Letter | Published:

Applying systems-level spectral imaging and analysis to reveal the organelle interactome

Nature volume 546, pages 162167 (01 June 2017) | Download Citation


The organization of the eukaryotic cell into discrete membrane-bound organelles allows for the separation of incompatible biochemical processes, but the activities of these organelles must be coordinated. For example, lipid metabolism is distributed between the endoplasmic reticulum for lipid synthesis, lipid droplets for storage and transport, mitochondria and peroxisomes for β-oxidation, and lysosomes for lipid hydrolysis and recycling1,2,3,4,5. It is increasingly recognized that organelle contacts have a vital role in diverse cellular functions5,6,7,8. However, the spatial and temporal organization of organelles within the cell remains poorly characterized, as fluorescence imaging approaches are limited in the number of different labels that can be distinguished in a single image9. Here we present a systems-level analysis of the organelle interactome using a multispectral image acquisition method that overcomes the challenge of spectral overlap in the fluorescent protein palette. We used confocal and lattice light sheet10 instrumentation and an imaging informatics pipeline of five steps to achieve mapping of organelle numbers, volumes, speeds, positions and dynamic inter-organelle contacts in live cells from a monkey fibroblast cell line. We describe the frequency and locality of two-, three-, four- and five-way interactions among six different membrane-bound organelles (endoplasmic reticulum, Golgi, lysosome, peroxisome, mitochondria and lipid droplet) and show how these relationships change over time. We demonstrate that each organelle has a characteristic distribution and dispersion pattern in three-dimensional space and that there is a reproducible pattern of contacts among the six organelles, that is affected by microtubule and cell nutrient status. These live-cell confocal and lattice light sheet spectral imaging approaches are applicable to any cell system expressing multiple fluorescent probes, whether in normal conditions or when cells are exposed to disturbances such as drugs, pathogens or stress. This methodology thus offers a powerful descriptive tool and can be used to develop hypotheses about cellular organization and dynamics.


To explore the spatiotemporal coordination among organelles, COS-7 cells were transfected to express fluorescent proteins targeted to lysosomes, mitochondria, endoplasmic reticulum (ER), peroxisomes, and the Golgi, and incubated with a dye to label lipid droplets (LDs) (Extended Data Fig. 1a). Images were acquired using a laser scanning confocal microscope equipped with a spectral detector (Extended Data Fig. 1b), and after applying a linear unmixing algorithm (Extended Data Fig. 1c–f), the six fluorophores could be separated into distinct compartments (Fig. 1a). Time-lapse images of single z-planes acquired every 5 s revealed the dynamics of all six labelled organelles within single cells (Supplementary Video 1). These images were then processed using an imaging informatics pipeline for quantitative analysis of inter-organelle contacts (Extended Data Fig. 1g).

Figure 1: Live-cell, six-colour confocal microscopy to characterize the organelle interactome.
Figure 1

a, Micrographs of a COS-7 cell expressing fluorescent fusion proteins LAMP1–CFP, Mito–EGFP, ss-YFP–KDEL, mOrange2–SKL, and mApple–SiT, and labelled with BODIPY 665/676. Micrographs are representative of 10 cells captured. b, LDs (outlined in white) were tracked, and their interorganelle contacts mapped. A blue line indicates that the LD was successfully tracked, and coloured lines indicate that the LD was within 1 pixel (97 nm) of the indicated organelle at the specified time point. Numbers on the micrographs represent time (s). For more examples, see Extended Data Fig. 2b. c, Top, matrix representation of the organelle interactome. The absolute numbers of organelle contacts in a single cell at a single time point are displayed as a graphical half matrix. Each row in the matrix represents the number of organelle contacts with each target organelle (columns), and is colour-coded from 0 to 600. Bottom, organelle interactome over time. Each half matrix represents the organelle interactome in a single cell at a specific time point. d, Network representation of the organelle interactome in all ten cells. All nodes (organelles) are connected and the length of the edges connecting two nodes represents the inverse of the number of contacts between those two organelles. Mito, mitochondria; perox, peroxisomes; lyso, lysosomes. Scale bars, 10 μm (a), 5 μm (b).

We tracked individual LDs and mapped their contacts with other organelles over time. Maps of three different LDs revealed near continuous contacts with ER, the major site of lipid synthesis, but transient/heterogeneous contacts with all other organelles (Fig. 1b, Supplementary Video 2). Histograms of LD–organelle contact duration revealed a higher fraction of long-lived LD–ER contacts (Extended Data Fig. 2a). Maps of LDs in a single cell further revealed that LDs were frequently associated with multiple organelles at the same time, and that some LDs associated with all five other organelles over the 300-s imaging period (Extended Data Fig. 2b), suggesting promiscuous exchange of lipids between LDs and other compartments.

We next characterized the number and pattern of contacts between all potential organelle pairs (that is, 15 pairs, Fig. 1c). A matrix analysis of organelle interactions in a single cell (Fig. 1c, top panel) revealed that different organelle pairs showed different frequencies of contacts. Despite the transient nature of individual organelle contacts (Fig. 1b), the overall pattern of organelle interactions was stable over five minutes (Fig. 1c, bottom panel, Extended Data Fig. 3a). This pattern of organelle contacts was consistent across ten cells (Extended Data Fig. 3a, b). These results suggest that a conserved organelle interactome co-exists with a highly dynamic pattern of individual organelle contacts, with the ER acting as the central node in the organelle interactome network (Fig. 1d).

