Coverslip cleaning and preparation
Table of Contents
For single-molecule imaging, high tolerance, Number 1.5 coverslips were purchased from Warner scientific (25 mm) and precleaned with a modified version of a previously described protocol57,58. Coverslips were sonicated overnight (approximately 12 h) in 0.1% Hellmenex II (Sigma), followed by five washes in 300 ml of distilled water. Coverslips were then transferred to a clean chamber of 300 ml of distilled water and sonicated overnight again, followed by five more washes. Coverslips were then ethanol sterilized in pure, 200 proof ethanol and air dried in a clean tissue culture hood. After cleaning, coverslips were stored in an airtight container until use, and were discarded if not used within 30 days of cleaning.
For FIB-SEM, sapphire coverslips (3 mm diameter, 50 µm thickness, Nanjing Co-Energy Optical Crystal Co. Ltd, COE) were cleaned for at least 1 h in a basic piranha solution (5:1:1 water:ammonium hydroxide:hydrogen peroxide) followed by several washes in distilled water. The bottom of each coverslip had a thin layer of gold sputtered onto the side regions of the coverslip to distinguish top from bottom in subsequent steps (sputter coater Desk II, Denton Vacuum). After sputtering, coverslips were rinsed several times in distilled water and stored under vacuum in a desiccation chamber until use.
Plasmids and reagents
ER-mRFP (Addgene no. 62236), mTagRFP-T2-Mito-7 (Addgene no. 58041) (referred to as mitoRFP in the text), mTagBFP2-N1 (Addgene no. 54566), mEGFP-N1 (Addgene no. 54767), mEGFP-C1 (Addgene no. 54759) and mEmerald-Sec61b-C1 (Addgene no. 90992) have been described previously, and were gifts from Erik Snapp, Michael Davidson, or generated in house. EGFP-VAPB59, HA-PTPIP51 (ref. 16) and pHAGE-Tet-STEMCCA60 have been previously described and were gifts from Pietro De Camilli, Kurt De Vos and Robert Tijan, respectively.
All insertions and cassette changes were performed using the NEBuilder implementation of Gibson Assembly (New England Biolabs) unless specified otherwise, taking care to leave appropriate restriction sites for later changes. All constructs were sequenced before use and will be available on Addgene unless prohibited by copyright. Specific strategies and resulting plasmid maps are linked in the Supplementary Information.
Cell culture and transfection
COS7 cells were purchased from ATCC and used within 40 passages. Cells were maintained in complete DMEM (phenol red-free Dulbecco’s modified Eagle medium (Corning) supplemented with 10% (volume/volume) fetal bovine serum (Corning), 2 mM l-glutamine (Corning), 100 U ml−1 penicillin and 100 µg ml−1 streptomycin (ThermoFisher)). Cells were cultured at 37 °C in 5% CO2, passaging was accomplished with 0.25% (weight/volume) trypsin (Corning) and care was taken never to let cells grow to more than 85% confluency or be seeded at less than 25% confluency, as they often become less flat after this.
For single-molecule imaging, coverslips were precoated with 500 µg ml−1 phenol red-free Matrigel depleted for growth factors (Corning) for 1 h before plating in a 35 mm tissue culture dish. Cells were seeded to ensure less than 60% confluency at the time of imaging to maximize regions of ER within the focal plane of the objective when focused just above the coverslip. Transfections were performed after letting the cells adhere to the glass for at least 12 h using Fugene6 (Promega) at a 3:1 Fugene (µl) to DNA (µg) ratio according to the manufacturer’s protocol. Each 35 mm dish was transfected with 1.5 µg of DNA using the following ratio: 750 ng PrSS-mEmerald-KDEL (ER structure label), 500 ng mitoRFP (mitochondria structure label) and 250 ng of the HaloTag construct used for sptPALM, except for cells transfected with PTPIP51 which were given an extra 500 ng of PTPIP51-IRES-mTagBFP2. Imaging was always performed at least 12 h after transfection, but always before 24 h and taking care to avoid cells showing morphological changes from ER and mitochondria label expression that become increasingly common at later time points. For starvation experiments, cells were incubated in HBSS for the last 8 h before imaging, but complete medium was replaced on the cells immediately before imaging.
