An open-access volume electron microscopy atlas of whole cells and tissues.

The initial datasets of seven common wild-type cultured cells (from cancer cells to immune cells) and three tissue samples (from mouse pancreatic islets to Drosophila neural tissues) are shown in Extended Data Table 1a . Detailed procedures are described below.

Cultured cell sample preparation

Cultured cells were cryo-fixed by high pressure freezing (HPF), which vitrified a sample on a millisecond time scale to best preserve any dynamic structural details. Although HPF sometimes can cause segregation artifacts (i.e. freezing damage) 18 , it minimizes any possibility of artifacts that might be of concern with chemical fixation. The subsequent freeze substitution labeled biomolecules with heavy metals to provide contrast for EM. The detailed procedures are presented in the Supplementary Information . HeLa cells (CCL-2), ID8 cancer, THP-1 macrophage, Jurkat cells were purchased from ATCC, SUM159 cells were gift from Dr. Tomas Kirchhausen laboratory at Harvard Medical School. Further authentication was not performed for this study. Mycoplasma contamination testing was performed on HeLa, THP-1 macrophage, Jurkat, and SUM159 cells and were negative. ID8 cancer cells were not tested for mycoplasma contamination. 1.

HeLa cells

HeLa cells

(CCL-2) were purchased from American Type Culture Collection (ATCC). HeLa cells were maintained in EMEM medium (ATCC, 30-2003) supplemented with 10% FBS (Corning, 35-011-CV) and 1X penicillin-streptomycin solution (Corning, 30-002-CI). Trypsinized HeLa cells were seeded on pre-cleaned and edge-coated sapphire coverslips. Immediately prior to freezing, live cells were inspected to ensure cell morphology and viability. After quality assurance the cells were transferred to a water jacketed CO 2 incubator (Thermo Fisher Scientific, Midi 40) kept at 37° C, 5% CO 2 , and 100% humidity while awaiting freezing. Each coverslip was removed from the incubator immediately prior to the freezing procedure, and dipped three times in the 25% w/v mixture of Bovine Serum Albumin (BSA) (B4287, Sigma Aldrich)) in the cell medium, which served as a cryo-protectant to prevent ice crystal formation 19 . The samples were then high-pressure frozen between two Aluminum planchettes (Technotrade International, 389 and 479) in Wohlwend HPF Compact 02 High Pressure Freezing Machine (Technotrade International). Once the samples were frozen, they could be stored indefinitely in liquid nitrogen for subsequent freeze-substitution (FS) and resin embedding (RE). FS was performed using automated FS machine (AFS2, Leica Microsystems): coverslips were transferred to cryotubes containing FS media (2% OsO 4 , 0.1% Uranyl Acetate (UA), and 3% water in acetone) under liquid nitrogen, followed by a programmed FS schedule. Resin embedding was performed immediately after FS. Samples were removed from the AFS2 machine, washed 3 times in anhydrous acetone for a total of 10 min and embedded in Eponate 12 which was polymerized for 48 hours at 60°C. Following EPON embedding, the coverslip was separated from the resin block containing the cells by sequential immersion in liquid nitrogen and hot water. EPON was chosen as the initial embedding resin due to less embedding artifacts. However, EPON generates streaking artifact during FIB-SEM imaging 1 . A thin layer (5–10 μm) of Durcupan on the specimen surface facing the FIB beam can effectively mitigate such streaks during FIB-SEM imaging. Therefore, once the coverslip was removed, the exposed surface was immediately re-embedded in Durcupan ACM resin (Sigma Aldrich, set 44610). The detail procedures of coverslip preparation, high pressure freezing, free substitution, and resin embedding are presented in the Supplementary Information . 