We quantified the fraction of globular organelles (LDs, peroxisomes or lysosomes) that made contacts with any of the other labelled organelles (Fig. 2a). Each of the globular organelles had a characteristic interaction repertoire. For example, 85% of LDs made contact with the ER. The second most common interaction partners for LDs were the mitochondria (21%) and Golgi (15%), whereas 10% of LDs contacted lysosomes or peroxisomes. We validated our six-colour imaging method to ensure that cell labelling and image acquisition parameters did not perturb organelle contacts (Extended Data Fig. 4a, b), and that our measurements were robust to errors in defining the edges of organelles (Extended Data Fig. 4c, d).

Figure 2: The organelle interactome depends on an intact microtubule cytoskeleton.
Figure 2

a, Box whisker plots showing the median fraction of LDs, peroxisomes or lysosomes contacting each of the other labelled compartments in control (− nocodazole) or nocodazole-treated (+ nocodazole) COS-7 cells. Although the total number of contacts between two populations of organelles was, by definition, the same, the fractions of each of the populations in contact were not always symmetric because some individual organelles made simultaneous contact with two or more organelles of the same type. b, Heat map comparison of control and nocodazole-treated cells to computer models of cells with random organelle distributions. The mean values for each interaction were calculated, then the mean value for random associations was subtracted, to give the frequency of associations above random for each binary organelle interaction in the absence or presence of nocodazole. n = 11 (nocodazole-treated) or n = 10 (untreated) cells from two experiments (a, b). *0.05 > P < 0.01, **P < 0.01 (unpaired two-tailed t-test). c, Plot of observed LD ternary contact frequency minus expected frequency, assuming the probabilities of all contacts are independent of each other. n = 10 cells. Error bars represent standard error of the mean (s.e.m.).

Given the primary role of microtubules in organizing organelles and inter-organelle contacts11,12,13,14 (Extended Data Fig. 5a), we next examined how the observed pattern of organelle interactions was affected by microtubule disruption by nocodazole treatment (Extended Data Fig. 5b). Except for the fraction of lysosomes contacting Golgi, which increased after nocodazole treatment, the fractions of most organelles making contacts with other organelles decreased after nocodazole treatment and were more similar to those observed in computer models with randomly placed organelles (Fig. 2a, b, Extended Data Fig. 5c and Supplementary Video 3). This indicated that microtubules have a vital role in establishing and maintaining multiple types of inter-organelle contacts. We confirmed these results using a pixel-based co-localization analysis15 (Extended Data Fig. 5d, e).

We tested the effect of starvation or excess fatty acids on LD–organelle contacts (Extended Data Fig. 6). In response to starvation, the fraction of LDs contacting mitochondria increased, consistent with the role of these contacts in transferring fatty acids from LDs to mitochondria for β-oxidation11,16. LD contacts also increased with lysosomes. This could aid the restoration of starvation-depleted fatty acid reserves within LDs, as lysosomes release fatty acids derived from autophagy of cell membranes16. In response to excess oleic acid, the fraction of LD–peroxisome contacts decreased, whereas that of LD–lysosome contacts increased. These changes may reflect increased lysosomal digestion of LDs under excess oleic acid conditions17.

We next measured the frequency of contacts among LDs and two other different organelles under complete media conditions. We compared the frequency of these tripartite contacts to that expected if all organelle contacts are independent of each other (Fig. 2c). We found that three regimes existed: some tripartite interactions occurred at the expected frequency, some occurred less than expected, and others occurred more frequently. The observed higher-than-expected frequency of LD contacts with ER and lysosomes, and of LD contacts with Golgi and lysosomes, could reflect the known coordination of lipid trafficking among these organelles1.

To acquire three-dimensional images of organelles in live cells with high spatial and temporal resolution, we next developed a lattice light sheet (LLS) implementation of multispectral imaging using an excitation-based linear unmixing approach (Extended Data Fig. 7a–c). This resulted in 3D images (Fig. 3a, Extended Data Fig. 7d, e) and 4D videos (Supplementary Video 4) in which six organelles were distinguished. The mean number, mean volume, total volume per cell of the organelles, and total cell volume in these images, as well as the speed of globular organelles from confocal images, are reported in Extended Data Table 1. These measurements revealed that the ER occupies approximately 37 times the volume of the Golgi and 9 times the volume of the mitochondria. The numbers of LDs, peroxisomes and lysosomes each ranged from around 90 to 190 per cell, and the maximum speed of movement of lysosomes was twice that of LDs and peroxisomes.

Figure 3: Live-cell, six-colour 4D LLS microscopy to characterize organelle distribution in space and time.
Figure 3

a, Maximum intensity projections of a COS-7 cell expressing CFP–SKL, Mito–EGFP, ss-YFP–KDEL, and mApple–SiT, and labelled with Texas Red dextran and BODIPY 665/676. b, zx images of segmented LLS images. c, d, Distributions of organelles in the axial (c) and lateral (d) dimensions of a single COS-7 cell, representative of 10 cells captured. e, Organelle dispersion in the cell over time. Voxels are colour-coded according to the time that an organelle last occupied that voxel. Shown are 2D projections where only the outer shell of the volume is visible. Cell was masked using the dispersion of the ER as a proxy for the cell boundary. Dashed white lines represent the 2D projected outline of the cell generated from the mask. See Supplementary Video 5 for volume exploration. f, Dispersion analysis. The summed fractional cytoplasmic volume (excluding the nucleus) occupied by each organelle is plotted as a function of time. Scale bars, 10 μm (a, e), 5 μm (b). Micrographs are representative of 10 cells.

COS-7 cells displayed characteristic organelle distribution patterns in both the lateral and axial dimensions (Fig. 3b–d). As shown for one cell, in the lateral dimensions, the ER had the widest distribution and Golgi the narrowest (Fig. 3d). In the axial dimension, Golgi displayed a narrow distribution in the centre of the cell, whereas ER, LDs, mitochondria and peroxisomes were localized throughout the cell, with concentrations near both the dorsal and ventral surfaces of the plasma membrane (Fig. 3c). These patterns were qualitatively consistent across the ten cells observed.

