Prediction Method Of Microbial Association Based On Time Series Data

Microorganisms play an important role in the ecological environment and human health.It is vital to explore the complex relationship between human and microbe as well as between microbe and microbe,and the inference of microbe association network through microbiome time series data is an pivotal link to explore the complex relationship between microbe and human disease.With the development of high-throughput sequencing technology,a large number of high-dimensional time series data describing the dynamic process of microorganisms have been produced.The construction of microbial interaction network from time series data can capture the dynamic relationship between different microbes or between microbes and environment,which can provide a more real performance for the interaction of microbes in nature.Based on the maximum entropy model and vector auto-regressive model,this thesis studies the dynamic interaction network based on microbial time series data.The main work is as follows:Firstly,a nonparametric model based on Chao-Shen entropy is proposed to infer microbial interactions.Based on the joint entropy formula,the conditional entropy was obtained by linear optimization of the joint entropy,and the pairwise association between microorganisms was deduced.In order to estimate entropy in high-dimensional data,Chao-Shen entropy estimation method is introduced into maximum entropy model.Compared with other entropy estimation methods,the inference performance of the maximum entropy model combined with Chao-Shen algorithm is verified in the time series sample data of intestinal microorganism disturbed by antibiotics.Second,a vector autoregressive graph regularization method using the global and local information of graphs is proposed to infer the interaction network between microorganisms.The regularized vector autoregressive model based on the Laplacian matrix only considers the global information of graphs,but ignores the local information of graphs.In this paper,a regularization method,LG-VAR model,which combines Laplacian matrix and vicus matrix,is introduced.Laplacian matrix can capture the global structure of the network,while vicus matrix can capture the local structure of the network.The regularization method combining the two matrices more accurately simulates the dynamic interaction network between microorganisms.Compared with several existing regularization methods such as Lasso,elastic net and Laplacian,the results show that the LG-VAR model performs better.The two improved models are suitable for high-dimensional microbial data and they are parametric model and parametric model respectively.Experimental results show that these two models have certain value and significance in the study of microbial dynamic correlation,and provide some new ideas for inferring dynamic interactive networks based on microbial time series data.