2. Cytotoxic T cell and T cell/Cancer cell conjugate (CTLs) Primary cytotoxic T cells in this study were isolated from female C57BL/6-Tg(TcraTcrb)1100Mjb/J mice (Jackson Labs), also known as OT-I mice, aged 8–12 weeks. Mice were housed in individually ventilated cages within animal rooms maintained on a 14:10-hour, light:dark cycle. Animal rooms were temperature and humidity-controlled, between 68–79°F and 30–70% respectively, with 10 to 15 room air exchanges per hour. Animals were maintained in accordance with the Guide for the Care and Use of Laboratory Animals . Genentech is an AAALAC-accredited facility and all animal activities in the research studies were conducted under protocols approved by the Genentech Institutional Animal Care and Use Committee (IACUC). To generate CTLs from OT-I mice, splenocytes were isolated and stimulated with 10nM OVA257-264 peptide (AnaSpec, Fremont, CA, USA) in 10% RPMI (RPMI 1640 plus 10% fetal bovine serum (Fisher), 2mM L-glutamine, 50U/mL penicillin/streptomycin, and 50μM β-mercaptoethanol. Following 3 days of stimulation, cells were resuspended in complete media plus 10 IU/mL recombinant human IL-2 (rHIL-2, Roche), and seeded in fresh media at 0.5×10 6 cells/mL every 48 hours. ID8 cancer cell line (ATCC) was grown in RPMI media (Gibco) with 10% FBS (Gibco). In-vitro activated CTLs from Ova-transgenic (OT-I) mice were combined with adherent ID8 murine ovarian cancer cells. To facilitate recognition of cancer cells by the OTI-I CTLs, ID8 cells suspended in complete medium (RPMI with 10% FBS, Gibco) were incubated with OVA 257–264 SIINFEKL peptide for 1 hour at 37°C. The cells were washed three times with complete medium, and 10 5 cells were added to each well of a 24-well plate. Each well contained a single sapphire coverslip coated with human Fibronectin (Corning). The ID8 cells were allowed to settle and adhere for 2 hours in an incubator at 37°C/5% CO 2 . At this time, the media in the well was replaced with 250uL phenol-red free RPMI (Gibco). Care was taken to ensure that media and cells were stored in 37°C/5% CO 2 for all incubation periods. Immediately after changing the media, 10 5 CTLs in 50uL complete medium were added to the well. The CTLs were allowed 7 minutes to find their targets and secrete lytic granules before fixation through high pressure freezing. The relatively short incubation time favors capture of CTL:target conjugates in an early-stage of interaction. The remaining cryofixation, FS, and resin embedding procedures were identical to those described for HeLa cells. Although there is a minor membrane damage in the area of immunological synapse, which most likely occurred during high-pressure freezing or freeze-substitution or resin embedding 18 , HPF remains the only viable option for non-chemical fixation of samples more than a few hundred nanometers thick. Tissue sample preparation 1.