Cells prepared for Airyscan imaging were plated in high tolerance commercially acquired 35 mm imaging chambers (MatTek Life Sciences). Briefly, coverslips were precoated with 500 µg ml−1 phenol red-free Matrigel depleted for growth factors (Corning) for 20 min. Cells were then transfected in solution using Fugene6 (Promega) according to a modified version of the manufacturer’s protocol. Briefly, cells were resuspended after pelleting in prepared transfection complexes made according to the manufacturer’s recommendations in OptiMEM (ThermoFisher). Cells were incubated for 15 min at 37 °C before being plated on the coated coverslips in 2 ml of complete Medium. Imaging was performed 18–24 h after transfection.
Halo labelling
Cells were labelled for sptPALM by incubating with 10 nM PA-JF646 (ref. 29) in OptiMEM (ThermoFisher) for 1 min followed at least five washes with 10 ml of PBS, performed while simultaneously aspirating and taking care to never let the cells come in direct contact with the air. The cells were then washed once with 10 ml of complete DMEM and left to recover in 2 ml of complete DMEM for 15 min before imaging.
Cells were labelled for Airyscan imaging by replacing the complete medium on the cells with complete DMEM supplemented with 10 nM JF635 (ref. 61). The highly fluorogenic nature of this JaneliaFluor (JF) dye compound removes the need for washing steps, and the sample can directly be imaged on the microscope.
Microscopy and imaging conditions
Single-molecule imaging was performed using a custom widefield microscope assembled in an inverted Nikon Ti-E outfitted with a stage top incubator to stabilize temperature, CO2, and relative humidity during imaging (Tokai Hit). The flat lamella of cells where sptPALM is possible (approximately 500 nm thick or less) were located using eye pieces to visualize the ER and mitochondria localization. To avoid bias, the experimenter was always blinded to the single-molecule tracers. In experiments where PTPIP51-IRES-mTagBFP2 was overexpressed, the cells were selected using the fluorescence of the mTagBFP2 in addition to the ER and mitochondria structure.
Excitation was performed using three fibre-coupled solid state laser lines (488 nm, 561 nm, 642 nm; Agilent Technologies) introduced into the system with a conventional rear-mount TIRF illuminator. The angle of incidence was manually adjusted for each cell beneath the critical angle to maximize the evenness of the illumination in the ER. The illumination in the 488 nm and 561 nm channel was manually adjusted based on the brightness of the sample to minimize fluorescent bleed-through, but the total power on each line was always kept less than 50 µW and 150 µW total in the back aperture, respectively. Single molecules were always imaged using a constant total power of 11.5 mW of 647 nm light in the back aperture. If necessary, a small amount of 405 nm light was introduced to tune the photoactivation rate of the molecules being tracked, but in practice this was rarely needed.
Emitted light was collected with a ×100 α-plan apochromat 1.49 numerical aperture oil immersion objective (Nikon Instruments) and focused through a MultiCam optical splitter (Cairn Research). The emission path was split onto three arms of the splitter using a 565LP and a 647LP dichroic mirror (Chroma) placed sequentially in the optical path to split the light from the 488 nm and 561 nm channels, respectively. These emission paths were additionally cleaned up by passing the emitted light through a 525/50 BP and a 605/70 BP filter (Chroma), respectively. The remaining light transmitted through the MultiCam represents the far-red signal where the single molecules of HaloTag-linked dye are imaged, and the path was passed through an additional 647LP filter to clean up any stray light in the system that could decrease the resolving power of the sptPALM approach. All three channels were collected from electronically synchronized iXon3 electron multiplying charged coupled device cameras (EM-CCD, DU-897; Andor Technology). To image quickly enough, the field of view was reduced to a 128 × 128 pixel square (20.48 µm × 20.48 µm). The location of the square was carefully chosen for each sample to contain the flattest region of ER possible while remaining near the centre of the camera chip, since the objective in use is chromatically corrected to high precision only near the centre of the field of view. Imaging was performed with 5 ms exposure times for 60–90 s at a time, and the timing of each frame was monitored using an oscilloscope directly coupled into the system (mean frame rate of approximately 95 Hz).
Airyscan imaging was performed using a commercially acquired Zeiss LSM 880 microscope with a live cell incubation system (Zeiss Microscopy). Briefly, labelled samples were sequentially excited with laser lines at 633 nm, 561 nm and 488 nm. Emission fluorescence is collected using a ×63 1.4 numerical aperture oil immersion objective (Zeiss Microscopy) with an open pinhole and passed through an appropriate custom bandpass filter based on the expected emission profile of the sample to the arrayed detector for the Airyscan unit (561 nm, 488 nm, BP495-550 + LP570; 633 nm, BP570-620 + LP645). Airyscan reconstruction and deconvolution was performed using the default settings (filter size = 6). Images were pseudocoloured and prepared using Fiji (NIH) for visibility.
