EVENTO
Development of Emprical Scoring Functions for Predicting Protein-ligand Binding Affinity
Tipo de evento: Defesa de Tese de Doutorado
INTRODUCTION. DockThor has obtained promising results in comparative studies with other well established receptor-ligand docking programs for predicting experimental binding modes. Despite useful for pose prediction, the current scoring function implemented in DockThor is not suitable for predicting binding affinities of protein-ligand complexes.OBJECTIVES: In this work, we develop several scoring functions from first-principles with interaction-driven features for predicting binding affinities of protein-ligand complexes trained with several machine learning techniques.MATERIALS AND METHODS: We implemented several empirical and physics-based features that represent the protein-ligand binding event. Then, we developed general and specific scoring functions for target-classes, the last to account for binding characteristics associated with a target class of interest, focusing on proteases, kinases and protein-protein interactions complexes (PPIs). The scoring functions were derived using linear regression (MLR) and seven machine learning techniques for nonlinear problems using the PDBbind refined set 2013 (N = 2959) for training and testing. We also trained and evaluated general scoring functions using docking poses obtained with DockThor.DISCUSSION AND RESULTS: The linear and the best nonlinear (random forest) general scoring functions trained with experimental structures obtained great performances when evaluated on the benchmark PDBbind core set 2013 (N = 195) (RMLR = 0.602 and RRF = 0.704). These results demonstrated that our scoring functions are competitive with the best scoring functions evaluated in such benchmarking studies, i.e. X-ScoreHM (linear) with R = 0.644 and RF::VinaElem (nonlinear) with R = 0.752. The scoring functions specific for target classes also obtained good performances on independent test sets: RSVM = 0.723 for proteases, RLWL = 0.702 for kinases and RSVM = 0.651 for iPPIs. Furthermore, the scoring functions trained with docking results obtained promising performances when evaluated in both experimental and docking structures, indicating that they are reliable to be applied in both cases.CONCLUSIONS: The development of the scoring functions implemented in this thesis is a crucial step to make the DockThor an even more competitive program, enabling the development of the virtual screening program and portal DockThor-VS.
Data Início: 15/07/2016 Hora: 10:00 Data Fim: 15/07/2016 Hora: 13:00
Local: LNCC - Laboratório Nacional de Computação Ciêntifica - Auditorio B
Aluno: Isabella Alvim Guedes - Laboratório Nacional de Computação Científica - LNCC
Orientador: Laurent Emmanuel Dardenne - Laboratório Nacional de Computação Científica - LNCC
Participante Banca Examinadora: Adriano Andricopulo - Instituto de Fisica de São Carlos - USP Carlos Mauricio Rabello de Sant'Anna - UFRRJ - UFRRJ Ernesto Raul Caffarena - PROCC - FIOCRUZ/RJ Helio José Corrêa Barbosa - Laboratório Nacional de Computação Científica - LNCC/MCTI Laurent Emmanuel Dardenne - Laboratório Nacional de Computação Científica - LNCC