Mouse pancreatic islets

Facilities for animal keeping and husbandry are certified and available with direct access on campus in Dresden (including facilities at Paul Langerhans Institute Dresden and Max Planck Institute of Cell Biology and Genetics). The animals were housed at 20–24 °C, with 45–65% humidity in an artificial day–night (12 hour) rhythm. All animal experiments were performed according to guidelines of the Federation of European Laboratory Animal Science Associations (FELASA) and recommendations and are covered by the respective licenses for those experiments from the local authorities. Licenses for animal experiments are approved by the State Directorate Saxony under license number DD25-5131/450/6; TV A 6/2018. Pancreatic islets of 9-week-old C57BL/6 female mice were isolated as previously described 20 . They were cultured overnight in standard culture media (RPMI 1640 (Gibco) with 10% FBS, 20 mM HEPES, 100 U/ml each penicillin and streptomycin) containing 5.5 mM glucose. Prior to high pressure freezing the islets were subjected to 1 hr incubation in Krebs-Ringer buffer containing either 3.3 mM or 16.7 mM glucose. Islets were frozen with a Leica EM ICE high pressure freezer (Leica Microsystems, Germany) and kept in liquid nitrogen until freeze substitution. Although freezing damage was observed in couple nuclei, the preservation of cytoskeletal elements benefits from this fixation. High pressure frozen islets were substituted as previously published 21 or according to a novel protocol: first the samples were substituted in a cocktail containing 2% OsO 4 , 1% UA, 0.5% glutaraldehyde, 5% H2O in Acetone with 1% Methanol at −90°C for 24 hours. The temperature was raised to 0°C over 15 hours followed by 4 washes with 100% acetone for 15 min each and an increase in temperature to +22°C. Afterwards the samples were incubated in 0.2% thiocarbohydrazide in 80% methanol at RT for 60 min followed by 6 × 10 min washes with 100% acetone. The specimens were stained with 2% OsO 4 in acetone at RT for 60 min followed by incubation in 1% UA in acetone + 10% methanol in the dark at RT for 60 min. After 4 washes in acetone for 15 min each they were infiltrated with increasing concentrations of Durcupan resin in acetone followed by incubation in pure Durcupan and polymerization at 60°C. For quality control the blocks were sectioned with a Leica LC6 ultramicrotome (Leica microsystems) and 300 nm sections were put on slot grids containing a Formvar film. Tilt series ranging from −63° to +63° were acquired with a F30 electron microscope (Thermo Fisher Scientific) and reconstructed with the IMOD software package 22 . 2. Drosophila brain samples (fan-shaped body and accessory calyx) While thin tissue samples (< 200-μm-thick) can be cryo-fixed, in many cases thick tissue specimens can not (e.g., large brain tissue) thus requiring various chemical fixation and staining protocols.