Channel registration and spectral analysis
At the time of imaging, a crude channel alignment was performed using a sparse distribution of Tetraspek beads on a coverslip, prepared and imaged as for other sptPALM samples. The angle of the dichroic mirrors was manually adjusted to get as much overlap between the channels in the main field of view as possible. In practice, this alignment was sufficient to support the manual steps in the tracking pipeline (see below), but applications requiring more precise alignment were accomplished using a custom subpixel alignment pipeline in Fiji.
Variation in expression level of the three markers (ER structure marker, mitochondria structure marker and single-molecule tracer) due to uneven transfection is not a trivial issue, and often required manual adjustment in the relative laser power for the 488 nm and 561 nm lines by the experimenter. Since this could in principle create artefacts in the automated analysis pipeline or introduce erroneous single-molecule localizations as a result of bleed-through, all of the samples were run through an automated spectral analysis pipeline that checked for fluorescence contamination from the blue-shifted channels. Any samples where the detectable bleed-through contamination was significant compared to the signal of single molecules were removed before downstream analysis was performed.
Localization and tracking
Localizations were identified in the single-molecule data sets using a previously described pipeline62 to estimate positions and precision of localization using a maximum likelihood estimation (MLE)-based fitting approach. The quality filter used in the downstream tracking pipeline limited analysable localizations to those identified with precision (as estimated from the Cramér–Rao lower bound) in the range of 20–30 nm.
Trajectories were assembled from single-molecule images using the TrackMate plugin in Fiji63,64. Linking parameters were experimentally selected for each dataset to minimize visible linkage artefacts as determined by eye. Resulting putative trajectories were then projected on to the simultaneously collected ER network structure, and manually curated to remove any trajectory linkages that are close in 2D but far from one another in the underlying organelle structure. This step proved crucial to assembling trajectories that moved within the structure without linkage artefacts. Resulting trajectories were exported from TrackMate and imported into MATLAB for subsequent analysis.
Spatial density analysis and contact site identification
Spatial probability density is mapped by choosing the spatiotemporal boundaries of the data to be analysed (x, y, t) and binning the resulting localizations into 30 nm square pixels. The resulting counts are normalized to the total number of localizations within the dataset, and as such probability represents solely the likelihood of a single molecule falling in a certain pixel if chosen at random (that is, a spatially defined probability mass function). This effectively minimizes the effects of differences in photoactivation efficiency or tagged protein expression level when identifying the boundaries of a contact site. Note that this analysis does not assume anything about the motion of the trajectories, the orientation or stability of the contact site, or the nature of molecular interactions—all of this information is analysed in subsequent steps.
The initial location of contact sites was identified from the spatially defined probability density when calculated over the entire image, but the location and boundaries often had to be manually refined, especially under conditions where contact sites move or change orientation (Extended Data Fig. 5).
Spatial clustering and diffusion landscape estimation
To generate a map of the diffusion landscape within contact sites, the space inside the contact site was divided into distinct compartments by Voronoi tessellation informed by the probability density at the site. The trajectories associated with the site were broken into single steps and assigned to a tessellation by the location of the beginning of the step32,33. Bayesian inference was then used to model the resulting distribution as an overdamped Langevin system within each tessellation, assuming single molecules in the same space at distinct times can be viewed as independent experiments probing the same molecular environment (see Supplementary Text, section 6c). The resulting diffusive component was reported for each tessellation as an effective 2D diffusion coefficient.
Identification of latent states in single trajectories
All trajectories longer than 500 steps (approximately 5.5 s) were analysed using a non-parametric Bayesian modelling technique (Hierarchical Dirichlet Process Modelling, HDP) coded using Python to estimate latent state changes in single-molecule behaviour30,31. Briefly, the system was treated as a switching linear dynamical system (SLDS). As in previous work30,31, an overdamped Langevin equation was used to interpret the parameters of the linear dynamical system used in the SLDS model. This approach removes the need for an upper bound on the number of potential states common in single particle tracking analysis approaches (Hidden Markov Models, and so on), which become intractable in a spatially complex environment like the ER (Supplementary Text, section 6d). Eigen-decompositions of the implied force and diffusion tensors for each determined state enables one-dimensional analysis through tensor diagonalization. Note this SLDS treats thermal fluctuations as a distinct component from measurement noise, allowing diffusive properties of the system to be quantitatively estimated independently of measurement noise (for example, localization errors).
