Correlative super-resolution fluorescence and electron microscopy using conventional fluorescent proteins in vacuo

1 Super-resolution light microscopy, correlative light and electron microscopy, and volume electron 2 microscopy are revolutionising the way in which biological samples are examined and understood. 3 Here, we combine these approaches to deliver super-accurate correlation of fluorescent proteins to 4 cellular structures. We show that YFP and GFP have enhanced blinking properties when embedded in 5 acrylic resin and imaged under partial vacuum, enabling in vacuo single molecule localisation 6 microscopy. In conventional section-based correlative microscopy experiments, the specimen must be 7 moved between imaging systems and/or further manipulated for optimal viewing. These steps can 8 introduce undesirable alterations in the specimen, and complicate correlation between imaging 9 modalities. We avoided these issues by using a scanning electron microscope with integrated optical 10 microscope to acquire both localisation and electron microscopy images, which could then be precisely 11 correlated. Collecting data from ultrathin sections also improved the axial resolution and signal-to- 12 noise ratio of the raw localisation microscopy data. Expanding data collection across an array of 13 sections will allow 3-dimensional correlation over unprecedented volumes. The performance of this 14 technique is demonstrated on vaccinia virus (with YFP) and diacylglycerol in cellular membranes (with 15 GFP). 16

Introduction electron images of ultrathin IRF sections are unique in that the information from both imaging 23 modalities is gathered from the same physical slice (50 to 200 nm thick). As such, the axial (z) 24 resolution corresponds to the physical slice thickness and the fluorescence images taken from IRF 25 sections are inherently better than the diffraction limit. However, the lateral (xy) plane will still be 26 diffraction limited if a widefield fluorescence (WF) microscope is used, so structures below ~200 nm 27 cannot be resolved. Intriguingly, it was recently shown that standard FPs blink in IRF sections when  To provide a baseline readout of FP blinking, we first imaged YFP-A3 vaccinia and GFP-C1 in whole 13 fully-hydrated HeLa cells that were fixed in 4% PFA and mounted in Citifluor AF4 using the SECOM light 14 microscope at atmospheric pressure in WF mode (Fig.1). Both YFP and GFP displayed blinking behaviour 15 when the laser power was increased to a density of 330 W/cm 2 , and progressive photobleaching was

27
Thin sections of ~200 nm were cut from the resin block (Fig.2, step 2). Fluorescent cells were located 28 using the SECOM light microscope in WF mode within the SEM at 200 Pa partial pressure of water 29 vapour, which we previously demonstrated to be optimal for WF imaging in vacuo (Brama et al., 2015).

30
The laser power was increased to a density of 330 W/cm 2 , to drive FP blinking, and sequences of 31 ~30,000 images were collected. The vacuum pressure was then decreased to 10 -3 Pa (high vacuum) for 32 optimal backscattered electron imaging. This WF-SR-EM cycle could be repeated sequentially to image 33 different cells in the same section, or the same cell across serial sections to build up a 3D volume 34 (Fig.2, step 3). The SR images were reconstructed using ThunderSTORM to generate individual 35 localisations (Fig.2, step 4). Lastly, overlays of WF-EM and SR-EM were created to identify fluorescently 1 labelled cellular structures (Fig.2,  Reconstructed SR images of YFP-A3 vaccinia-infected HeLa cells in IRF sections in vacuo were of higher 5 resolution than WF images of the same section ( Fig.3A; SMovie 3). Resolution was measured using 6 Fourier Ring Correlation (FRC) (Nieuwenhuizen et al., 2013) to be 85.5±13.1 nm, assessed using 10 7 patches of size 2.7 µm 2 across the image and twenty statistical repeats. Only patches that contained a 8 significant amount of data were used in the evaluation since there were substantial areas of the image 9 with very low information content. While this resolution was slightly lower than that typically achieved   individual virus particles (Fig.3D). The precision with which the YFP signal could be localised to virus 20 particle structure in WF-EM overlays compared with SR-EM overlays was assessed using a standard 2-21 colour Pearson's correlation coefficient measurement; a more negative coefficient reflecting a stronger correlation, since the viral particles were darker in the SEM image, and brightness was 23 therefore anti-correlated. The WF-EM correlation was found to be -0.091, -0.002 and -0.053 (Fig.3C,   24 left to right, respectively), whereas the SR-EM correlation was considerably higher, at -0.165, -0.135, 25 and -0.147 (O) (Fig.3D, left to right, respectively), confirming the visual observation that in these 26 images the SR signal was more strongly localised to viral particles.