To map out total cytoplasmic volume explored by each organelle over time, we made time-dependent volume renderings of our 3D images (Fig. 3e, Supplementary Video 5). At time 0, the ER occupied just over 35% of the cell volume, excluding the nucleus, but quickly explored over 97% of it within 15 min, whereas lysosomes occupied a small fraction of the cell volume and explored just over 15% in the same time period (Fig. 3f). The Golgi remained largely peri-nuclear, but traced through a considerable volume. The ER exploration of virtually the entire cytoplasmic volume may explain its ability to rapidly sense and respond to overall changing cellular needs5.

We observed complex organelle morphologies, such as LDs contacting mitochondria, ER or Golgi (Fig. 4a). These morphologies were especially notable when all six channels were viewed at once (Fig. 4b). Quantification of organelle interactions in LLS images yielded comparable results to those obtained by single z-plane confocal microscopy (Extended Data Fig. 8a), with the ER showing the highest contact rate. The preponderance of organelle–organelle contacts occurred in the juxtanuclear region (Fig. 4c). The exception to this was the fraction of the population of LDs, peroxisomes or lysosomes contacting ER, which was similar in the juxtanuclear region to elsewhere in the cell (Fig. 4c). As with our confocal data, we ensured that image-acquisition parameters did not perturb organelle contacts and that our measurements were robust to errors in segmentation (Extended Data Fig. 8b–d). We also compared object-based and pixel-based analysis methods, again obtaining complementary results (Extended Data Fig. 9).

Figure 4: LLS analysis of the organelle interactome in 3D.
Figure 4

a, Examples of LD, peroxisome and lysosome interorganelle contacts in segmented LLS images. For clarity, only two channels are shown. b, Four examples of complex interorganelle contacts and organization in segmented LLS images. The ER (transparent yellow) is shown in the right panels only. c, Box whisker plots showing the median fraction of LDs, peroxisomes, or lysosomes contacting each of the other labelled compartments in the juxtanuclear or peripheral regions of the cell. n = 10 cells. **P < 0.01 (paired two-tailed t-test). d, Fields of view from volume rendered images of mitochondria (magenta) and sites of mitochondrial contact with five other organelles (green) in LLS images at discrete time points. e, Fields of view from volume-rendered images of mitochondria (grey) and sites of mitochondrial contact with all five other organelles. f, Percentage of segmented mitochondria voxels that contact other organelles over time in the cell shown in (d, e). g, Fields of view from volume-rendered images of ERMCSs (magenta) and sites of contact with four other organelles (green). h, Fields of view from volume rendered images of ERMCSs (grey) and sites of contact with all five other organelles. i, Percentage of ERMCS voxels that contact other organelles over time in the cell shown in (g, h). Scale bars, 2 μm (a, b), 5 μm (d, e, g, h). Micrographs are representative of 10 cells captured.

We next visualized individual organelle contacts with mitochondria in three dimensions over time (Fig. 4d and Supplementary Video 6) and as a group (Fig. 4e and Supplementary Video 7). This revealed that multiple organelle types simultaneously interact with mitochondria, and the ER is the most prominent interacting partner, followed by Golgi and LDs (Fig. 4f). We then treated the extensive interaction sites between ER and mitochondria (that is, ER mitochondria contact sites, ERMCSs) as a ‘metaorganelle’, visualizing its contacts with other organelles (Fig. 4g, h and Supplementary Video 8). The ERMCS metaorganelle made most contact with Golgi, followed by peroxisomes and LDs (Fig. 4i). ERMCS association with Golgi could be required for efficient cholesterol transport between ER, mitochondria and Golgi18.

Our visualization and quantification of dynamic contacts between six different organelles will allow targeted research into the molecular mechanisms that guide these relationships. We anticipate the use of live-cell multispectral imaging in investigating organelle organization and interactions in cells exposed to drugs, pathogens, and other stressors, as well as during cell migration, division and differentiation. This approach should also be useful for identifying proteins that mediate or regulate contact site formation (for example, tethers)19. Developments in making brighter and more photostable fluorescent proteins, as well as improvements in genome editing and computational tools for automated image analysis, should enable multispectral imaging approaches to discriminate even more than six molecular species over time in single cells.


Data reporting

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.

