Predicting protein residue-residue contacts and disorder
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Predicting a protein's three dimensional structure from its corresponding sequence has long been an extremely important and challenging problem in the field of Structural Bioinformatics. A principle difficulty has been in efficiently exploring the large number of possible shapes, or conformations, that a protein's chain can assume. To gain traction on this problem, the use of additional sources of structural information has been shown to be of use in navigating the conformation space. This work represents three methods to predict facets of protein structure solely from sequence. Two of the methods presented, DNcon and PROPcon, are used to predict residue-residue contacts and the other, DNdisorder, predicts the order/disorder state of a residue. This predicted information can be used directly by protein structure prediction pipelines to better navigate the complex and large protein conformation search space as well as be used to rank and assess the quality of predicted protein structures. All three methods, DNcon, PROPcon and DNdisorder, are built upon a novel combination of boosting and deep learning. By leveraging both of these machine learning techniques along with the processing power offered by graphical processing units, it was possible to train and test very large classifiers in a relatively short amount of time. Both DNcon and DNdisorder were benchmarked in the 10th round of the Critical Assessment of Protein Structure Prediction experiments and achieved at or near state-of-the-art performance.
Access is limited to the campuses of the University of Missouri.