Um estudo sobre a predição da estrutura 3D aproximada de proteínas utilizando o método CReF com refinamento

Abstract

One of the most important problems in Structural Bioinformatics is to understand how the information coded in linear sequence amino acids, or primary structure, is translated into the three-dimensional structure of a protein. Many algorithms proposed solutions to this complex problem of NP-complete class. One of them is the CReF method (Central Residue Fragment-based) which makes prediction of approximate 3-D structure of proteins and polypeptides. The method uses data mining techniques to group data structures, showing good secondary structure prediction, good performance at low machine cost, but has problems in the prediction of turns and loops regions and usability. Valuing the different characteristics of CReF and seeking to evolve it, this work proposes improvements to CReF. After the initial stage of understanding the tool and making changes to turn it executable on the current state of data banks and support tools, two categories of improvements to make were identified. The technical improvements aimed to automate CReF, adapting it to the environment and emphasizing usability. In the method‟s improvements variations on the amount of groups were tested for data mining with the Expectation Maximization algorithm in Weka. Tests indicated that the best results for the initial conformation were for four and six groups, hence we decided to allow the user to select the amount of groups. A new mapping of the data in the Ramachandran plot indicated some problems that had to be fixed. In the analysis of data mining results, we decided that groups in regions not allowed would be discarded. The new version of CReF generated by the implementation of these improvements standardized the method of secondary structure prediction to use Porter. As a consequence, the rules of selection of data mining groups to represent each amino acids have been changed and extended. The new version has the same initial performance of CReF in prediction and execution, however, the problem of correct predictions of turns and loops remained. This problem was addressed through a refinement protocol, based on simulations by the molecular dynamics method, which presented a significant result for the target protein 1ZDD.