Drosophila brain tissues from 5-day-old adult

(Genome type: iso Canton S G1 × w1118 iso 5905) were prepared according to Progressive Lowering of Temperature and Low Temperature Staining (PLT-LTS) progressive heavy metal enhancement protocol described previously 23 . In details, after tissue dissection and pre-fixation, we osmicated tissue in 1% OsO 4 , then 1.5% K ferrocyanide, followed by a complete wash. Afterwards, a transfer to 1% thiocarbohydrazide for 15 min at 22°C, then a complete wash, followed by 2% osmium for 30 min at 22°C. After osmication, we stained in lead aspartate for 30 min at 55°C first, then 1 h at 22°C. Finally, the tissue was dehydrated in a Leica AFS freeze-substitution chamber: the temperature was dropped from 4°C to −25°C, and the concentration of acetone or ethanol was increased for 20 min in each of 10%, 30%, 50%, 70%, 80%, 90%, and 97%. The subsequent low temperature en bloc staining was performed in either 0.2% uranyl acetate in acetone, or 1% EPTA in 97% ethanol. Specimens were infiltrated and embedded in Durcupan. Preparation for FIB-SEM After Durcupan re-embedding, a 3D X-Ray tomogram of the entire block was taken using an XRadia-510 Versa micro X-Ray system (Carl Zeiss X-ray Microscopy, Inc.). The 3D X-Ray tomograms allowed for robust Region of Interest (ROI) selection. Specifically, it identified cells with good morphology (for example a properly shaped cell, a cell in a desired stage of mitosis, a T-cell attacking a cancer cell, etc.), and also determined the proper orientation for the tissue samples. Once a potential ROI was identified, the sample block was then re-mounted to the top of a 1 mm copper post (using Durcupan) which was in contact with the metal-stained sample for better charge dissipation, as previously described 1 . Each sample was oriented on the copper post to allow the shortest distance along FIB milling direction. A small vertical sample post was trimmed to the ROI with a width of ~100 μm and a depth of 60–80 μm in the direction of the ion beam for each sample. Sample trimming used an ultramicrotome (EM UC7, Leica Microsystems), was done in few iterations with X-Ray tomogram collected after each step. Once the desired ROI was trimmed, a thin layer of conductive material of 10-nm gold followed by 100-nm carbon was sputtered using a Gatan PECS 682 High-Resolution Ion Beam Coater. The coating parameters were 6 keV, 200 nA on both argon gas plasma sources, 10 rpm sample rotation with 45-degree tilt. FIB-SEM system advances A Zeiss Gemini 500 and multiple Zeiss Merlin SEM systems were customized for this work to enable the two orders of magnitude improvement in imageable volume (from ~500 μm 3 to greater than 100,000 μm 3 ) compared to our prior work 1 . Such improvement primarily comes from: 1) higher precision and stability of FIB milling control to extend reliable long-term acquisition to 4-nm voxels, and 2) enhanced SEM signal detection using secondary electrons to achieve faster imaging, hence reduce electron radiation energy density (from ~400 to ~80 keV/nm 3 ) that further improves FIB milling control. Specifically, to ensure precise and stable milling at 4-nm sampling, we integrated a more stable FIB column (Capella from Zeiss) and repositioned at 90 degrees to the SEM column ( Fig. 1a ). The Capella FIB column provided higher practical milling currents up to 30 nA compared to 7 nA of the Magnum FIB column (Thermo Fisher Scientific, previously FEI) used in our prior work 1 . We further tightened its FIB emission current control band from ± 100 pA to ± 50 pA to obtain a consistent milling beam profile. We also re-configured the milling closed-loop control (Fig. 10 in Xu et al., 2017 1 ) to further reduce the milling variation between FIB reheat cycles thus enabling long-term acquisition at high electron radiation energy density. This FIB milling optimization sufficiently reduced milling artifacts hence allowed for the switching of SEM signal detection from backscattered electrons to secondary electrons for faster imaging (600 V vs. 0 V specimen bias in Fig. 12 of Xu et al., 2017 1 ). The new SEM signal detection scheme lowered electron radiation energy density by 5x (from ~400 to ~80 keV/nm 3 ), which in turn further mitigated FIB milling artifacts and instability. In addition, customized 100-μm-thick molybdenum apertures were laser machined to define more accurate and consistent FIB milling currents, and to achieve more than 2x longer lifetime; and the custom NI LabVIEW control software was re-designed: to include new configurations of FIB milling closed-loop control which improve milling variation; and to switch from the Zeiss RemCon serial port communication protocol to the Zeiss API which reduced the overhead by several seconds per imaging/milling cycle. The optimization of SEM imaging conditions was then guided by the following principles. The isotropic resolution limits are convoluted by the waist size of the incoming primary electron beam and the scattering volume of the penetrating primary electrons. While spherical and chromatic aberrations of the electromagnetic lenses can be mitigated by smaller beam currents at modest numerical aperture to achieve the best beam focus and x-y resolution, lower primary beam landing energy generates fewer scattering events, resulting in smaller scattering volume which improves resolutions along all three axes. However, any further reduction below the optimal energy will result in a reduced scattering ratio of the heavy metal stain atoms over lighter background atoms of the sample resin, thus fading contrast. Through investigating a multi-dimensional operation space, we found that a beam current of 200 to 300 pA and a landing energy of 700 to 900 eV are optimal for isotropic 4-nm imaging. Such positive synergy between faster SEM scanning and robust FIB milling has extended reliable imaging acquisition at 4-nm isotropic voxels from less than a week to months. Primarily limited by time, the maximum volume can be seamlessly extended.