Cell culture and transfection

African green monkey, Cercopithecus aethiops, COS-7 cells were obtained from the American Type Culture Collection (CRL-1651). A PCR-based assay to detect species specific variants of the cytochrome c oxidase I gene (COI analysis) was used to rule out inter-species contamination, and cells were tested for mycoplasma contamination by the American Type Culture Collection. COS-7 cells were maintained in Dulbecco’s modified eagle’s medium (DMEM, Life Technologies) supplemented with 10% fetal bovine serum (Corning), at 37 °C and 5% CO2. For confocal imaging, cells were cultured in 10 μg ml−1 fibronectin (Millipore)-coated LabTek 8-well Chambered Cover glass dishes (ThermoScientific), and transfected to express the following fluorescent proteins, either alone or in combination: lysosomes were labelled with lysosomal-associated membrane protein 1 fused to CFP (LAMP1–CFP, G. Patterson, constructed as described for PA-GFP-lgp120 (ref. 20)); mitochondria were labelled with the mitochondrial-targeting sequence of cytochrome c oxidase subunit VII fused to EGFP (Mito-EGFP, A. Rambold); the ER was labelled with a fusion of the bovine prolactin signal sequence with YFP and a KDEL ER retention sequence (ss-YFP–KDEL, E. Snapp, constructed as described for ss-GFP–KDEL21); peroxisomes were labelled with mOrange2 fused to the peroxisome-targeting sequence SKL (mOrange2–SKL); and the Golgi was labelled with mApple fused to the 15 N-terminal amino acids of sialylytransferase (mApple-SiT). For experiments shown in Extended Data Fig. 5a, b, mApple-SiT was replaced with mApple fused to the 10 C-terminal amino acids of microtubule-associated protein 4 (mApple-MAP4-C10). Transfection was performed using Lipofectamine 2000 (Life Technologies), according to the manufacturer’s instructions. Briefly, cells were transfected in Opti-MEM (Life Technologies) with 50–200 ng of each plasmid and 0.5 μl Lipofectamine 2000 per well. Cells were incubated in the transfection mixture for 6 h, and then the medium was replaced with standard culture medium. BODIPY 665/676 (50 ng ml−1, Life Technologies) was added to the medium for 16 h before imaging and was present during imaging. Where indicated, cells were treated with 5 μM nocodazole (Sigma-Aldrich) on ice for 2 min, warmed to 37 °C, and imaged 1 h after the addition of nocodazole. For nutritional perturbation experiments, cells were incubated in Hanks’ balanced salt solution (HBSS, Gibco 14025) or complete medium supplemented with 300 μM oleic acid (oleic acid-albumin from bovine serum, Sigma-Aldrich) for 18 h before imaging. For LLS imaging, cells were cultured on 10 μg ml−1 fibronectin-coated 5-mm coverslips, and transfected with CFP–SKL (P. Kim, constructed as described for RFP–SKL22), Mito–GFP, ss-YFP–KDEL, and mApple-SiT as described above. BODIPY 665/676 (50 ng ml−1, Life Technologies) was added to the medium for 16 h before imaging and was present during imaging. Texas Red dextran (10,000 MW, 10 μg ml−1, Life Technologies) was added to the medium for 14 h, washed 3 times, and chased for 2 h in standard culture medium (including BODIPY 665/676) before imaging.

Image acquisition and unmixing

Laser scanning confocalImages were acquired on a Zeiss 780 laser confocal scanning microscope equipped with a 32-channel multi-anode spectral detector (Carl Zeiss) using a 63×/1.4 NA objective lens, at 37 °C and 5% CO2. All fluorophores were excited simultaneously using 458, 514 and 594 nm lasers and a 458/514/594 nm main beam splitter, and images were collected onto a linear array of 32 photomultiplier tube (PMT) elements in lambda mode at 9.7 nm bins from 468 to 687 nm. Images were acquired with 5.0 s intervals between frames for 60 frames. We acquired images of cells labelled with single-fluorophore reporters (each of the six reporters used to label multi-labelled cells) using the same image acquisition settings as for multi-labelled cells (same laser power and master gain settings). Spectra were defined using images from singly labelled cells, and images from multi-labelled cells were subjected to linear unmixing using Zen software (Carl Zeiss). To estimate the error in assigning a pixel to the wrong organelle after unmixing, cells expressing single fluorescent proteins or incubated with BODIPY only were subjected to image segmentation, then the mean pixel intensity for every pixel in the foreground was computed for each of the six fluorophores (Extended Data Fig. 1f). In all cases, the correct fluorophore intensity was at least an order of magnitude higher than any incorrectly assigned fluorophore. In general, with well-acquired reference spectra, the error in unmixing is assumed to be a result of noise in the image. For time-lapse imaging, a single frame was first acquired and subjected to linear unmixing, in order to ensure that the cell was expressing all five fluorescent proteins at roughly equal levels. As a general rule of thumb, for successful linear unmixing the intensity of the fluorophores in the sample should not differ by more than one order of magnitude. Approximately 5–10% of the transfected cells met this criteria, making them suitable for time-lapse imaging.

Unmixed images were then spatially deconvolved using Huygens software (Scientific Volume Imaging) using theoretical point-spread functions. In principal, spatial deconvolution could have been performed on the images before linear unmixing. Because hyperspectral images generally contain many more wavelength channels before unmixing than they do after unmixing (for example, our confocal data set consisted of 26 spectral channels before unmixing and only six after), it may be computationally more efficient to perform linear unmixing on these data sets first, before applying a highly iterative spatial deconvolution procedure.

Lattice light sheetThe LLS instrument illuminates the specimen with an ultrathin light sheet derived from 2D optical lattices10. The thinness of the sheet leads to high axial resolution while simultaneous illumination of the entire field of view permits imaging at hundreds of planes per second, leading to full 3D live cell imaging of cells with near isotropic resolution. A recent implementation of spectral light sheet imaging used a diffraction grating to project dispersed light onto a camera, allowing fine sampling of emitted light across the visible spectrum, but still required scanning of the sample in two dimensions, y and z, making it relatively slow23. To avoid this drawback, we adapted a LLS instrument for spectral imaging with a multispectral excitation-based approach (Extended Data Fig. 7a). This obviated the need to scan in y, giving fast acquisition rates and allowed imaging at subcellular resolution over time.

