Using Homology Information From PDB to Improve The Accuracy of Protein β-turn Prediction by NetTurnP*: Using Homology Information From PDB to Improve The Accuracy of Protein β-turn Prediction by NetTurnP*

Abstract

茁-Turn is a secondary protein structure type that is important in protein folding, protein stability and molecular recognition processes. To date, various methods have been put forward to predict 茁-turns, but none of them have tried directly to map the structures of pre-existing homologues from structural databases like RCSB PDB to the protein to be predicted. Given the large size of PDB (>70 000 structures), it is actually of high possibility to find a structural homologue for a newly identified sequence. In this work, we present a new method that predicts 茁-turns by combining homology information extracted from PDB with the results predicted by NetTurnP. Two datasets, the golden set BT426 and the self-constructed dataset EVA937, are used to assess our method. For each sequence in both datasets, only homologues deposited earlier than the sequence in PDB are employed. We have achieved Matthews correlation coefficients (MCCs) of 0.56, 0.52 respectively, which are higher than those obtained by NetTurnP alone of 0.50, 0.46, and the prediction accuracies (Qtotal) obtained using our method are 81.4% and 80.4% separately, while NetTurnP alone achieves 78.2% and 77.3% . The results confirm that combining the homology information with state-of-the-art 茁-turn predictors like NetTurnP can significantly improve the prediction accuracy. A Java program called BTMapping has been written to implement our method, which is freely available at http://www.bio530.weebly.com together with the related datasets.