Discussion
Mitochondria are not only the central organelles for carrying out oxidative metabolism, but also the site of production of steroid hormones and oxysterols from the precursor cholesterol. Exchange of metabolites, ions and lipids between mitochondria and other organelles can take place via MCSs, but few tools are available to study  www.nature.com/scientificreports/ such MCSs quantitatively in living cells. This contrasts with quantitative analysis of mitochondrial morphology, for which several image analysis toolboxes have been developed in recent years, e.g. 17,[34][35][36][37] . We present two novel approaches to quantify the spatiotemporal relationship between endo-lysosomes and mitochondria from livecell imaging data. Here, we devise a novel analysis strategy based on object segmentation, distance transforms and probabilistic modeling to automatically assess distances between mitochondria and endo-lysosomes. We developed two methods based on this approach; (1) centroid-based and (2) surface-based distance measurement between both organelles types. The first method measures the distance between the centroid positions of all LE/ LYSs to their nearest mitochondria and thereby allows for identifying a subpopulation of endo-lysosomes forming close contact to mitochondria. Using 3D confocal microscopy, we demonstrate that such contacts between both organelle types are not due to 2D projections of the cellular volume. The second method for distance measurements quantifies the distance of each surface pixel of the individual endo-lysosomes to nearby mitochondria in 3D. This method provides information of the surface fraction of all LE/LYSs being in contact with mitochondria. Both methods have been automatized and are available as Macros to ImageJ from the authors. We show that both methods provide complementary information about the extent of contact site formation. To study the dynamics of MCSs between LE/LYSs and mitochondria in living cells, we employ color segmentation combined with object tracking to identify a subpopulation of all MCSs, which remains stable for at least 10 and up to 100 s. Such long-lasting MCSs move slowly and in strongly confined areas, well described by a model for anomalous subdiffusion. Whether the subdiffusive motion of MCSs is due to spatial confinement of bound endo-lysosomes in a constricted space of the cytoplasm and/or the consequence of viscoelastic effects on the mitochondria-vesicle complexes remains to be shown in future analyses. To assess mitochondrial ultrastructure and resolve MCSs at the nanoscale, we employ X-ray microscopy. SXT provides three-dimensional information with isotropic resolution ≤ 50 nm, and since its contrast is based on linear absorption of X-rays, the extent of 'darkness' in reconstructed transmission images is a quantitative measure of X-ray absorption by cellular structures 38,39 . By SXT we can resolve the internal structure of mitochondria, revealing the organization of cristae and detect areas of close apposition of vesicular organelles, resembling LE/ LYSs. The complex mitochondrial ultrastructure, we observe by SXT in human fibroblasts is in accordance with recent X-ray and electron microscopy studies of mitochondria in muscle cells 39,40 . It supports the potential of these methods in delineating mitochondrial structure and morphology in human diseases.
Mitochondria have been proposed to receive cholesterol either from the PM or from LE/LYSs and additionally from lipid droplets 2 . Evidence for such transport pathways comes from biochemical experiments on purified mitochondria 41,42 . Alternatively, measurements on formation of pregnenolone, the first step in steroid hormone synthesis from cholesterol taking place in mitochondria 9,15 or analysis of synthesis of 27-hydroxycholesterol, an oxysterol formed from cholesterol in mitochondria 43 can report about cholesterol transport to mitochondria. Such studies have shown, that NPC2 is needed to deliver cholesterol from endosomes to mitochondria, while NPC1 is not essential 9,15 . Supporting this model, Karten and co-workers showed that NPC2 mutant proteins which can transfer sterol between membranes in vitro but cannot bind to NPC1, can rescue pregnenolone production in NPC2 deficient cells 9,44 . Mitochondria from NPC1-deficient cells were even found to be enriched in cholesterol which was accompanied by various metabolic dysfunctions 44,45 . Sterol export from LE/LYSs containing NPC2 to mitochondria could proceed via the sterol transporter MLN64 (also called StARD3), which locates to the same endo-lysosomes as the NPC proteins, and whose expression is upregulated in NPC1 deficient cells [46][47][48] . To study sterol transport, we used the fluorescent