Cluster-based local modeling approach to protein secondary structure prediction

Protein secondary structure can be used to help determine the nanotechnologically relevant tertiary structure of a protein molecule via the fold recognition method. Starting from the primary amino acid sequence, protein secondary structure prediction (PSSP) has been widely studied using a large variety of algorithms. These include support vector machine (SVM) which has been successfully applied to many prediction problems, also PSSP. In this paper, we attack the PSSP problem from another perspective by using local modeling based on clustering. Most previous PSSP solutions improve the prediction accuracy by using more informative encoding schema, better prediction algorithms, and possibly finer methodology such as dual-layer classifiers or consensus voting mechanism. These approaches all adopted a global modeling technique to build a single classifier. Based on the successful applications of local modeling in many fields, we propose a hybrid approach to solve the PSSP problem by preprocessing the protein sequences with a genetic algorithm based clustering before building an individual SVM model for each cluster. Extensive analysis of several datasets of protein sequences and using statistical hypothesis testing, it seems preferable to cluster the sequence data before a classification step is performed in the PSSP problem. The improved prediction of protein secondary structure is important for advanced nanotechnology applications, like biomolecular machines using proteins as their components.