Integração de algoritmos de aprendizado de máquina ao docking molecular para planejamento e realização de ensaios in vitro de inibidores da enzima enoil-ACP-redutase NAD(P)H-dependente (FabI) de Staphylococcus aureus e de Escherichia coli.

• Ano de defesa: 2023
• Autor(a) principal: Gabriel Corrêa Veríssimo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais
• Link de acesso: https://hdl.handle.net/1843/76020
• topic: Aprendizado de máquina; Docking molecular; Planejamento de antibacterianos; Expressão e purificação de enzimas; Enoil-ACP-redutase (FabI)

The emergence of multidrug-resistant bacteria is a severe health problem worldwide. The World Health Organization (WHO) categorizes carbapenem-resistant strains of Escherichia coli and methicillin/vancomycin-resistant strains of Staphylococcus aureus as critical and high-level priorities for antibacterial drug development. Computational methods, as molecular docking, are accelerating the development of new antibacterials; however, predictability of molecular docking can be improved with machine learning (ML). In this context, the aim with this work is to integrate ML models with molecular docking studies to enhance predictive ability of ligand inhibitory activity against both S. aureus and E. coli FabI enzymes. Additionally, the aim is also to obtain recombinant FabI enzymes from both species for in vitro assays using compounds selected through virtual screening., and to obtain the S. aureus e E. coli FabI recombinant enzymes for in vitro assays with selected compounds based on ML-based virtual screening. Therefore, 2,352 docking protocols were validated using redocking, crossdocking, and ROC curve analyses. Eleven ML algorithms were deployed, generating 220,856,328 classification models correlating inhibitory activity with interaction fingerprints, calculated based on docking poses. We selected the best models for each enzyme using several classification metrics. The top three models for each protein were then utilized in a virtual screening to select substances that were subsequently tested against both E. coli and S. aureus cells. To obtain the recombinant proteins, we transformed Escherichia coli BL21(DE3) cells via electroporation with pET28a-saFabI and pET29a-ecFabI plasmids, each containing the coding sequences for the proteins of interest. These cells were cultured at 37 ºC in Luria-Bertani medium supplemented with MgSO4 and kanamycin. IPTG was used to induce expression for a duration of 18 hours at 18 ºC. Cell lysis was carried out by sonication, and protein purification was performed using affinity, desalting, and size exclusion chromatography. Both proteins were obtained in the form of highly pure and stable tetramers. With regard to the machine learning results, the top three models of each enzyme yielded MCCint values ranging from 0.567 to 0.846, and MCCext values ranging from 0.638 to 1.000. The Support Vector Machine (SVM) and Multilayer Perceptron (MLP) algorithms produced the best results, outperforming docking in all metrics These models were used in a virtual screening, allowing nine active compounds to be obtained in in vitro assays against E. coli and S. aureus cells. Subsequent STD-NMR studies will be conducted to verify the interaction of these substances with the obtained proteins.