cholesterol analog CTL, which resembles cholesterol closely and accumulates in endo-lysosomes of NPC1 (Fig. S10) and NPC2 deficient cells 13 . By live-cell imaging of CTL, we observe sterol transport to mitochondria in dependence of NPC2 but independent of NPC1, strongly supporting such biochemical studies. Involvement of MLN64 as an acceptor for sterol delivered by NPC2 in the endo-lysosomal membrane is further supported by the observation, that ectopic expression of MLN64 in NPC2deficient cells did not rescue the cholesterol storage phenotype, suggesting that both proteins operate in the same pathway 49 . Based on these and our results, we suggest that the MCSs we observed between LE/LYSs containing NPC2 and mitochondria could serve the function of sterol transfer between both organelles. Since in our labeling procedure, a lot of the CTL resides in the PM, we suggest that CTL traffics from the PM first to LE/LYSs containing NPC2 and from there to mitochondria in dependence of NPC2 function. Supporting such a model we have shown recently, that cycling of sterol between PM and LE/LYSs depends on NPC2 14 . While we believe Figure 5. Centroid-based 3D quantification of distances between LE/LYSs containing NPC2 and mitochondria. NPC2−/− fibroblasts were incubated with 100 nM Alexa546-NPC2 for 72 h in LPDS medium and labeled for 30 min with MitoTracker Green before imaging in three dimensions at a spinning disk confocal microscope. Selected frames (A) and sum projections of xz-views along the optical axis (B) of color overlays of both channels are shown, revealing the close apposition of LE/LYSs containing Alexa546-NPC2 and of mitochondria containing MitoTracker Green (arrows in B). C, endo-lysosomes were segmented and are shown as colorcoded entities together with their centroid positions for the same field as shown in A. D, from the 3D volume of segmented LE/LYSs the mean distance of each centroid position to the surface was calculated and plotted as histogram together with a kernel density estimation of the distance distribution. E, cumulative histograms of the same distances, resembling the mean radii of endo-lysosomes analyzed for each cell separately. F, correlation plot of these mean radii against the measured volume of the LE/LYSs. G, cluster analysis of the pooled histogram of distances between centroids of endo-lysosomes and mitochondria for all five cells, determined using a Gaussian mixture model as in Fig. 3F. The x-axis shows the mitochondria-lysosome distance in µm. The y-axis of the main plot shows the three identified cluster populations, while that of the upper histogram shows the number of counts. See text for further details. Bar, 20 µm in panel A and 3 µm in panel B. www.nature.com/scientificreports/ that this is the most likely scenario, our study does not rule out that other sterol trafficking pathways contribute to cholesterol transport to mitochondria, like transport from droplets or via de novo cholesterol synthesis. Also, a recent study found NPC2 in a mitochondria-associated population of autophagosomes and implicated NPC2 in Toll-like receptor mediated activation of autophagy, in adipocytes 50 . Since autophagosomes are formed from LE/LYSs and control mitochondria turnover 51 , an additional role of MCSs between mitochondria and endolysosomes containing NPC2 could be to control autophagy and mitophagy. While NPC2 is important for sterol transport from LE/LYSs to mitochondria, we found that neither NPC1 nor NPC2 is involved in establishing contact sites between both organelle types (Fig. 6). Thus, other proteins must provide the molecular machinery for organelle tethering, such as specific binding partners for the rab GTPase rab7 30 . Constitutively active rab7 increased the number and duration of contacts, while the rab7 GTPase activating protein TBC1D15 terminated contacts between endo-lysosomes and mitochondria in HeLa cells 30 . Interestingly, NPC1 has been implicated in MCSs between the ER and endo-lysosomes, where rab7-NPC1 interactions via the C18orf8-Mon1-Ccz1 guanidine exchange factor were shown to control lysosomal cholesterol export 52 . In light of these and our results, involvement of NPC1 in membrane tethering seems to be organelle and eventually cell-type specific, and an interesting question would be to identify the rab7 interaction partners for contact formation to mitochondria. Some of the discrepancy concerning the involvement of NPC1 in formation of interorganelle contact sites could also originate in different imaging modalities and employed image quantification protocols. The automated cell-wide image analysis methods, we provide here will allow one to address this and other questions in a user-unbiased manner in future studies.

