Development And Application Of Theoretical Methods For Protein Directed Design

Directed design of proteins is of great value for many fields such as enzyme engineering,drug design,protein folding and antibody design.This paper focuses on the following research:1.Efficient prediction of the contribution of single point mutation to protein thermostability was achieved by developing the vd WP method.Through molecular dynamics simulation of the crystal structures of the wild-type proteins,mutation and structural optimization were applied on the wild-type structures.27 protein systems and 853 single point mutation data were tested and the result of correlation from vd WP is comparable to that of FEP.At the same time,unlike the FEP method,which requires a lot of computational cost,the vd WP method can reduce the cost of predicting the free energy change before and after single point mutation greatly.And vd WP can play an important role in the saturation mutation design process in the early stage of protein design.2.Due to the large error in the calculation of the solvation free energy in the development of the vd WP method,we developed the solvation free energy decomposition method named IESR in the explicit water model.The IESR method enables accurate calculation of the free energy contribution of each component in the solvation free energy.The contributions of solute-solvent interaction free energy and solvent reorganization free energy can be calculated separately from short-term MD simulations of common small molecules.The comparison between the solvation free energy,enthalpy and entropy with the experimental results shows that IESR method is accurate for the prediction of the contribution of each part of the solvation free energy.The IESR method will help to accurately calculate the solvation free energy and the contribution of each part of solvation in the explicit water model.And IESR will also provide effective information for improving the hydrophilicity of small molecules in the design of drug molecules.3.Due to the large error in the calculation of electrostatic interaction energy was found in the development of the vd WP method.We focus on developing deep learning models for predicting the polarized charges of proteins.By modifying the code of the open-source software package named Deepmd-kit,we introduced a descriptor considering long-range interactions and verified the validity of this long-range descriptor on the prediction of polarized charges.With the addition of long-range descriptors,Deepmd-kit model could predict the polarized charges and electrostatic potentials of proteins accurately.Efficient and accurate prediction of polarized charges facilitates the analysis of polarizing effects and long-range interaction contributions for protein directed design.4.Since protein design based on the method of MD requires a certain computational cost,we also developed the machine learning model of protein design for the prediction of protein-protein interactions.The Se BPPI model was firstly developed with the sequences as input for qualitative prediction of protein-protein binding.By combining the existing pretrained models of proteins and the recurrent convolutional neural networks,Se BPPI outperforms the existing machine learning methods on 5independent test sets.We have also integrated the Se BPPI model on web server named icdrug for academic users: http://www.icdrug.com/ICDrug/Se BPPI.5.With the Se BPPI model,one could predict the probability of protein-protein binding qualitatively,and accurate prediction of protein-protein binding free energy is also very important.Therefore,Deepbind,a neural network model was developed for accurate prediction of the contribution of single point mutation to the protein-protein binding free energy.Validation on 5 independent test sets showed that the Deepbind model outperforms existing machine learning models in predicting protein-protein binding free energy.The model is also integrated in the web server named icdrug for academic users: http://www.icdrug.com/ICDrug/XBPPI.The above research provides some useful new computational tools and analytical methods for the research of directed design of protein,and also lays the foundation for further development on directed design methods of protein.