The team of Professor Bi Guoqiang of the University of Science and Technology of China/Shenzhen Advanced Institute of the Chinese Academy of Sciences has made an important breakthrough in the development of cryo-EM in situ imaging technology

The team of Bi Guoqiang, a dual-appointed professor from the University of Science and Technology of China and the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences, cooperated with Professor Zhou Zhenghong of the University of California, Los Angeles to develop a set of cryoET data processing algorithm and software IsoNet based on deep learning The problems of missing cone effect and low signal-to-noise ratio in cryoET imaging are effectively solved, and the relevant research results are based on Isotropic reconstruction for electron tomography with deep learning Published in Nature Communications on October 29, 2022

The researchers built a set of iteratively optimized self-supervised deep learning artificial network algorithms, and used the 3D reconstruction data of cryoET tomography after rotation processing as the training set to correct the missing cone of the cryoET tomography 3D reconstruction data
.
At the same time, in the process of IsoNet algorithm, the noise reduction process is added, so that the same artificial neural network can complete the missing information and reduce the noise of the fault three-dimensional reconstruction data at the same time
.

Figure 1.
The basic principle and process of cryoET imaging data missing cone correction and noise reduction based on deep learning

The three-dimensional structures of apoferritin and ribosome simulating missing cones were processed separately using the IsoNet algorithm, and the processed results were comparable to low-resolution atomic models
.
It also treated real HIV capsids, the paraflagellar rods, and synapses in cultured nerve cells synapse) 3D reconstruction data of faults, and obtained very good results
.
It is particularly noteworthy that after using the IsoNet algorithm to process the 3D reconstruction images of cell-level thick samples containing a large number of proteins, membranous organelles and cytoskeletons, the three-dimensional structural information of synaptic vesicles, mitochondria, microtubules, microfilaments, cell membranes and protein complexes has been well recovered
.

Figure 2 The effect of using IsoNet algorithm to process the synaptic cryoET 3D reconstruction data, and the three-dimensional visual rendering of the ultrastructure of the synapse after processing by the IsoNet algorithm based on the real electron density

After the release of the preprint bioRxiv, the IsoNet algorithm has attracted extensive attention and in-depth discussion in the field, one of which is how the IsoNet algorithm achieves missing cone correction.
One of the main speculations is that during the training process, the artificial neural network can learn the structural features of biological structures such as proteins at different angles in three-dimensional space, and supplement this information to the direction of the missing cone, similar to the three-dimensional average of
single-particle cryo-EM.
Therefore, by continuously optimizing the artificial neural network architecture and expanding the training sample set, the IsoNet algorithm will be able to recover the high-resolution three-dimensional structure information of each protein molecule in the cell, thus laying the foundation
for truly visualizing the high-resolution three-dimensional structure and tissue distribution of each protein molecule in situ.
As experts in the field, Dimitry Tegunov and others commented on Twitter, the idea of the IsoNet algorithm will be the future development direction
of cryoET technology.

The first authors of this paper are Liu Yuntao, a doctoral student at the University of Science and Technology of China (now a postdoctoral fellow at the University of California, Los Angeles), and Zhang Heng, an undergraduate student of the class of 2018 (now a graduate student at Peking University), and the corresponding authors are Professor Bi Guoqiang and Professor
Zhou Zhenghong.
Dr.
Tao Changlu, associate researcher of Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, and Wang Hui, a doctoral student at the University of California, Los Angeles, also participated in the work of
this paper.
This work was supported
by the Ministry of Science and Technology, the National Natural Science Foundation of China, and the Chinese Academy of Sciences.

Article link: style="line-height:150%">

(National Research Center for Microscale, Department of Biomedicine, Department of Scientific Research)