Images were acquired on a custom LLS instrument10 equipped with a 25×/1.1 NA imaging objective and digital camera with a scientific complementary metal oxide silicon (sCMOS) sensor (Hamamatsu Orca Flash 4.0) and six solid-state lasers (MPB Communications) emitting at 445, 488, 532, 560, 590 and 642 nm wavelengths. 3D image stacks of 140 planes per cell were acquired every 9.2 s for 100 time points, n = 10 cells. Image stacks were acquired sequentially at each laser wavelength, with emission collected through a series of four interference filters (Semrock), which functioned to block excitation wavelength bands and transmit emission wavelength bands to the CMOS detector. These filters were a notch filter with transmission minima at 405, 488, 561 and 635 nm, a long-pass filter with transmission edge at 442 nm, a notch filter with transmission minimum at 532 nm and a notch filter with transmission minimum at 594 nm. We acquired images of cells labelled with single-fluorophore reporters using the same laser power and exposure time as for multiply labelled cells. To estimate the error in assigning a pixel to the wrong organelle after unmixing, singly labelled cells were subjected to image segmentation, then the mean pixel intensity for every pixel in the foreground was computed for each of the six fluorophores (Extended Data Fig. 7c). Lateral and axial resolution at each of the excitation wavelengths, measured as the FWHM intensities of 100 nm diameter beads, are as follows. 445 nm: 0.294 μm (lateral) × 0.649 μm (axial); 488 nm: 0.312 μm (lateral) × 0.666 μm (axial); 532 nm: 0.375 μm (lateral) × 0.731 μm (axial); 560 nm: 0.359 μm (lateral) × 0.771 μm (axial); 589 nm: 0.370 μm (lateral) × 0.789 μm (axial); and 642 nm: 0.370 μm (lateral) × 0.947 μm (axial).

Linear unmixing was performed using a custom algorithm in Mathematica. Organic fluorophores including fluorescent proteins have characteristic emission and excitation spectra. When excited with different wavelengths of light, different fluorophores emit different numbers of photons on the basis of their probability of absorbing a photon and transitioning to an excited state at that particular wavelength. The recording of differential emission intensities with different wavelengths thus approximates the excitation spectrum of each fluorophore and is revealed to be characteristic for each fluorophore (Extended Data Fig. 7b). Our excitation-based unmixing algorithm uses this excitation information as reference spectra, rather than using emission data. With linear unmixing, the observed spectrum at every pixel in a spectrally resolved digital image and the known spectra of the fluorophores present in the specimen can be treated as a system of linear equations of the general form:where the column vector with elements labelled yi (i {1, 2, …, q}) is the observed pixel spectrum, the matrix with elements mi,j are the known fluorophore spectra, the column vector with elements xj (j {1, 2, …, p}) are the abundances of all of the fluorophores used to label the specimen and the column vector with elements ni is the noise in the observed spectrum, p is the number of fluorophores used in the experiment, and q is the number of spectral channels acquired. Equation (1) can be simplified as:the vector x is solved for with the method of least squares by applying an inverse or pseudo inverse operation to the matrix, M:In practice, a non-negativity constraint is usually imposed upon the solution for x and is not shown in this generalized expression. We implemented the least squares algorithm using the Mathematica optimization function, ‘FindMinimum’ and included a non-negativity constraint. For computational efficiency, the unmixing problem was first solved without the non-negativity constraint. If all answers to x were positive, the solution was kept. If any of the elements of x were negative, the solution was discarded and the problem was solved a second time with the non-negativity constraint imposed. Before unmixing, images were corrected for chromatic aberration with a linear pixel transformation, using values derived from images of ‘tetra-spec’ multi-labelled subresolution fluorescence beads (Life Technologies). After unmixing, images were spatially deconvolved using experimentally derived point spread functions for each excitation wavelength.

Image analysis

Laser scanning confocalUnmixed images were segmented in Mathematica using algorithmic, histogram-based approaches for defining intensity thresholds. The optimal threshold values for all organelles except ER were determined using the Mathematica implementation of the method from ref. 24. ER was segmented using a local threshold determined using the Mathematica function: ‘LocalAdaptiveBinarize[]’. Globular organelles were tracked using the TrackMate plugin (an ImageJ implementation of the linear assignment problem (LAP) tracking algorithm)25. The segmented images and tracking data were then imported into Mathematica for organelle interactome analysis. We wrote a custom-feature-based co-localization analysis program. Because globular organelles were segmented in two different software environments (Mathematica for contact analysis and ImageJ for tracking), the segmented images were first merged in Mathematica before analysing contacts. Organelle contacts in confocal images were defined as an overlap of 3 or more contiguous pixels between segmented features, with the target feature dilated 1 pixel (equivalent to 97 nm). We recognize that this definition may lead to an overestimate of contacts, as the distance between membranes at membrane contact sites studied by electron microscopy has generally been reported in the range of 15–30 nm (ref. 26). However, the maximum distance between membranes that allows for a functional interaction (such as for the transport of lipid molecules) is unclear, and may be even greater than 100 nm owing to the long lengths of organelle–organelle tethering proteins26.

For analysis of the organelle interactome, the first frames of segmented images of cells were analysed for regions of overlap between different organelles. The number of globular organelles that overlap with each of the five other organelles was then computed and compared to models of randomly distributed organelles within images of cells. To construct models, the segmented images of all five organelles not of interest were used as a base image on which the model globular organelles of interest (either peroxisomes, lysosomes or LDs) were distributed using a random number generator for centroid coordinates. Parameters for the model images, for example, number and size of globular organelles of interest, were derived from the measurements of segmented images of multi-labelled cells.

Network diagrams were constructed from the mean number of contacts for each organelle pair from all cells observed for each condition in Mathematica. The length of edges connecting all nodes was calculated as the inverse of the number of contacts between those two organelles—shorter edges means higher number of contacts. The six-dimensional network was then rendered as a diagram in 2D space to minimize the global error in total edge length.

