Desenvolvimento e aplicação de funções escore otimizadas para previsão de afinidade entre proteínas e ligantes

Abstract

Kinases are the most intensively studied protein in drug design and development. Among kinases, non-specific serine/threonine protein kinase represents an interesting protein system for modeling purposes due to the availability of structural and functional experimental data. Non-specific serine/threonine protein kinase comprises an important class of protein targets used to develop drugs to treat cancer. In this study, we describe the creation of machine learning models to predict protein-ligand binding affinity for this enzymatic class. We make use of energy terms available in classical scoring functions such as Autodock4 and AutoDock Vina. We use these terms to build a novel scoring function targeted to a dataset comprised of nearly 100 protein-ligand complexes for which experimental crystallographic structure and inhibition constant data are available. We also applied a hybrid mass-spring method to determine binding affinity using the program Taba. We carried out predictive performance analysis of all scoring functions. Our study clearly shows that machine learning models present superior predictive performance when compared with classical scoring functions. Also, our machine learning models can capture structural features responsible for the binding affinity against non-specific serine/threonine protein kinases.