FIB-SEM imaging

As summarized in Extended Data Table 1a , all samples were imaged by a 0.14–0.3 nA electron beam with 0.7–1.2 keV landing energy at 0.1–2.0 MHz. In most cases, both backscattered and secondary electron signals were collected by InLens detector to provide better signal-to-noise ratio. The time to acquire these volumes ranged from two to five weeks of uninterrupted imaging. Faster SEM scanning rates (5x–10x scanning speed improvement) were applied on Drosophila samples with stronger staining contrast while maintaining similar image quality, thus achieving maximum volume of 175,000 μm 3 within three-week imaging. For all samples, the x-y pixel size was set at 4 nm. A subsequently applied focused Ga+ beam of 15 nA at 30 keV strafed across the top surface and ablated away 4 nm of the surface. The newly exposed surface was then imaged again. The ablation – imaging cycle continued for two to five weeks of uninterrupted imaging to complete FIB-SEM imaging one sample. The sequence of acquired images formed a raw imaged volume, followed by post processing of image registration and alignment using a Fiji plugin based on Scale Invariant Feature Transform (SIFT) algorithm 24 to form isotropic 4 nm voxels. The final aligned stack consisted of an isotropic volume, containing multiple complete cells, which can be viewed in any arbitrary orientations.

Evaluation of FIB-SEM resolution

In order to evaluate the actual resolution in the acquired FIB-SEM datasets we analyzed the transitions of the edges of the ribosomes abundant within the cell volumes. Ribosomes are macromolecular machines consisting of RNA and associated proteins. Their small size (20–30 nm), abundance in the cytoplasm of living cells, and the fact that they are stained with high EM contrast makes them good candidates for resolution evaluation. First, we selected volumes of approximately 1000–2000 × 1000 × 1000 pixels in cultured cell datasets. These volumes were selected in the middle of datasets in areas with large number of ribosomes. We then calculated histograms of grey level values in these volumes and subtracted ~0.8 * histogram max to bring the average cytoplasmic material signal level close to 0. We then used the Laplacian of Gaussian (LoG) algorithm (part of Python scikit-image package 25 ) to select blobs. We set the limit of standard deviation for LoG between 1 and 2.5 pixels (all these datasets have 4-nm voxels) to ensure that selected blobs were predominantly ribosomes. Filters were used to exclude the following blobs: average value below zero (this meant LoG algorithm failed), edge value above 0.4 times of max value or 37%–63% transition longer than 15nm (the latter two happen where there is another blob or another feature with high signal level in close proximity—making subsequent edge transition analysis inaccurate), amplitudes below certain threshold, so that at least approximately 3000 blobs per dataset were remained for analysis. The edge transitions in all three directions were analyzed for these remaining blobs, in particular we determined 37%–63% transitions (the value used by Zeiss in their resolution estimation) and 20%–80% transitions (close to 1 sigma value). The summaries of the 37%–63% transitions and 20%–80% transitions for the 3D ribosome blobs selected in cultured cell datasets are plotted in Fig. 1b . The histograms of the 37%–63% transitions for these datasets are shown in Extended Data Fig. 3 . The examples of 3D blobs for each dataset are presented in Extended Data Fig. 4 – 9 . For each dataset the cross-sections and transition analysis for 3 brightest and 3 dimmest ribosomes are shown. Accurate determination of FIB-SEM milling rate The position of a sample block face was measured by the SEM beam and the FIB beam independently. An in-line auto focus routine 7 kept the SEM images in focus throughout the entire image acquisition. While the working distance of the SEM beam detected the block face position by measuring its distance from the SEM objective lens, the FIB milling position under closed-loop control indicated the block face position directly from the orthogonal angle. As the sample block face being imaged then milled away layer-by-layer, the SEM working distance and the FIB beam position tracked the block face location accordingly. Therefore, we estimated the total thickness milled away by FIB based on the changes of either SEM working distance or FIB milling position from the beginning to the end of the FIB-SEM acquisition. The average milling rate of z-step was then calculated as the total milling depth divided by the number of image frames. Here we report the average z-scaling factors derived from both measurement approaches in Extended Data Table 1b , which can be used to scale the dataset to generate real isotropic voxels. For example, HeLa Cell 2017-06-21 dataset has a z-scaling factor of 1.31 which means the actual z-step is averaged at 5.2 nm instead of 4.0 nm.

Image emulation to visualize the value of finer resolution

To emulate 8-nm images, we first applied a 6-nm Gaussian blur filter and then added Gaussian noise with a standard deviation of 22 to 4-nm images. The higher shot noise at 8-nm sampling was expected due to the 10x lower electron dose of ~12 e/nm 3 , compared to that of ~120 e/nm 3 at 4-nm sampling. The 6-nm Gaussian blur and added noise ( Fig. 1d , g , and j ) were underestimated compared to that of the real 8-nm images ( Fig. 1e , h , and k ) collected on the same type of specimens. The 8-nm emulated images with comparisons to 4-nm images and real 8-nm images are shown in Fig. 1c – k and Supplementary Video 1 . Statistics and Reproducibility In Fig. 1c – k , Fig. 2b – d , Fig. 3b – d , Fig. 4a – l , and Extended Data Fig. 2a – c , the sub-panels represent different locations within the dataset that demonstrate similar staining and resolution, showing reproducibility of our imaging quality.

Supplementary Material 1790170_Sup_Info 1790170_SI_Guide 1790170_SD_Fig_1 1790170_SD_ED_Fig_4 1790170_SD_ED_Fig_5 1790170_SD_ED_Fig_6 1790170_SD_ED_Fig_7 1790170_SD_ED_Fig_8 1790170_SD_ED_Fig_9 1790170_SD_ED_Fig_10 1790170_Sup_Vdo_1 1790170_Sup_Vdo_2