Two-Stage Support Vector Machines for Protein Secondary Structure Prediction

摘要

Neural network approaches using Multi-Layer Perceptrons (MLPs) and Support Vector Machines (SVMs) for protein secondary structure prediction are presented. A two-stage SVM approach is proposed to capture the contextual relationship of secondary structure elements. The proposed technique yielded higher accuracy than the PHD (Profile network from HeiDelberg) method that cascades two MLPs. We also demonstrate that it is feasible to improve current single-stage approaches to protein secondary structure prediction by adding a second-stage prediction scheme to capture the contextual information among secondary structural elements and thereby improving the accuracy of prediction. Two-stage SVM approach achieved prediction accuracies of 72.4% and 76.7% on two databases of 126 and 513 nonhomologous globular proteins, respectively.