Desenvolvimento de funções de pontuação para predição de afinidade proteína-ligante utilizando Redes Neurais Profundas

• Autor(a) principal: Vidal, Lincon Onório
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/349
• topic: Planejamento de fármacos; Aprendizado por computador; Redes neurais (Computação); DockTDeep (Programa de computador); CNPQ::CIENCIAS BIOLOGICAS::FARMACOLOGIA

Computer-aided drug design (CADD) methodologies have been used successfully in researching new and safer pharmaceutical drug candidates, helping to reduce the high costs involved in developing new compounds. One of the main CADD methods is structure-based virtual screening (SBVS), in which large libraries of compounds and molecular targets are computationally evaluated in order to classify them biologically as active or inactive compounds. In this context, a scoring function (SF) is used to predict the affinity value between a compound bound to a protein target of interest, playing a crucial role in SBVS methods. More recently, Machine Learning techniques, and particularly Deep Neural Networks, have been employed in the development of SF for affinity prediction. In this work, we present the development of SFs based on Convolutional Neural Networks and Residual Neural Networks models using the DockTDeep program, developed by GMMSB, aiming to enable the use of such functions in SBVS applications. Several analyses of the influence of the main elements used in the construction of the predictive model, such as voxel grid-based representation and treatment of protein-ligand complex data, cost functions for model optimization, regularization strategies and data augmentation techniques, as well as analyses regarding the biases of the data sets and their representations, are also performed. The developed models obtained competitive results compared to other scoring functions present in the literature, reaching 𝑅 = 0.721 in CASF-2016 and 𝑅 = 0.718 in CASF-2013, while starting from a simpler construction that dispenses the use of descriptors manually selected by molecular modeling experts. Such results demonstrate the practical feasibility and potential of using Deep Neural Network-based scoring functions for SBVS applications.