Research And Application Of Secondary Structure Detection From Cryo-EM Images

The protein structure with atomic resolution was resolved from cryo-EM maps,so it can make the protein structure stable.If the secondary structure can be identified from the cryo-EM maps,not only the detailed information of the electron density map can be further extracted,but also the structural stability of the protein can be further improved.At first,the segmentation of secondary structure from cryo-EM maps mainly relied on human subjectively defined features,which limited its use.Later,the features were automatically extracted through machine learning models,which effectively improved the accuracy of secondary structure detection.There are a lot of noise in cryo-EM maps and the information contained in the secondary structure was different at different resolutions,so it is challenging to detect the secondary structure from cryo-EM maps.Therefore,the 3D Unet model can be constructed to automatically extract low-level spatial features and high-level abstract features from cryo-EM maps,then these features was merged to predict the secondary structure of each voxel in the cryo-EM map.The dice?score of the model were more than 70% in various types of secondary structure,especially the alpha helix structure was about 77%.In order to use large-scale data for training,it is necessary to search the appropriate protein structure from the PDB database,and then the EMAN2 tool was used to simulate the electron density map as the training data.In addition,compared with the 3D CNN model and the 3D Deep Lab V3 model,it was7.49% higher than them on average in the secondary structure of the simulated testing dataset.Finally,in order to further explore its application,the identified secondary structure was divided into secondary structure fragments by the algorithm of spatial search,then the structure optimization of the protein was further researched.The experimental shows that 90% of the protein structure accuracy has been significantly improved in 20 protein structures.Under the three-dimensional Unet image segmentation framework,the secondary structure elements were extracted effectively from cryo-EM maps,which can further explore protein structure optimization from cryo-EM maps.Finally,it can also be applied to de novo protein structure prediction construction.