Nonlinear Discriminant Analysis Based On Kernel Selection And Its Application In Protein Subcellular Localization

With the completion of the human genome sequencing,the research on life science has entered into the era of Post-genome project.Now,as the quantity of protein sequences increases rapidly,using traditional techniques to predict protein subcellular locations will lead to high cost,time wasting and low resolution.In recent years,the prediction of protein subcellular location based on machine learning has gradually become a hot topic and researches on it include the feature expression of protein sequences,dimension reduction algorithm and the classifier.Now,many methods have been proposed to solve this problem.This thesis mainly researches the kernel discriminant analysis which is a dimension reduction algorithm and its application on protein subcellular localization.Biological data used in protein subcellular location prediction generally have the property of non-linear.So using these linear dimensionality reduction algorithms can't extract these nonlinear characteristics,which will decrease the accuracy.In order to solve this problem,the researchers proposed nonlinear reductive dimension algorithms that combine kernel tricks with traditional linear reductive dimension algorithms.In this thesis,kernel discriminant analysis(KDA)is used to predict the protein subcellular locations.Kernel parameters will have a great influence on the performance of the kernel discriminant analysis.There are some traditional algorithms being usually used to choose the appropriate parameter,such as the genetic algorithm,the grid search method.However,these methods require a large amount of calculation.In this thesis,we propose a new method based on reconstruction error and use it to select the appropriate kernel parameter.Compared with these traditional algorithms,this algorithm will spend less time on the selection of kernel parameters and this is an innovation for this thesis.Now,studies on kernel discriminant analysis(KDA)are mainly based on one kind of kernel function,especially in prediction of protein sub-cellular location.However,single kernel function can't extract the data feature allsidedly,which leads to a limitation of kernel linear discriminant analysis based on single kernel function in sub-cellular localization.Combined kernel function has characteristics from all the components,which can improve the accuracy of the algorithm.So the KDA with combined kernel function is used to predict the protein sub-cellular locations in this thesis and is proved to be more efficient than the KDA with single kernel function.This is another innovation for this thesis.