Scientific Reports
Apart from cholesterol transport, MCSs could play a role in ion homeostasis between endo-lysosomes and mitochondria. For example, delivery of iron from internalized transferrin into mitochondria has been shown to require transient tight contacts to endosomes 53,54 . Similarly, MCSs between LE/LYSs and mitochondria regulate mitochondrial calcium dynamics in a process involving the lysosomal cation channel TRPML1 55 . TRPML1 has been shown to be less active in lysosomal storage disorders including NPC1 deficiency and calcium signaling has been reported to be disturbed in NPC1 deficient cells 56,57 . An attractive hypothesis is, therefore, that NPC2 also plays a role in calcium signaling between endo-lysosomes and mitochondria, either directly or indirectly, via its function as cholesterol carrier protein delivering sterols to NPC1 for export from the LE/LYSs. This hypothesis can be tested in future studies using the tools developed here.

Material and methods
Reagents. Fetal bovine serum (FBS) and DMEM were from GIBCO BRL (Life Technologies, Paisley, Scotland). Other chemicals including human lipoprotein depleted serum (LPDS), EMEM (51417C) and DHE were from SIGMA Chemical (St. Louis, MO). MitoTracker Green, rhodamine-dextran (Rh-dextran, 70 kD) and succinimidyl esters of Alexa488, Alexa546 and Alexa647 were purchased from Invitrogen/Molecular Probes (Inc. USA). NPC2 was purified from bovine milk and conjugated with an succimidyl ester of Alexa488 (emission in green) or of Alexa546 dye (emission in red) as described previously 21 . Transferrin was iron-loaded and subsequently labelled with the Alexa647 dye as described previously 14 . Cholestatrienol (CTL) was synthesized as described 58 .
Cell culture. Human skin fibroblast from healthy male donor (Coriell Institute #GM08680 (referred to as control)) and from NPC2 patients (Coriell Institute #GM18455) were purchased from Coriell Institute for Medical Research (NJ, USA). They were cultured in T25 culture flasks, at 37 o C in an atmosphere of 5% CO 2 in complete DMEM culture medium supplemented with 1% glutamax, 1% Penicillin-Streptomycin and 10% or 20% FBS respectively. Primary fibroblasts from NPC1 patients (Coriell Institute #GM03123) were cultured under same conditions, in 15% FBS in EMEM supplemented with 1% Penicillin-Streptomycin. Cells were checked daily and split with trypsin when a confluency of 90% were reached. Prior to fluorescent microscopy, cells were placed on 35 mm microscope dishes with glass bottom (P35G-1.5-50-C, MatTek) (coated with poly-D-lysine for NPC2 diseased cells) and allowed to settle for 48-72 h in their culture medium. All live cell imaging was carried Figure 6. Surface-based 3D quantification of distances between LE/LYSs and mitochondria and comparison between different fibroblast types. A, NPC2−/− fibroblasts were incubated with 100 nM Alexa546-NPC2 for 72 h in LPDS medium and labeled for 30 min with MitoTracker Green before imaging in three dimensions at a spinning disk confocal microscope. NPC2-containing endo-lysosomes were segmented in 3D, and surface pixels of identified objects are shown for selected frames (A). Euclidian distances calculated in 3D based on the corresponding MitoTracker stacks were defined for the surface pixels of the LE/LYSs using a binary mask and are given as intensity values (B). A zoom of the endosome in the inset in B (next to the white star) is shown in the right half of C, while the corresponding color-overlay of the same endo-lysosome is shown in the left half of C (compare Fig. 5A). 3D Euclidian distances of 0 µm refer to regions of the vesicle surface, were the endo-lysosome is in contact with mitochondria. The centroid-and surface-based distance measurement gave comparable results, as inferred from the respective histograms calculated for the cell shown in A (D and E).   Image analysis of fluorescence microscopy data. All image analysis was done using ImageJ (http:// rsb. info. nih. gov/ ij), and data was further analyzed using self-developed Jupyter notebooks (https:// jupyt er. org/) and R-scripts (https:// www.r-proje ct. org/; available upon request).