Lattice light sheetUnmixed images were segmented and organelle contacts analysed using Imaris software (Bitplane). Segmentation was performed with the surfaces tool, using smoothing and background subtraction, and manual thresholding. Split touching objects was used for the globular organelles (LDs, peroxisomes, and lysosomes). Objects smaller than 10 voxels were excluded. Distance transformations were performed to calculate the distance between objects, and organelles with a minimum distance of 0 nm between edges were considered to be interacting. For analysis of juxtanuclear versus peripheral organelle interactions, the region of the cell with a thickness greater than approximately 2 μm was considered juxtanuclear, and was masked manually. For analysis of organelle distribution in 3D, the centre of the nucleus was estimated visually.

Automated pixel-based quantification of organelle co-localizationA custom MATLAB program was written to capture co-localizations among pixels in the thresholded images. This approach is useful in measuring interactions among organelles with variable and non-globular morphology that would otherwise be difficult to accurately characterize. Boundary regions for individual cells in each image were identified manually using ER as a marker for the full cell. Each channel image was median-filtered using a 5-pixel support. An adaptive threshold was computed using Otsu’s method24 and the images were binarized. Mathematical morphology27 was used to fill in a two-pixel radius around each object for co-localization measurements with a dilation operator. For each organelle image Io,x we define the set of foreground pixels obtained from this processing as . The co-localization S between two organelle channels is then calculated from the normalized intersection cardinality,S takes on values in the range [0,1]. S achieves a value of one when the organelles are perfectly co-located, that is when all the foreground pixels from the processed channel images intersect each other following the morphological processing, and a value of zero when organelles are not co-localized anywhere. These co-localization measures were computed between each organelle pair in each image frame, creating a six-by-six matrix that is symmetric, with zeros on the diagonal. The co-localization measures from each image frame were combined for each experimental condition to form a distribution of co-localization.

The object-based contact analysis and pixel-based co-localization analyses are not expected to always yield similar results, as seen in Extended Data Fig. 5d. This is because the object based method measures the frequency of objects touching, whereas the pixel-based method measures the statistical similarity between different image channels. Object-based methods provide a reliable quantification of organelle contacts but require an assumption that the image segmentation faithfully identifies the edges of organelles. The pixel-based method used here to validate our object-based analysis does not require this underlying assumption regarding identification of organelle edges, nor does it answer the specific question: where do organelle edges make contacts with other organelle edges. The same pixel-based approach was applied to both 2D confocal and 3D LLS multi-channel time-lapse images. A software tool called LEVER 3D28 originally developed for characterizing neural stem cell interaction with blood vessels15 was used here to visualize the interaction between mitochondria and the other organelles (Fig. 4d, e), or the interaction between ERMCSs and other organelles (Fig. 4g, h).


All statistical analyses (Student’s t-tests and Student’s paired t-tests) were performed in Mathematica. For all box whisker plots, the white line in the centre of each box represents the median value, the upper and lower edges of the boxes represent the 75th and 25th quantile of the data, respectively, and the upper and lower fences represent the 95% confidence level of the distribution. Closed circles represent near outliers defined as points beyond 3/2 times the interquantile range from the edge of the box and open circles denote far outliers defined as points beyond 3 times that range.

To assess the variance between cells across time in the number of organelle contacts (as in Fig. 1c and Extended Data Fig. 3a), we performed an analysis of variance (ANOVA). The 15-element vector of each frame was normalized to a sum of 1 and then averaged across time, and a correlation distance was computed between each image frame-15-element vector and the average vector. The variance between frames, as determined by a correlation distance from the average vector over the 60 image frames, was then evaluated as a measure of within-cell temporal variation of the organelle interactome. The variance between cells was evaluated in the same way, with each cell being represented by their average vector. The results show that the variance between cells is significantly larger than the variance within an individual cell across time (P < 1 × 10−37). To quantify the similarity of organelle contact patterns between cells, we performed a cluster analysis using the correlations between organelle associations shown in Extended Data Fig. 3b. The gap statistic29,30 is a widely used technique for estimating the number of clusters in a data set. The gap statistic compares the average intra-cluster dispersion of the given data to that of uniformly distributed randomly generated data on the same range, and picks the number of clusters as maximizing the improvement of the given data compared to the random data. For the organelle association data in Extended Data Fig. 3a, the gap statistic found a single cluster in the data, with no meaningful differences to separate the organelle associations for the ten cells.

Code availability

All Mathematica code for object-based organelle interactome analysis, as well as instructions for its use and example images, is available for download at http://organelle-interactome.sourceforge.net.

Data availability

The authors declare that all data supporting the findings of this study are available within the paper and the Supplementary Information files.


  1. 1.

    , & Lipid droplet-organelle interactions: emerging roles in lipid metabolism. Curr. Opin. Cell Biol. 35, 91–97 (2015)

  2. 2.

    ., . & No peroxisome is an island—peroxisome contact sites. Biochim. Biophys. Acta (2015)

  3. 3.

    , , & Signals from the lysosome: a control centre for cellular clearance and energy metabolism. Nat. Rev. Mol. Cell Biol. 14, 283–296 (2013)

  4. 4.

    , & Mitochondrial lipid trafficking. Trends Cell Biol. 24, 44–52 (2014)

  5. 5.

    & Structure and function of ER membrane contact sites with other organelles. Nat. Rev. Mol. Cell Biol. 17, 69–82 (2016)

  6. 6.

    , , & ER contact sites define the position and timing of endosome fission. Cell 159, 1027–1041 (2014)

  7. 7.

    et al. ER tubules mark sites of mitochondrial division. Science 334, 358–362 (2011)

  8. 8.

    et al. Arf1/COPI machinery acts directly on lipid droplets and enables their connection to the ER for protein targeting. eLife 3, e01607 (2014)

  9. 9.

