Title: Improving Protein Disorder Prediction by Incorporating Evolutionary Information and Optimizing Knowledge Representation

Talk by Kang Peng

Abstract: A dominant view in molecular biology is that protein functions depend on the protein 3-D structure determined by amino acid sequence. However, it turns out that lots of disordered proteins, or proteins without unique 3-D structure, still carry out important functions. Our previous work on predictions of disorder from sequence information at about 70% position by position out of example accuracy compared to the 50% expected by chance for the balanced datasets, supported the hypothesis that amino acid sequence determines three-dimensional structure as well as lack of three-dimensional structure. Recently, the prediction accuracy has been boosted to 82.6% by using larger dataset and better knowledge representation. Using the same method, we were able to achieve accuracy of 83.6% on an even larger new dataset in this study. We also propose 2 new methods designed to incorporate evolutionary information. The first one directly uses homologous sequence segments of the true disordered regions in training disorder predictors. The second one builds disordered predictors from family profiles built by PSI-BLAST. Both methods achieve improved prediction accuracy of 85%