Image preprocessing. Multichannel fluorescence images were routinely deconvolved using a theoretical point spread function of the appropriate wavelength, numerical aperture, refractive index and pixel spacing and by applying 30 iterations of the Richardson-Lucy algorithm including background correction in the ImageJ plugin DeconvolutionLab 60 . Fluorescence of CTL was separated from cellular autofluorescence in the UV channel by acquiring bleach stacks of labeled cells, followed by subtracting the last image resembling autofluorescence (frame 50), from the first being comprised of CTL and autofluorescence (frame 1). This procedure was validated by measuring bleaching kinetics relative to unlabeled cells using PixBleach, as described previously 61 .

Quantification of distances between LE/LYSs and mitochondria.
To determine positions of LE/LYSs containing Alexa546-NPC2 a Macro was written which implements the FindMaxima routine in ImageJ excluding edge maxima. Identified peak positions were saved as regions of interest (ROIs). Using another Macro, the corresponding mitochondria images were first segmented using an intensity threshold to get binary images followed by skeletonization. The skeletonized images were inverted, and the Exact Euclidian Distance Transform (3D) plugin of ImageJ was called from the Macro to get an EDT map for each mitochondria image. The calculated distances were converted to µm, and the positions of all endo-lysosomes containing Alexa546-NPC2 were mapped onto the EDT. The resulting distance values of LE/LYSs from six cells (n = 3560 distances) were read-out to an ImageJ Results table from which the distribution of inter-organelle distances was determined for statistical inference of the underlying distribution. This inference consisted of (a) fitting the experimental distribution with the sum of three Gaussian functions (Fig. S4A) and (b) application of a Gaussian mixture model using the scikit learn library for Python (https:// scikit-learn. org/ stable); see Fig. 3). For 3D segmentation of LE/LYSs from confocal z-stacks, we used the 3D ImageJ suite developed by Dr. Thomas Boudier 62 and available at https:// image jdocu. tudor. lu/ plugin/ analy sis/ 3d_ analy sis/ start. This plugin provides statistical measures of organelle sizes but also the surface pixels of identified objects in three dimensions. The centroids and surface pixel maps were calculated for LE/LYSs labeled either with Alexa546-NPC2 or with Rh-dextran. Using an intensity threshold, the surface pixel map was binarized in 32-bit format, background pixels were set to NaN, and this binary mask was applied to the 3D-EDT thereby allocating 3D distances to neighboring mitochondria as intensity values to the surface pixels. All pixel intensities were measured in an automatized manner in an ImageJ Macro script and further analyzed in a Jupyter notebook. www.nature.com/scientificreports/ Segmentation and tracking of MCSs. Corresponding image stacks of fibroblasts labeled with Alexa546-NPC2 and MitoTracker Green were overlayed in an RGB color stack. Color-thresholding implemented in ImageJ was applied followed by conversion to a binary mask of segmented MCSs and blurring with a Gaussian filter with one-pixel width. Using the ImageJ plugin ThunderStorm developed for localization microscopy 63 , the coordinates of all identified MCSs were determined and mapped onto a time projection (i.e. average shifted histogram using a scaling factor of 5). The identified dynamic clusters were slightly blurred with a Gaussian filter to merge sub-resolution clusters, and their sizes were determined using binarization and particle counting in ImageJ (see Fig. 8D and E). Tracking of MCSs was carried out either with SpatTrack, a MatLab-based program developed by us 21 , or using the Python library Trackpy developed by D. Allan and co-workers 64 . Both programs implement the Crocker and Grier algorithm for particle tracking 65 , but Trackpy is Python-based and allows additionally for correction of drift which can take place during image acquisition. Time-and ensemble MSD as well as velocity distributions were calculated using Trackpy in self-developed Jupyter notebooks. The MSD was fitted to an anomalous diffusion model of the form 26 : Here, D α is the (anomalous) diffusion constant, and α is the anomaly parameter. For α < 1, the model describes subdiffusion, for α = 1, normal (free) diffusion and for α > 1, superdiffusion.
