An electron counting algorithm improves imaging of proteins with low-acceleration-voltage cryo-electron microscope

Relative to the 300-kV accelerating field, electrons accelerated under lower voltages are potentially scattered more strongly. Lowering the accelerate voltage has been suggested to enhance the signal-to-noise ratio (SNR) of cryo-electron microscopy (cryo-EM) images of small-molecular-weight proteins (<100 kD). However, the detection efficient of current Direct Detection Devices (DDDs) and temporal coherence of cryo-EM decrease at lower voltage, leading to loss of SNR. Here, we present an electron counting algorithm to improve the detection of low-energy electrons. The counting algorithm increased the SNR of 120-kV and 200-kV cryo-EM image from a Falcon III camera by 8%, 20% at half the Nyquist frequency and 21%, 80% at Nyquist frequency, respectively, resulting in a considerable improvement in resolution of 3D reconstructions. Our results indicate that with further improved temporal coherence and a dedicated designed camera, a 120-kV cryo-electron microscope has potential to match the 300-kV microscope at imaging small proteins.

a. Fourier shell coefficient (FSC) curves of clusters by their sizes at 120 kV. Clusters from "small/medium/large-sized" are displayed in the same colour of black, blue and red. MCF algorithm was used to filter each cluster, and alignment was the same. Each reconstruction contains roughly the same number of clusters. b. Fourier shell coefficient (FSC) curves of "small-sized" clusters in Coxsackievirus A10 dataset at 120 kV with MCF, PVF and GCF counting algorithm. Each pair of reconstructions holds roughly the same number of clusters (~81.4% of all "small-sized" clusters). The resolution of each counting algorithm mark on the graph. c. FSC curves of "mediumsized" clusters in Coxsackievirus A10 dataset at 120 kV with MCF, PVF and GCF counting algorithm. Each pair of reconstructions holds roughly the same number of clusters (~71.3% of all "medium-sized" clusters). The resolution of each counting algorithm mark on the graph. d. FSC curves of "large-sized" clusters in Coxsackievirus A10 dataset at 120 kV with MCF, PVF and GCF counting algorithm.
Each pair of reconstructions holds roughly the same number of clusters (100% of all "large-sized" clusters). The resolution of each counting algorithm mark on the graph. e. FSC curves of 3 types of clusters in Coxsackievirus A10 dataset at 200kV with MCF and PVF counting algorithm. Each pair of reconstructions holds roughly the same number of clusters (18.3% of "small-sized" clusters, 51.2% of "medium-sized" clusters and all "large-sized" clusters). The resolution of PVF counting mark on the graph. f. FSC curves of 3 types of clusters in apo-ferritin dataset at 300 kV with MCF and PVF counting algorithm. Each pair of reconstructions holds roughly the same number of clusters (12.5% of "small-sized" clusters, 24.8% of "medium-sized" clusters and all "large-sized" clusters). The resolution of PVF counting mark on the graph. g. SNR ratio between "large-sized" clusters of PVF and "small-sized" clusters of MCF at 120 kV. The dataset is Coxsackievirus A10. No alignment was changed a. Shape of SNR-weighted PVF for "large-sized" clusters. The SNR-weight was applied in Fourier space, therefore some negative value was on "Weighted PVF" graph which shown in real space. b. FSC curves of "pow" parameter testing on all clusters at 120 kV. Every cluster was used for reconstruction. c. FSC curves of "pow" parameter testing on "small-sized" clusters at 120 kV. Every cluster was used for reconstruction. d. FSC curves of "pow" parameter testing on "medium-sized" clusters at 120 kV. Every cluster was used for reconstruction. e. FSC curves of "pow" parameter testing on "large-sized" clusters at 120 kV. Every cluster was used for reconstruction. f. Comparison of SNR for "large-sized" cluster between "pow=0" and "1". The dataset is Coxsackievirus. g. FSC curves of "pow" parameter testing on "large-sized" clusters at 200kV. Every cluster was used for reconstruction. The dataset is Coxsackievirus A10. h. FSC curves of "pow" parameter testing on "large-sized" clusters at 300 kV. Every cluster was used for reconstruction. The dataset is apoferritin. curves of "small/medium/large-sized" clusters from PCA method. Three curves were normalized so that their intensity at near zero frequency is 1. The Vertical artefacts on each curve come from the bad channels on Falcon III camera we used. d. SNR and NPS ratio between "large" and "small-sized" clusters in logarithm scale at 120 kV.

Supplementary
PCA was used. e. SNR ratio after applying SNR weight to "large-sized" clusters  The reference is 300 kV. X/Y-axis scaling factors were measured according to

Supplementary Note 3. Measuring K3 readouts at three voltages.
Using the fluorescent screen current as a reference, the readouts of K3 was measured at different voltages. In each measurement, the beam was adjusted so that it completely covered the fluorescent screen and the strong Fresnel rings was outside of the screen while keeping parallel illumination. The SC value was noted, and the dose rate of K3 was measured by DigitalMicrograph with each exposure time of 0.2~0.3s, 5 times each. The results were linear-fitted and shown on Fig. 1a Where 1 , 2 is signal of F1, F2 and 1 , 2 is the noise. We use The star-file of Hybrid counting method after refinement was symmetry-expanded from O-symmetry to C1 by relion_particle_symmetry_expand. We then used relion_particle_subtraction program to subtract the expanded star-file with masks.
Each mask contains 3/4 to 1/8 of total mass and low-passed to 15 Å ( Supplementary   Figure 11 a). "--ignore_class --data expanded_data.star" was used for subtraction to force subtracting to given expanded star-file. After subtracting Hybrid images, particle images from MCF replaced Hybrid images and the same processes were done one again. Since the subtraction is not perfect and many positive or negative features remained, new masks low-passed to 15 Å were used during Auto-refinement after iteration 3. The refinements were local, starting with angle steps of 1.8 degree. After refinement, SNR was calculated and compared followed latter part of Supplementary Note 7. We found the refined result of 1/8 (56 kD) is dramatically worse, shown on Supplementary Figure 11c, thus we excluded this result.