High pressure freezing and freeze-substitution
Immediately before freezing, cells were manually inspected using an inverted widefield microscope to ensure reasonable morphology and viability. Cells were then transferred to a water jacketed incubator (ThermoFisher, Midi 40) where they were kept at 37 °C in 5% CO2 and 100% humidity until ready for freezing. Each sapphire coverslip was removed one at time from the incubator, overlaid with a 25% (weight/volume) solution of 40,000 MW dextran (Sigma), loaded between two hexadecane-coated freezing platelets (Technotrade International), and placed in the HPF holder. Freezing was then performed using a Wohlwend Compact 2 high pressure freezer, according to the manufacturer’s protocol. Frozen samples were stored under liquid nitrogen until freeze-substitution was performed.
Freeze-substitution was performed with a modified version of a previously described protocol22,65. Briefly, frozen samples were transferred to cryotubes containing freeze-substitution media (2% OsO4, 0.1% Uranyl acetate and 3% water in acetone) and placed in an automated freeze-substitution machine (AFS2, Leica Microsystems). A freeze-substitution protocol was used as previously described23, and the samples were then washed three times in anhydrous acetone and embedded in Eponate 12 (Ted Pella, Inc.). The sapphire coverslip was then removed and the block was re-embedded in Durcapan (Sigma) resin for FIB-SEM imaging.
FIB-SEM
FIB-SEM was performed essentially as previously described21,23,66. Briefly, a customized FIB-SEM using a Zeiss Capella FIB column fitted at 90 degrees on a Zeiss Merlin SEM was used to sequentially image and mill 8 nm layers from the Durcapan-embedded block. Milling steps were performed using a 15 nA gallium ion beam at 30 kV to generate two sequential 4 nm steps. Data was acquired at 500 kHz pixel−1 using a 2 nA electron beam at 1.0 kV landing energy with 8 nm xy resolution to generate isotropic voxels. Data sets were registered postacquisition using a SIFT-based algorithm67.
Voxel classification and surface determination
Although several automated segmentation protocols exist for reconstruction of organelles from FIB-SEM data10,21, we found that minor errors in voxel classification within the contact sites themselves obscured our ability to analyse the local curvature in sufficient resolution for our needs (see Supplementary Text, section 3a, for discussion). Consequently, we selected a few small volumes containing mitochondria and manually classified the voxels for the ER using Amira (ThermoFisher). We used a modified watershed algorithm to classify the mitochondrial membranes but performed a manual curation to remove artefacts. Potential contact sites on the ER surface were identified as regions of ER membrane within 24 nm (approximately three pixels) of the OMM, as measured by dilation of the OMM (see Supplementary Text, section 3b, for discussion of contact site distances).
Triangulation, smoothing and curvature analysis
Triangulated surfaces were fit to the voxel classifications using a marching cubes-based algorithm implemented directly in Amira. To avoid voxel-step artefacts in the surface, gaussian smoothing was applied to the voxel data using a local likelihood measure selected over a kernel size relevant for the expected curvature of the underlying membrane (Supplementary Text, section 3c). The resulting triangulated surfaces were rendered for use in the figures using Amira, and they serve as the scaffold for subsequent 3D curvature analysis.
Mean local curvature of the ER was computed as a scalar field over the triangulated surface using a 20-layer neighbourhood to fit a quadratic form along the two principal curvature axes. Note this is different from gaussian curvature, and the resulting value is negative in strictly concave regions and positive in regions that are strictly convex. Scalar fields were calculated and mapped using the curvature field module in Amira. Details are given in the Supplementary Text, section 3d–e.
Statistics and reproducibility
Each single-particle tracking dataset contains a total of at least 16 regions (20.48 μm × 20.48 μm) divided over at least two experiments, each selected from a different cell. Not all of these contained mitochondria-associated contact sites, though all had contact sites of some kind (Extended Data Fig. 2). FIB-SEM data was visually examined in three COS7 cells, but all data shown or quantified in the paper comes from a single representative cell. Single contact site analysis was performed on each of the hundreds of contact sites analysed, each of which contains anywhere from 1–50 VAPB trajectories. All representative images of correlated sptPALM data throughout the paper are a single image of at 16–24 similar acquired, the number of which is listed in the associated quantifications. ERMCS structures shown from FIB-SEM data are representative examples of 25 similar contact sites, except where explicitly stated to be otherwise, such as unusual examples. Airyscan images are representative of at least 30 cells, and Airyscan experiments were performed five times with multiple combinations of fluorescent labels with indistinguishable results (see Supplementary Text, section 8, for a more complete discussion of reproducibility).
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.