    & Optimizing imaging parameters for the separation of multiple labels in a fluorescence image. J. Microsc. 213, 46–62 (2004)

  10. 10.

    et al. Lattice light-sheet microscopy: imaging molecules to embryos at high spatiotemporal resolution. Science 346, 1257998 (2014)

  11. 11.

    et al. AMPK activation promotes lipid droplet dispersion on detyrosinated microtubules to increase mitochondrial fatty acid oxidation. Nat. Commun. 6, 7176 (2015)

  12. 12.

    , , , & ER sliding dynamics and ER-mitochondrial contacts occur on acetylated microtubules. J. Cell Biol. 190, 363–375 (2010)

  13. 13.

    , & Interplay between microtubule dynamics and intracellular organization. Int. J. Biochem. Cell Biol. 44, 266–274 (2012)

  14. 14.

    , , , & Endoplasmic reticulum-endosome contact increases as endosomes traffic and mature. Mol. Biol. Cell 24, 1030–1040 (2013)

  15. 15.

    et al. Visualization and correction of automated segmentation, tracking and lineaging from 5-D stem cell image sequences. BMC Bioinformatics 15, 328 (2014)

  16. 16.

    , & Fatty acid trafficking in starved cells: regulation by lipid droplet lipolysis, autophagy, and mitochondrial fusion dynamics. Dev. Cell 32, 678–692 (2015)

  17. 17.

    et al. Autophagy regulates lipid metabolism. Nature 458, 1131–1135 (2009)

  18. 18.

    Cellular cholesterol trafficking and compartmentalization. Nat. Rev. Mol. Cell Biol. 9, 125–138 (2008)

  19. 19.

    , , & A tether is a tether is a tether: tethering at membrane contact sites. Dev. Cell 39, 395–409 (2016)

  20. 20.

    & A photoactivatable GFP for selective photolabeling of proteins and cells. Science 297, 1873–1877 (2002)

  21. 21.

    , , & Monitoring chaperone engagement of substrates in the endoplasmic reticulum of live cells. Proc. Natl Acad. Sci. USA 103, 6536–6541 (2006)

  22. 22.

    , , & The origin and maintenance of mammalian peroxisomes involves a de novo PEX16-dependent pathway from the ER. J. Cell Biol. 173, 521–532 (2006)

  23. 23.

    , , , & Hyperspectral light sheet microscopy. Nat. Commun. 6, 7990 (2015)

  24. 24.

    A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man Cybern. 9, 62–66 (1979)

  25. 25.

    et al. Robust single-particle tracking in live-cell time-lapse sequences. Nat. Methods 5, 695–702 (2008)

  26. 26.

    & Piecing Together the Patchwork of Contact Sites. Trends Cell Biol. 27, 214–229 (2016)

  27. 27.

    ., . & Digital Image Processing Using MATLAB , 2nd ed. (Gatesmark Publishing, 2009)

  28. 28.

    , , , & LEVER: software tools for segmentation, tracking and lineaging of proliferating cells. Bioinformatics 32, 3530–3531 (2016)

  29. 29.

    , , , & Automatic summarization of changes in biological image sequences using algorithmic information theory. IEEE Trans. Pattern Anal. Mach. Intell. 31, 1386–1403 (2009)

  30. 30.

    , & Estimating the number of clusters in a data set via the gap statistic. J. R. Stat. Soc. Series B Stat. Methodol. 63, 411–423 (2001)

  31. 31.

    , & Spectral imaging: principles and applications. Cytometry A 69, 735–747 (2006)

  32. 32.

    et al. Systems-level analysis of microbial community organization through combinatorial labeling and spectral imaging. Proc. Natl Acad. Sci. USA 108, 4152–4157 (2011)

  33. 33.

    , & Multiplexed spectral imaging of 120 different fluorescent labels. PLoS One 11, e0158495 (2016)

  34. 34.

    , , , & Multispectral labeling technique to map many neighboring axonal projections in the same tissue. Nat. Methods 12, 547–552 (2015)

  35. 35.

    , & A guide to choosing fluorescent proteins. Nat. Methods 2, 905–909 (2005)

  36. 36.

    et al. A far-red fluorescent protein evolved from a cyanobacterial phycobiliprotein. Nat. Methods 13, 763–769 (2016)

  37. 37.

    & (eds) The Molecular Probes HandbookA Guide to Fluorescent Probes and Labeling Technologies (Life Technologies, 2010)

  38. 38.

    , , , & Multi-spectral imaging and linear unmixing add a whole new dimension to laser scanning fluorescence microscopy. Biotechniques 31, 1272–1278 (2001)

  39. 39.

    , & NIH Image to ImageJ: 25 years of image analysis. Nat. Methods 9, 671–675 (2012)

Download references


We are grateful to P. Sengupta and other members of the Lippincott-Schwartz laboratory for helpful discussions. This work was supported by the Intramural Research Program of the National Institutes of Health (A.M.V., S.C., and J.L.-S.) and the Howard Hughes Medical Institute (W.R.L., E.B., and J.L.-S.), by a Postdoctoral Research Associate (PRAT) Fellowship from the National Institute of General Medical Sciences to A.M.V. and by NIH grant R01AG041861 from the National Institute on Aging to E.W. and A.C.

Author information

Author notes

    • Alex M. Valm
    •  & Sarah Cohen

    These authors contributed equally to this work.

