Local structure prediction can facilitate ab initio structure prediction, protein threading, and remote homology detection. However, previous attempts for local structure prediction suffer from poor prediction accuracy. In this paper, we propose a knowledge-based prediction method. We use a measure called match rate to estimate the confidence of our knowledge-based method. Accuracy of this method is shown to be positively correlated with the match rate. We use our knowledge-based method to predict the local structure of proteins with high match rate. For proteins with low match rate, we propose a neural network prediction method. Thus, we have a hybrid prediction method that combines our knowledge-based method with a neural network method.
        Our methods are extensively tested on three different structural alphabets: SAH (Structural Alphabet of HMMSTR), PB (Protein Blocks) and STR. Prediction results are evaluated by two prevailing evaluation criteria, MDA (maximum deviation of backbone torsion angle) and QN, which is similar to Q3 in secondary structure prediction. All of the tests demonstrate significant improvement to the previous studies.