Advanced machine learning method for prediction of protein secondary structure

Abstract

A new method based on Markov process to encode the protein sequences has been introduced. With this simple method, input vectors that contain the essential features of protein sequence can be extracted and efficiently used to train SVM classifiers. Our method achieved the remarkable result that out-performs other advanced methods at present. Using a seven-folded cross validation on the data set of 513 non-homologous protein chains (CB513), the SVM together with Markov transition matrix encoding scheme produces a three-state overall per-residue accuracy(Q3) of 82.49 percent and a segment overlap accuracy(SOV) of 77.18 percent. That is the next improving step to reach the theoretical limitation.