28
GFP-C1 blinks in vacuo, enabling SR CLEM localisation of the DAG in cell membranes 29 Since GFP has previously been shown to blink on a longer timescale than YFP (Bagshaw and Cherny,

31
GFP was feasible in sections of resin-embedded HeLa cells expressing the GFP-C1 construct. We found 32 that GFP also blinks well in IRF sections in vacuo (SMovie 4). Reconstructed SR images of GFP-C1 had a 33 resolution of 87.5±27.7 nm, measured by assessing seven 2.7 µm 2 patches using FRC (Fig.4A); correspondence between GFP signal and membrane structures in SR-EM overlays compared with WF-EM 1 overlays. In the SR images, it was possible to localise DAG to membrane subdomains within the Golgi 2 apparatus, cristae and perimeter membranes of mitochondria, endoplasmic reticulum and the 3 perimeter membrane of a putative double-membrane autophagosome (Fig.4C,D). These organelle 4 localisations could not have been identified from fluorescence data alone, and required EM to add 5 ground-truth to the SR reconstructions.  To assess the effect of vacuum pressure on FP blinking, we acquired SR images at atmospheric pressure 9 and 10 -3 Pa to complement the data gathered at 200 Pa (Fig.5). Had we collected three sequential SR 10 datasets from the same region at different vacuum pressures, photobleaching would influence the 11 comparison between reconstructions. Therefore, we collected one dataset from each of three

25
We then showed that it was possible to modulate the blinking of GFP by cycling through different 26 vacuum pressures sequentially whilst maintaining continuous illumination (Fig.6, SMovie 5). After 27 establishing high vacuum, the chamber could be repeatedly cycled from 10 -3 Pa to 200 Pa and back 28 again, each cycle taking just over 3 minutes to complete (Fig.6A). SR image acquisition began when the 29 laser power was increased to a power density of 330 W/cm 2 in high vacuum, and images were acquired

3
During the course of these cycling experiments, there was a noticeable delay in the recovery of GFP 4 blinking when moving between vacuum pressures. When cycling from 10 -3 Pa to 200 Pa, this delay was 5 very short and synchronised with the introduction of water vapour into the SEM chamber. When moving 6 from 10 -3 Pa to atmospheric pressure (relative humidity of around 70%), GFP blinking recovered quickly 7 once the chamber door was opened, after a delay of some 2-3 seconds. This may be attributable to the 8 time taken to rehydrate the fluorophores by diffusion of water into the chamber and section, adding to 9 the evidence that water is essential for GFP blinking. On venting the chamber with dry nitrogen gas,

13
We next investigated how electron contrast affected GFP blinking. We previously reported that IRF 14 sample preparation resulted in a reproducible gradient of electron contrast, from heavy membrane-15 staining at the edge of the cell pellet exposed directly to the uranyl acetate stain, to lighter

27
Finally, we imaged dry-mounted 200 nm IRF sections containing YFP-A3 vaccinia using an N-STORM 28 microscope system, to compare the performance of a commercial SR microscope to the SECOM 29 platform (SFig.3, SMovies 7, 8 and 9). We found that the frequency of FP blinking was similar during 30 the initial phase of data collection (SFig.3 frames 1 and 50, SMovies 7, 8 and 9). However, FP blinking 31 frequency reduced more quickly in the N-STORM dataset when compared to the SECOM at 200 Pa 32 (SFig.3 frames 200 and 2,000, SMovies 7 and 9). This is illustrated by maximum intensity projections 33 from early and late points in the data collection (frames 300-500 vs. frames 2,100-2,300). The signal to noise ratio also worsened in the later stages of data collection at atmospheric pressure, further dataset recorded at 200 Pa contained ~2.27x more localisations than were detected using N-STORM.

