Homogeneous multifocal excitation for high-throughput super-resolution imaging

Abstract

19Super-resolution microscopies, which allow features below the diffraction limit to be 20 resolved, have become an established tool in biological research. However, imaging 21 throughput remains a major bottleneck in using them for quantitative biology, which requires 22 large datasets to overcome the noise of the imaging itself and to capture the variability 23 inherent to biological processes. Here, we develop a multi-focal flat illumination for field 24 independent imaging (mfFIFI) module, and integrate it into an instant structured illumination 25 microscope (iSIM). Our instrument extends the field of view (FOV) to >100x100 µm 2 without 26 compromising image quality, and maintains high-speed (100 Hz), multi-color, volumetric 27 imaging at double the diffraction-limited resolution. We further extend the effective FOV by 28 stitching multiple adjacent images together to perform fast live-cell super-resolution imaging 29 of dozens of cells. Finally, we combine our flat-fielded iSIM setup with ultrastructure 30 expansion microscopy (U-ExM) to collect 3D images of hundreds of centrioles in human 31 cells, as well as of thousands of purified Chlamydomonas reinhardtii centrioles per hour at 32 an effective resolution of ~35 nm. We apply classification and particle averaging to these 33 large datasets, allowing us to map the 3D organization of post-translational modifications of 34 centriolar microtubules, revealing differences in their coverage and positioning. 35 (STED) 5 . Initial implementations of these methods were relatively slow, in part due to the 42 use of the time domain to separate single molecules (SMLM), the need to collect multiple 43 images to cover Fourier space (SIM), or the scanning of a reduced excitation volume 44 3 (STED). In both wide-field and patterned illumination, the imaging throughput -defined as 45 the area imaged per time -can be increased by parallelizing the acquisition, as long as the 46 necessary illumination can be maintained over a larger surface in the sample plane. 47Extending the array of patterned excitation has been used to increase the speed or FOV of 48 confocal 6 , STED 7,8 and SIM imaging 9 , but ensuring the quality of the extended pattern 49 across a larger FOV remains a limiting factor. For SMLM, flat-fielding approaches made it 50 possible to extend super-resolution imaging over ~100x100 µm 2 FOVs 10,11 , but their transfer 51 to patterned illumination remains limited. 52An entirely different approach to effectively achieve super-resolution modifies 53 sample preparation rather than image acquisition. By physically increasing the size of the 54 specimen in an isotropic manner via expansion microscopy 12 (ExM), overall imaging 55 throughput can be increased by opting for faster diffraction-limited microscopes 12-14 . 56 Alternatively, combining ExM with existing super-resolution microscopies allows even 57 further improvement in resolution [14][15][16][17][18] , since the resolvable scale is reduced by the 58 expansion factor. However, expansion exacerbates limita...

3D Structure from 2D Microscopy images using Deep Learning

“…The consensus elongated cylindrical structure produced by our method is also consistent with the known structure of glutamylated tubulin in centrioles [19] (See Supplementary Data: Figure S2).…”

“…The second data-set is derived from expansion microscopy experiments to image labelled glutamylated tubulin in centrioles purified from Chlamydomoanas reinhardtii [19]. The images are segmented and presented as tiff stacks of size 128x128x84 in xyz.…”

“…To validate our method, we also applied it to a separate experimental dataset [19] obtained using a different imaging technique. We analysed SIM / expansion microscopy images of glutamylated tubulin in purified centrioles.…”

Data-driven microscopy allows for automated targeted acquisition of relevant data with higher fidelity

“…With automated image analysis pipelines, HCS enables researchers to extract rich and unbiased information from datasets that would otherwise overwhelm any human operator (17, 18, 19, 20, 21, 22, 23). In other settings where high-spatiotemporal information is necessary and can be acquired, HCS has successfully been integrated (24, 25). The alternative to HCS in high-spatiotemporal settings, acquiring data from selected points, typically leads to the loss of population context and risk of bias, especially since data selection is often left to human operators.…”