Protein Secondary Structure Prediction with Hydrophobicity and Hydrophobic Moment

Protein secondary structure prediction has been satisfactorily performed by machine learning techniques such as support vector machines (SVM’s). We discuss a special technique to include hyrophobicity information to further improve the classification results. Hydrophobicity or hydrophobic moment measure of each amino acid is included within a given window length in the protein secondary structure prediction using support vector machines. The input data is divided into two groups, which is subsequently classified by an SVM. By including hydrophobicity or hydrophobic moment, the classification accuracy is increased. Comparing the accuracy between using 1 SVM and 2 SVMs. 2 SVMs method has 3-9% higher accuracy than 1 SVM method. INTRODUCTION Protein structure prediction is a research topic of growing interest. The number of protein sequences deposited in the Protein Data Bank (PDB) grows faster than the numberof known protein structures. It is very time-consuming to crystallize each protein and use X-ray or nuclear magnetic resonance to analyze its structure. Higher accuracy in secondary structure prediction by using machine learning techniques may help predict tertiary structure more precisely. Such research is usually initiated by using basic local alignment search tool (BLAST) or position-specific iterated BLAST (PSIBLAST) to find similar homologous proteins. Then, the sequences are aligned, and the position-specific scoring matrix (PSSM) is calculated. This is typically followed by a machine learning algorithm such as a neural network (NN) or a support vector machine (SVM) to do secondary structure prediction. It is known that discriminating features may help improve classification accuracy. In this paper, we investigate how best to include new features to improve secondary structure prediction. We focus on several membrane proteins which have special distributions with respect to input features. In a previous paper on secondary structure prediction with support vector machines (Ibrikci et al., 2005), the protein structure was classified into 8 classes. In another paper (Rost et al., 2003), the protein structure was classified into 3 classes as alphahelix, beta sheet and coil. In this paper, we first classify the data into 2 classes, such as alpha-helix and non-alpha-helix. Later by combining the results of three binary classifiers, the three-class classifier is obtained. The window size was