Analysis of mitochondrial length and morphology. The free available Mitochondrial Network Analysis (MiNA) toolset for Fiji was used to analyze the length and branching of mitochondria in single cells, of deconvolved images 17 .
Co-localization analysis of Rh-dextran and and Alexa647-Tf with Alexa488-NPC2. Deconvolved widefield images were imported into SpatTrack 21 , and particle-detection based colocalization analysis was carried out using a pixel size of 0.233 µm and a suitable tolerance for particle displacement between image acquisition. Intensity-thresholding of the corresponding DHE image allowed for identifying cell borders upon binarization. In addition to calculating the fraction of co-localized endo-lysosomes in each channel, SpatTrack allows for calculating the nearest neighbor distance between both markers. It also calculates the radial distribution function (RDF) within the cell geometry given by the mask image to identify cell borders and carries out a simulation of randomly placed vesicle populations for each channel to normalize the RDF 21 . The RDF gives thereby the spatial scale at which different endo-lysosome populations coincide relative to a random particle distribution (Fig. S7) 21 .
Image simulations. MCMC simulations of endo-lysosome distributions relative to mitochondria in the cell geometry. The MCMC simulation was implemented as Macro to ImageJ. Two populations of vesicles were simulated, interacting by a Lennard Jones potential and additionally by a Morse-type potential to mimic the impact of the nucleus as recently described 20 , and illustrated in Fig. S5.
An additional contribution to the energy function came from a scaled version of the EDT-map. The strength of the interaction with mitochondria is modeled with a scaling factor f = 100 multiplied with the EDT map for the red particles, meaning that the potential energy contribution for the EDT map varies in this case map between 0 kcal/mol exactly at positions of mitochondria (i.e., 0 Å) to 1000 kcal/mol at the largest distance of 10 Å. Additionally an equally sized population of particles with weaker interaction to mitochondria (f = 10 or f = 0 for the green particles) was included. Particles were randomly displaced with distances drawn from a Gaussian distribution, which was implemented using the Box-Muller procedure as ImageJ Macro, as described 66 . Moves were accepted/rejected based on the Metropolis criterion. Soft X-ray tomography. Cell preparation. R 2/2 grids (QUANTI-FOIL, 100 Holy Carbon Films, Grids: HZB-2 Au) were tapped to objective glasses, that had been cut into appropriate pieces to fit the bottom of 12 well plates and autoclaved. Before seeding NPC2 diseased fibroblasts, the grids were coated with poly-D-lysine. The cells were allowed to grow in their culture medium to reach a confluency of approximate 70%, before they were treated with 100 nM NPC2 protein for 48 h in LPDS medium. Prior to fixation with 4% paraformaldehyde at room temperature, the cells were pulse labeled with TopFluor cholesterol (TF-chol) from a methylβ-cyclodextrin complex (TF-chol/MCD) for 3 min and chased for 2 h in M1 media at 37 °C. The signal from TF-chol were used to localize cells before acquiring tomograms. The cells were kept in PBS until cryo-plunge freezing with liquid ethane and subsequently storage in liquid nitrogen. Before plunge freezing, a small volume of 270 nm gold beads was added to the grids to serve as fiducial markers for subsequently tomographic alignment and reconstruction.
Transmission X-ray microscopy. SXT was performed at beamline U41-PGM1-XM at the electron storage ring BESSY II operated by Helmholtz-Zentrum Berlin. During imaging, the plunge frozen cells on the grids were kept at liquid nitrogen temperature. The cells were imaged over a tilt range of 120-125° with 1° tilt steps on a fullfield transmission X-ray microscope, with an X-ray photon energy of 510 eV and a 25 nm zone plate. The image pixel size was 9.8 nm. A light microscope incorporated in the X-ray microscope, with a Zeiss LD EC Epiplan Neofluar 100 × NA 75 DIC was used to collect the corresponding fluorescent signal 67 .
Image processing and analysis. The freely available software B-Soft and Tomo3D were used to align and reconstruct the SXT data, respectively 68,69 . The segmentation and subsequent rendering were performed in SuRVoS 70 . www.nature.com/scientificreports/