Font Size: a A A

Protein Structure Evaluation Based On 3DCNN

Posted on:2020-04-30Degree:MasterType:Thesis
Country:ChinaCandidate:L DengFull Text:PDF
GTID:2370330575977354Subject:Computer technology
Abstract/Summary:PDF Full Text Request
The core of protein biology is to understand the mapping between structure and function.As the protein structure data increases,the mapping of protein spatial structure to functional relationship can be obtained by some calculation methods.However,the effect of these calculation methods is mainly determined by the quality of the protein structure.Most existing calculation methods use existing structural and functional relationships.Such methods are better in most cases,but we cannot get an accurate function for a specific location structure.As the demand for protein structure data grows,the work on protein structure evaluation becomes more and more important.In the structure of the protein,the interaction between the amino acids and the atoms will occur,so that in addition to the different amino acid sequences,other forces will lead to a large difference in the spatial structure of the protein.Every atom in the structure is affected by other atoms around it.This part is a tiny environment.We can analyze the structure of the protein through a tiny microenvironment.In this paper,we present a method for applying three-dimensional convolutional neural network(3DCNN)techniques to protein structure analysis.This method automatically extracts specific features from the original distribution of protein structures,driven by supervised tags.As a pilot study,we used our neural network model to analyze the local protein microenvironment around 20 amino acids and predict the most compatible amino acids in the protein structure.By this method we finally evaluate the protein structure.purpose.To further verify the functionality of our method,we used a large number of protein spatial structures and abstracted them into data in accordance with the algorithm for extracting the microenvironment as an input sample for model training.In order to verify the effectiveness of the proposed method,the proposed algorithm and the LDDT algorithm were compared on the 2012 dataset of the CASP competition.The experimental results show that the proposed algorithm evaluates the protein structure score lower than the LDDT algorithm,but the score has a positive correlation with the LDDT algorithm score.Compared to the conventional protein evaluation method,our three-dimensional convolutional neural network scores on protein structure evaluation on average not only to indicate the structure of the protein,but also to successfully summarize the known information about similar and different microenvironments.Conventional protein structure evaluation methods require the true and false structure of the protein as input to calculate the score,which is unevaluable for the predicted protein structure without knowing its true structure,and our model just makes up for it.This is not enough.We use a large amount of protein structure data as a sample to train the three-dimensional convolutional neural network,and adjust the training parameters of the model according to the results.When the model is trained,only the predicted structure can be used as input to obtain a protein structure score.In this paper,the three-dimensional convolutional neural network is trained,and the obtained results are analyzed and compared.We find that our model score has a positive correlation with the GDT score.This conclusion enables us to have no real structural conditions.A better structure can be selected among the many predictive structures.This conclusion also shows that the end-to-end well-trained deep learning network is always superior to the traditional method of evaluating protein structure,indicating that the 3DCNN framework is very suitable for the analysis of protein microenvironment and can be used for protein structure evaluation.
Keywords/Search Tags:Protein structure, Microenvironment, 3DCNN, Structural analysis
PDF Full Text Request
Related items