Learning kernel matrix from gene ontology and annotation data for protein function prediction

During the last few years, Kernel methods have gained considerable attention for analyzing biological data for protein function prediction. Based on biological processes annotation of Yeast and GO(gene ontology), we constructed a kernel matrix to predict protein functions. We used measurement method about semantic similarity on GO and adaptive Hausdorff distance to successfully obtain protein similarity matrix, and furthermore, transformed protein similarity matrix to a undirected graph. Then, We developed a novel method that can learn optimal diffusion kernel from graph by maximizing kernel-target alignment. Experimental results illustrate that the kernel matrix generated by our formula has larger AUC value than ordinary diffusion kernel and those proposed before. Our method can even learn a common optimal kernel matrix for multiple predict tasks at one run. Furthermore, it can also be directly used to learn from various biolobical networks.