Improving Continuous-Rotation Tomography of Cellular Specimens

This is the authors' abstract. We don't add key points for this paper.

Cryo electron tomography (cryo-ET) has developed into a powerful technique for visualizing the near-native ultrastructure of cells and tissues [1]. The demand for structural information within its native biological context is driving the rapid growth of cryo-ET. However, cryo-ET often requires large data sets, due to the intrinsic heterogeneity of cellular specimens and to the radiation-damage-limited signal-to-noise ratio (SNR) of individual tomograms. Since each dataset typically requires tens of minutes to acquire, there is an urgent need to develop faster acquisition schemes to maximize throughput and enable larger-scale experiments. Continuous-rotation acquisition, similar to the method of acquisition in micro-crystal electron diffraction (microED), has the potential to dramatically boost throughput while also potentially improving data quality [2, 3]. However, early attempts at continuous-rotation tomography have shown limited reconstruction quality compared to conventional tilt-series acquisition [3]. Here, we have used an ultra-fast electron counting camera to acquire continuous-rotation cellular tomography data of cortical neurons. Using this system, we have developed and investigated a range of acquisition and processing techniques to improve reconstruction quality to match or exceed that of conventional tomography. Cortical neurons were surgically isolated from 8-10 pooled E18 Rattus Norvegicus embryos and mechanically digested via trituration. Cells were seeded in 35 mm2 glass bottom plates coated with 0.1 mg/mL of poly-D-lysine and incubated at 37 °C and 5% CO2 for 6 or 20 days. Prior to seeding and coating, 4-5 Quantifoil R2/2 gold 200 mesh grids were placed in each dish and UV irradiated for 30 minutes to sterilize. Neurobasal media supplemented with 2 mM L-glutamine, 1× B-27, 100 U/mL penicillin, and 100 mg/mL streptomycin was used for culturing, and one third of media was replaced with fresh media every 3-4 days. Neurons were subsequently frozen in liquid ethane cooled with liquid nitrogen at -190 °C and 65% humidity using the Leica EM GP2 automatic plunge freezer. Grids were back blotted for 15 seconds prior to plunge freezing. Continuous tilt tomograms were collected on a Thermo Fisher Scientific Glacios TEM (200 kV) using a Direct Electron Apollo camera, which has a novel event counting direct detection sensor, which was necessary to maximize SNR and maintain linearity across a wide range of exposure rates [4]. Continuous rotation data was collected using a script in SerialEM [5] using a variety of imaging conditions: 11,000× or 45,000× magnification, resulting in 8k×8k super-resolution frames with 2.90 or 1.45 Å/pixel, respectively; defocus of -10 μm at 0° tilt; as a single movie with unidirectional rotation from -50° to +50° or from -60° to +60°, or as two bidirectional movies with -20° to +60° in the first movie and -20° to -60° in the second movie; with a stage rotation rate of either 1°/s or 2°/s. The total exposure ranged from 60 to 120 e-/Å2. The Apollo camera outputs dose-fractionated super-resolution frames at 60 frames per second (fps) with no dead time between frames. The total frames acquired and saved for the acquired data sets ranged from 2990 to 7176, acquired in 50 to 120 seconds. Filling fractions of each of the movies were assessed, and the images were downsampled to physical Nyquist, losslessly compressed, and transferred from local disk to a high-end Linux workstation. Approximately 500 frames from each data set were copied to separate files and used to create a gain reference image [6]. All frames were then gain-corrected with this reference image. Simple cross-correlation between successive single frames did not yield unambiguous alignments of the frames. We therefore calculated the cross-correlation between each frame and the running average of frames surrounding that frame, resulting in unambiguous alignments with reasonable shifts between adjacent frames. After alignment, sums of multiple consecutive frames were generated and used with the standard tomographic reconstruction procedure in EMAN2 [7]. Results were assessed by visual inspection of the cellular features in the 3D reconstruction, and by comparing the power spectra of averages of the aligned stacks of frames to unaligned frames. The aligned versions show enhancement of signal at intermediate frequencies. A reconstruction algorithm able to reconstruct directly from the aligned stacks without calculating sums of consecutive frames is currently underway [8]. Continuous-rotation tomography data of a cortical neuron, acquired at 45K magnification (2.9 Å/physical pixel), -10 μm defocus, unidirectional from -50˚ to +50˚, acquired in 62 seconds, with a total exposure of 79 e-/Å2. (a) A single summed image of 30 consecutive frames from the data set, representing a tilt range of ∼1.3˚ and ∼1 e-/Å2. (b) A central slice from the resulting 3D tomographic reconstruction. Scale bars each represent approximate 340nm