New Sparse Methods For Metabonomic Data Analysis

Metabolomics is a new interdisciplinary subject based on metabolite analysis,through high-throughput detection and data analysis,and aiming at information modeling and systematic integration.It has been one of the most hot-spots in life science research.However,due to high dimensional data it acquired and multiple technological platform it involved,the data analysis of metabolomics meet great challenges.One of the most effect methods to solve this problem is data sparsity and this thesis focus on some of the issue in sparse data analysis of metabolomics.The main contents are as follows:Firstly,a sparse partial least squares algorithm based on sure independence screening(SIS)method is proposed.In this method,sparse regression is achieved by impose SIS procedure in solving the direction vector of ordinary PLS method.Theoretic studies showed that this method enjoyed good asymptotic properties and numeric studies from both simulated dataset and real dataset indicated the proposed algorithm could obtain a relative sparse model with rather good performance.Secondly,A sparsity method of multi-source data analysis based on complemented variables is proposed.In this method,significant correlated and complemented variables selection is achieved via a G2-test and a extended log-likelihood ratio test combining with multiple test respectively.Numeric studies and real multi-source dataset is used to evaluate the validation of proposed method.Thirdly,a software aiming at quantitative analysis of nuclear magnetic resonance spectroscopy is implemented for the sparsity of spectroscopy.This software with a friendly interface is convenient at its operation.The procedures and functions of the software was illustrated through an example of processing of serum samples.