    • Michael W. Davidson



  1. Eunice Kennedy Shriver National Institute for Child Health and Human Development, NIH, Bethesda, Maryland 20892, USA

    • Alex M. Valm
    • , Sarah Cohen
    •  & Jennifer Lippincott-Schwartz
  2. Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, Virginia 20147, USA

    • Wesley R. Legant
    • , Eric Betzig
    •  & Jennifer Lippincott-Schwartz
  3. School of Biomedical Engineering, Science and Health Systems, Drexel University, Philadelphia, Pennsylvania 19104, USA

    • Justin Melunis
    •  & Uri Hershberg
  4. Department of Microbiology and Immunology, College of Medicine, Drexel University, Philadelphia, Pennsylvania 19104, USA

    • Uri Hershberg
  5. Department of Electrical and Computer Engineering, College of Engineering, Drexel University, Philadelphia, Pennsylvania 19104, USA

    • Eric Wait
    •  & Andrew R. Cohen
  6. National High Magnetic Field Laboratory and Department of Biological Science, The Florida State University, Tallahassee, Florida 32313, USA

    • Michael W. Davidson


  1. Search for Alex M. Valm in:

  2. Search for Sarah Cohen in:

  3. Search for Wesley R. Legant in:

  4. Search for Justin Melunis in:

  5. Search for Uri Hershberg in:

  6. Search for Eric Wait in:

  7. Search for Andrew R. Cohen in:

  8. Search for Michael W. Davidson in:

  9. Search for Eric Betzig in:

  10. Search for Jennifer Lippincott-Schwartz in:


A.M.V., S.C., and J.L.-S. conceived and designed the study. A.M.V., S.C., and W.R.L performed the experiments. M.W.D. contributed novel reagents. A.M.V., S.C., J.M., E.W., U.H., and A.R.C. analysed the data. A.M.V., S.C., and J.L.-S. wrote the manuscript, with input from all co-authors. J.L.-S. and E.B. supervised the project.

Competing interests

The authors declare no competing financial interests.

Corresponding author

Correspondence to Jennifer Lippincott-Schwartz.

Reviewer Information Nature thanks S.-H. Shim and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Supplementary information


  1. 1.

    Video 1: Point-scanning confocal, 6-colour time-lapse images

    COS-7 cells expressing fusion proteins targeted to the lysosomes (cyan), mitochondria (green), ER (yellow), peroxisomes (red), and Golgi (magenta), and labelled with BODIPY 665/676 to stain LDs (blue) were imaged as described in Fig. 1a. Images were acquired every 5 s, for a total of 60 frames (5 min). In frames 21-40, the ER channel (yellow) is omitted so that the other channels can be seen more clearly. Video plays at a rate of 6 frames per s. Scale bar, 10 µm.

  2. 2.

    Video 2: Time-lapse images of tracked LDs

    Videos of the tracked LDs (outlined in white) shown in Fig. 2b. Confocal images were acquired every 5 s, for a total of 60 frames (5 min). Video plays at a rate of 6 frames per s. Scale bar, 5 µm.

  3. 3.

    Video 3: Time-lapse images of a cell treated with nocodazole.

    A COS-7 cell expressing fusion proteins targeted to the lysosomes (cyan), mitochondria (green), ER (yellow), peroxisomes (red), and Golgi (magenta), and labelled with BODIPY 665/676 to stain LDs (blue), was incubated on ice for 2 min and treated with 5 μM nocodazole for 1 h. After the nocodazole treatment, confocal images were acquired every 5 s, for a total of 60 frames (5 min). In frames 21-40, the ER channel (yellow) is omitted so that the other channels can be seen more clearly. Video plays at a rate of 6 frames per s. Scale bar, 10 µm.

  4. 4.

    Video 4: Lattice light sheet, 6-colour time-lapse images

    Volume rendering of COS-7 cells expressing fusion proteins targeted to the peroxisomes (cyan), mitochondria (green), ER (yellow), and Golgi (red), and labelled with Texas Red dextran (lysosomes, magenta) and BODIPY 665/676 (LDs, blue), imaged as described in Fig. 3a. Image stacks of 140 planes were acquired every 9.2 seconds, for a total of 100 frames (15.3 min). Video plays at a rate of 6 frames per s. Scale bar, 10 µm.

  5. 5.

    Video 5: Organelle dispersion through the cytoplasm over time

    Volume rendering of 6 organelles in a COS-7 cell. Voxels are colour-coded according to the time that they were last occupied by the organelle from blue to red.

  6. 6.

    Video 6: Montage of mitochondria-organelle contacts in time lapse, lattice light sheet images

    Volume rendering of mitochondria (magenta) in COS-7 cells expressing fusion proteins targeting 3 other organelles, and labelled with Texas Red dextran and BODIPY 665/676. Contacts between mitochondria and other organelles are coloured green. Image stacks of 140 planes were acquired every 9.2 seconds, for a total of 100 frames (15.3 min). Scale bar, 5 µm.

  7. 7.

    Video 7: Mitochondria-organelle contacts in time lapse, lattice light sheet images

    Volume rendering of mitochondria in COS-7 cells expressing fusion proteins targeting 3 other organelles, and labelled with Texas Red dextran and BODIPY 665/676. Contacts between mitochondria and other organelles are coloured yellow (ER), cyan (peroxisomes) red (Golgi), magenta (lysosomes) and blue (LDs). Image stacks of 140 planes were acquired every 9.2 seconds, for a total of 100 frames (15.3 min). Scale bar, 10 µm.

  8. 8.

    Video 8: Mitochondria-ER metaorganelle contacts with other organelles

    Volume rendering of mitochondria-ER contacts in COS-7 cells expressing fusion proteins targeting 2 other organelles, and labelled with Texas Red dextran and BODIPY 665/676. Contacts between mitochondria-ER metaorganelle and other organelles are coloured cyan (peroxisomes) red (Golgi), magenta (lysosomes) and blue (LDs). Image stacks of 140 planes were acquired every 9.2 seconds, for a total of 100 frames (15.3 min). Scale bar, 10 µm.

About this article

Publication history






Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.