GRaSP: uma estratégia de aprendizagem supervisionada baseada em grafos de vizinhança de resíduo para previsão de sítio de ligação

Abstract

Proteins are macromolecules crucial for the maintenance of biological systems and participate in vital processes for the cell. Protein activity is performed through physicochemical interactions between the protein and other molecules called ligands. These ligands comprise organic compounds, metal ions, nucleic acids or even other proteins, in which attach to the protein so that its activity is properly performed. The region on the protein where these interactions take place is called binding sites. The identification and characterization of these regions is crucial to determine the function of a protein, which is one of the necessary steps in areas such as the design and development of new drugs. Due to experimental issues, the location of these regions may not be trivial, requiring the support of automatic methods to assist in their identification. In this thesis, the GRaSP is proposed, a machine learning-based strategy for binding site prediction that uses residue neighborhood graphs as input. From experiments using databases of different protein structures, GRASP proved to be robust, presenting compatible or better results in relation to the tools already consolidated in the literature. Furthermore, due to the simple and informative modeling provided by the graphs, the algorithm proved to be efficient. While already consolidated methods take around 5 hours of processing for protein structures with approximately 300 residues, the GRASP is able to process them in an average time of 20 seconds.