Method for predicting the viability of proteins docking using Deep Learning techniques

• Grant number: 21/04626-8
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: November 01, 2021
• End date: October 31, 2022
• Field of knowledge: Interdisciplinary Subjects
• Principal Investigator: Rafael Plana Simões
• Grantee: Ramon Hernany Martins Gomes
• Host Institution: Faculdade de Ciências Agronômicas (FCA). Universidade Estadual Paulista (UNESP). Campus de Botucatu. Botucatu , SP, Brazil

Abstract

Molecular docking is a theoretical-computational method for predicting the interaction between the surfaces of two molecules. Various software for molecular docking with different search algorithms and scoring functions are available, and they have been used extensively for research and innovation activities. However, despite all the consolidated scientific development that underlies the traditional docking methodologies, some problems limit the use of these technologies. The main problem of the docking techniques is the low correlation between the binding energies predicted by classical approaches with the energies determined empirically. In addition, it should be considered that the classical methodologies do not allow infer about the viability (or existence) of the complex in a binary way: EXIST or NO EXIST, thus leaving the problem of predicting the viability of the complexes generated by docking. This project's main objective is to develop a methodology for the validation of the molecular complexes inferred by traditional docking methods using a Deep Learning approach. For this purpose, two datasets will be created: one containing experimentally validated molecular structures that form a protein-protein complex (obtained from a protein databank) and the other containing molecular structures with a low probability of forming molecular complexes. This set of instances will be expanded using molecular dynamics techniques. Then, an algorithm will be developed to convert these molecular complexes into two-dimensional images (matrix of pixels) that will contain relevant information about the interaction of the proteins. These images will compose the training and test sets for the Deep Learning algorithm. From this, a model will be inferred using Deep Learning techniques for binary prediction (EXIST, or NO EXIST) of solutions obtained by conventional docking methods. Finally, an evaluation of the predictive performance of the proposed methodology will be carried out using a test set.(AU)