A Protein Secondary Structure Prediction Tool Using Two-Level Strategy to Improve the Prediction Accuracy of Secondary Structures and Structure Boundaries

An important limitation of current protein secondary structure prediction tools is the bad performance in locating the secondary structure boundaries. Efficiently utilize the residue position-specific preference around secondary structure boundaries can help to resolve this problem. TLSSP (two level secondary structure predictor), proposed in this study, used a two-level strategy to utilize these properties efficiently and find the optimal global secondary structure. In TLSSP a set of binary classifiers were designed to recognize the boundaries of helices and strands firstly, then a global model based on condition random fields (CRFs) was built to predict the secondary structures. Five-fold cross-validation test on EVA dataset (containing 3744 proteins provided by EVA service) indicated that, TLSSP can get quite good performance on both boundaries prediction and global secondary structure prediction.