1
The results at atmospheric pressure were more comparable, with ~1.46x more localisations detected in 2 the SECOM dataset. Although the absolute number of individual molecule localisations cannot be 3 directly compared because of differences in the sections imaged and the microscope systems, these 4 findings provide further evidence that FP blinking in IRF sections can be driven more effectively at 200 5 Pa.

16
This workflow delivers several advances in super-resolution imaging. The overlaid SR-EM images provide 17 a new 'ground truth' method for testing super-resolution reconstruction algorithms and parameters.

18
For example, we determined that closely opposed virus particles sometimes appear as elongated, 19 curved structures rather than punctate point sources in SR images, information that can be used to 20 improve reconstruction software to remove sample-specific errors. Furthermore, we propose that serial 21 IRF sections could be imaged using a standard fluorescence microscope equipped with a vacuum stage 22 (as commonly used in materials research and available off-the-shelf) and lasers, enabling multi-colour 23 3-dimensional SR imaging using standard FPs without the need for multiple buffers tailored to each 24 fluorophore. In this case, sample preparation would be simplified as electron contrast would be 25 unnecessary, and the blinking activity of FPs would likely be improved in the absence of metal staining.

26
Automation of this process could lead to the generation of 3D SR molecular localisations through large 27 volumes of cells and tissues.

28
The workflow also delivers advances in correlative microscopy. In post-embedding CLEM experiments 29 on thin sections, it is necessary to treat the section for each imaging modality. In practice, this usually 30 means wet mounting of the sections for light microscopy, followed by rinsing and additional contrasting 31 for electron microscopy. Each of these steps can potentially alter the specimen, causing section 32 distortion or shrinkage, or subtle changes in molecular composition. The physical transfer step between 33 imaging modalities also makes the correlation of data significantly more challenging, particularly in 34 cases where the relationship between protein location and structure is unknown. Collecting sections 35 directly on a rigid conductive coverslip and imaging dry within the partial vacuum of the ILSEM means 1 that it is possible to avoid these complicating factors. The high precision with which SR and EM images 2 can be correlated when using the ILSEM is therefore a major strength of our approach. It also 3 eliminates many of the complexities of pre-embedding CLEM image alignment. Furthermore, the 4 workflow has great potential for multi-colour molecular localisation through large cell and tissue tomography'.

7
This study also revealed some unexpected properties of resin-embedded FPs. First, YFP and GFP display 8 strong stable blinking in vacuo when embedded in resin after freeze substitution with a small 9 percentage of water. Second, YFP and GFP blink better in resin at 200 Pa partial pressure of water vapour than they do at either atmospheric pressure, or at 10 -3 Pa. This was unexpected because the 11 intensity of YFP and GFP in widefield imaging mode was higher at atmospheric pressure than at 200 Pa     Fixed mounted whole cells containing either GFP-C1 or YFP-A3 were cultured directly on glass 1 coverslips, fixed as described above, and mounted using Citifluor AF4.

26
For SEM imaging, the system was pumped to high vacuum (~10 -3 Pa). The vCD backscatter detector (FEI 27 Company, Eindhoven) was used at a working distance of 7.3 mm, and inverted contrast images were 28 acquired (2.5 keV, spot size 3.5, 30 µm aperture, and pixel dwell time of 60 µs for a 3,072*1,536 image  2 Several freely available reconstruction packages were tested during preliminary super-resolution data 3 acquisition in order to determine which worked better with the data generated by the ILSEM-based

19
Reconstructions were generated using averaged shifted histograms at a magnification of 10x (for low

12
The SR reconstruction represents ~25,000 localisations. C: An area matching that shown in A and B 13 from the next section in the series, acquired with the chamber pumped to high vacuum (10 -3 Pa). The  figure 4; cropped areas shown in figure 5). The first 500 frames were excluded from analysis as the 9 EMCCD was close to saturation and FP blinking was difficult to observe (indicated by vertical line).

12
Supplementary movie 8 -YFP blinking in an IRF section at atmospheric pressure acquired using the 13 SECOM platform.
14 Supplementary movie 9 -YFP blinking in an IRF section at atmospheric pressure acquired using a Nikon 15 N-STORM microscope system.