The Research On The Modeling Of Gene Regulation Network Based On Multiple Models

In the era of precision medicine,diagnostics and medical services based on genomics are becoming more and more widely used in clinical practice with the significant decrease in cost and the popularization and application of sequencing technology,including prenatal examination,single gene genetic disease examination,oncology individualized treatment,susceptibility gene examination,whole group sequencing and so on.Nowadays,we can easily use the advanced high-throughput sequencing technology to sequenced and analyze 3 billion bases in the individual genome.According to the latest scientific research progress and the unique genetic information,it provides a health management plan for the whole life cycle,such as personalized health advice,reproductive guidance,medication tips,dietary advice,nutrition guidance and so on.The gene regulation network controls the activity of the cell by regulating the expression of the gene,and then regulates the life activities of the organism.The understanding of the inherent laws of the regulatory network of gene expression helps to understand the nature of life activities and the inherent causes of the disease,also helps people to better treat diseases and maintain a healthy life.A variety of mathematical models have been used to construct gene regulatory networks and have achieved some results since the 60 s of last century.The widely used models include Boolean networks,mutual information models,Bayesian networks,differential equations,neural networks and so on.However,some shortcomings have been found in the process of use,such as the existing models are directly used to build a gene regulatory network which cannot accurately predict the expression level of the gene and identify the relationship between genes,and it is often necessary to adjust the specific network to the specific network.In addition,the application scope of a single model is narrow and the robustness is not strong.Multi model fusion construction of gene control network can make up the advantages of a single model,make the model have better applicability and robustness,and achieve higher prediction accuracy.Two major contents in the research of gene regulation network modeling are accurately predicting the level of gene expression and accurately identifying the regulatory relationship between genes.Based on the two main tasks in gene regulatory network modeling and a systematic study of basic gene regulatory network models,this paper under the guidance of multi model fusion theory separately puts forward the Ensemble of Flexible Neural Tree and Ordinary Differential Equations for Inferring Gene Regulatory Networks and the Learning Bayesian Networks Structure based Part Mutual Information for Reconstructing Gene Regulatory Networks.The integration model based on ordinary differential equation and flexible neural tree has high accuracy and is suitable for simulating complex nonlinear relationship and forecasting nonlinear time series.But the relationship between variables can only be indirectly deduced through the prediction accuracy of data because of the model's complex internal structure,so this paper mainly focuses on using the model to improve the prediction accuracy of gene expression level.The Bayesian network model based on partial mutual information can easily identify the direct regulation relationship and indirect regulation relationship between genes because of its application to the relationship between variables.It can easily identify the causal relationship between genes and overcome the false positive problems of the mutual information model and the false negative problems of the conditional mutual information model in the construction of the gene regulatory network with good intuition,so it lays particular emphasis on improving the recognition accuracy of the relationship between genes.The two models proposed in this paper improve the accuracy of gene regulatory network reconstruction from different aspects respectively.The gene regulation network based on the ordinary differential equation and the flexible neural tree integration model separately uses the probabilistic incremental program evolution and particle swarm optimization algorithm to train the structure and parameters of the two sub models.Then the output of the two sub models that are well trained is integrated with weighted average.The experimental results show that the predictive accuracy of the model to gene expression level is higher than that of previous studies with 30%-40%.The Bayesian network model based on partial mutual information first uses partial mutual theory to get a more accurate undirected initial network,then uses the K2 search algorithm based on BDE scoring function to train the final gene regulation network.The experimental results show that this method improves the efficiency of Bayesian network training.Compared with some other models,the false positive of the proposed model is within acceptable range,and the true positive and accurate rate has a good promotion.