| Two-dimensional ultraviolet(2DUV)spectroscopy of the protein backbone is a powerful tool for identifying the details of protein secondary structure.However,its practical utilization has so far been limited by the its complex spectral patterns,complicated analysis process and low accuracy.This thesis mainly relies on machine learning,a cutting-edge advanced technology in the field of computer science,to build a structure prediction model and analysis process based on 2DUV spectrum.We calculated the secondary structure spectrum data set of 87993 samples by applying molecular dynamics and molecular mechanics(MD/MM)methods.Then a convolutional neural network(CNN)protocol was used to build a mapping model from2 DUV spectrum mode to secondary structure.The model is being tested on the testing set reached a nearly perfect classification result.And then we tried to explain the classification basis of the classifier.In order to verify the transferability of the model,we also tested the homologous protein data set and the non-homologous protein data set.Combined with the transfer learning technique,the accuracy of the model for secondary structure recognition of non-homologous protein fragments was more than92%.Finally,we compared the CNN model with other commonly used machine learning algorithms on the data set.The results show that the convolutional neural network applied to 2DUV spectrum structure recognition has high accuracy and interpretability.In addition,the model we built also has good migration prediction capabilities.This thesis uses unsupervised learning methods to explore the relationship between different protein chemical environments and the 2DUV spectrum patterns.We found that the α-helix and β-sheet datasets have five similar spectral patterns.Importance analysis shows that the relative distances and orientation between adjacent segments are the decisive features governing 2DUV patterns.Our work proves the great potential of machine learning models in the field of2 DUV spectroscopy,and provides new ideas for subsequent application research of2 DUV spectroscopy. |
