Multiple Kernel Learning For Predicting Protein Secondary Structures

With the unprecedented explosive expand of protein sequencedatabase, the technology of predicting the protein secondarystructure based on sequence information began to replace theexperimental methods which are time consuming and laborious.Among these methods, kernel methods have been applied inprotein secondary structure prediction for years. But unfortunately,classifiers with one single kernel can get use of only one kind ofproperty from protein sequence. Although in the field of predictingprotein secondary structure, all the information we get is acid aminosequence, there do exist several features abstracted from thesequence, which contain hidden pattern and relevance with structurein their own ways.For the sake of mixing the features form different platformstogether, a multiple kernel learning framework which mergesseveral protein properties together is performed to predictsecondary structures.We tested multiple kernel learning method on a protein data setof2421chains with less than30%sequence identity, and a comparedresult shows that the method gives more accurate and stableprediction both than single feature kernel and features linked kernel. Therefore, Multiple Kernel Learning method can act as animprovement of the current kernel classifiers to make a stepforward.