In silico discovery of GPCR ligands using graph-based signatures and auxiliary features

• Ano de defesa: 2022
• Autor(a) principal: João Paulo Linhares Velloso
• Orientador(a): Douglas Eduardo Valente Pires , David Benjamin Ascher
• Banca de defesa: Lucas Bleicher, Rafaela Salgado Ferreira, Rubens Lima do Monte Neto, Wandré Nunes de Pinho Veloso
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Bioinformatica
• País: Brasil
• Palavras-chave em Português: Aprendizado de Máquina; Desenvolvimento de Medicamentos; Receptores Acoplados a Proteínas G
• Link de acesso: http://hdl.handle.net/1843/52800

GPCRs are crucial receptors for many vital physiological processes including control of cell division and proliferation, regulation of ion transport, modulation of neuronal firing, homeostasis, modulation, and modification of cell morphology. They are also involved in many pathological processes, such as in Alzheimer’s and Parkinson’s disease, cardiovascular disorder, asthma, depression and diabetes. Given their biological importance, over a third of FDA approved drugs target GPCRs. Nonetheless, GPCRs lead compound development suffers from high attrition rates, with poor in vivo efficacy being the primary contributor, resulting in only 7% of all drugs (for other receptors as well) in phase I studies being marketed. This thesis focused on the development of machine learning models capable of predicting bioactivity of small molecules when interacting with GPCRs as means to support the discovery of novel leads through ranking compounds on drug discovery investigations, which would enable enriching screening libraries with compounds more likely to be active. The developed models (composing the pdCSM-GPCR tool) rely on deriving a range of molecular signatures from known ligands, associating them to bioactivities, and modelling them as regression problems, making them independent of receptor structural information. Because of this characteristic, the same approach can be used for any GPCRs which already had been screened for ligands, and also other important targets, including kinases, and ligand-gated ion channels. Our models make up the most comprehensive computational resource for prediction of GPCR bioactivity to date, including support for drug development for orphan GPCRs. Our approach achieved Pearson’s correlations of up to 0.89, across 10-fold cross- validation and blind tests. We significantly outperformed previous methods. pdCSM-GPCR was made freely available via a user-friendly web server at http://biosig.unimelb.edu.au/pdcsm_gpcr. We also investigated the properties of small molecules with high affinity for GPCRs in order to identify molecular determinants of recognition. Overall, potent ligands possess nitrogen- containing fragments and aromatic rings, features common in ligands across all classes of GPCRs. The outcomes of this research provide powerful tools for GPCR drug discovery and valuable biological insights into the characteristics that make up GPCR ligands.