3D-Tri : um algoritmo de indução de árvore de regressão para propriedades tridimensionais - um estudo sobre dados de docagem molecular considerando a flexibilidade do receptor

Abstract

With the growth of biological experiments, solving and analyzing the massive amount of data being generated has been one of the challenges in bioinformatics, where one important research area is the rational drug design (RDD). The interaction between biological macromolecules called receptors, and small molecules called ligands, is the fundamental principle of RDD. In in-silico molecular docking experiments it is investigated the best bind and conformation of a ligand into a receptor. A docking result can be discriminated by a continue value called Free Energy of Binding (FEB). We are attempting on mining data from molecular docking results, aiming at selecting promising receptor conformations to the next docking experiments. In this sense, we have developed a comprehensive repository to store our molecular docking data. Having such repository, we were able to apply preprocessing strategies on the stored data and submit them to different data mining tasks. Among the techniques applied, the most promising results were obtained with regression model trees. Although we have already addressed important issues and achieved significant results, there are some properties in these experiments turning it difficult to properly select conformations. Hence, a strategy was proposed considering the three-dimensional (3D) properties of the receptor conformations, to predict FEB. This thesis presents the 3D-Tri, a novel algorithm able to handle and treat spatial coordinates in a x,y,z format, and induce a tree that predicts FEB value by representing such properties. The algorithm uses such coordinates to split a node in two parts, where the edges evaluate whether the atom being tested by the node is part of a given interval [(xi,xf );(yi,yf );(zi,zf )], where i indicates the initial position of the coordinate, and f its final position. The induced model can help a domain specialist to select promising conformations, based on the region of the atoms in the model, to perform new molecular docking experiments