Falcão, André Osório e Cruz de AzerêdoGuedes, Rita Alexandra do Nascimento CardosoOliveira, Pedro Frederico Matos Regalado de2024-02-122024-02-1220242023http://hdl.handle.net/10451/62588Trabalho de Projeto de Mestrado, Engenharia Informática, 2024, Universidade de Lisboa, Faculdade de CiênciasIn silico drug development is increasingly being seen as a valuable approach, offering a faster pace for drug discovery while concurrently driving down costs. Ligand-based methods, anchored on the chemical and biological attributes of recognized ligands, are effective in predicting interactions with extensively studied target proteins. However, a hurdle presents itself when these methodologies are tasked with making predictions for targets that haven’t been previously explored. Addressing this shortcoming, a novel methodology is introduced that melds structural data of target proteins—sourced from AlphaFold 2’s predictive structures—into machine learning models, thereby elevating their ability to predict protein-molecule interactions. This innovative strategy refines and extracts 3D structural protein fingerprints and then amalgamates them with the structural data of ligands This enriched dataset trains our machine learning model to discern the nuanced dynamics between ligand attributes and distinct structural nuances of myriad target proteins, facilitating predictions for previously uncharted molecules and protein targets. To assess the efficacy of the introduced model, comprehensive datasets, encompassing two distinct bioactivity types and detailing the entirety of available information for Human G-Protein Coupled Receptors, were employed. The derived insights highlighted that this innovative strategy parallels the prowess of contemporary traditional ligand-based methodologies. It was found that, in some cases, the model exhibited a unique ability to accurately predict interactions for target proteins that were not included in its training phase. This ability hinges on the presence of some level of similarity between these external proteins and those within the training set. Such capabilities underscore the approach’s expansive potential, offering a promising avenue to advance drug research into previously uncharted target proteins. In summary, this work introduces a novel in silico drug discovery technique, adeptly merging ligand-centric methodologies with structural data integration to predict protein-molecule interactions. By integrating both protein and ligand structural information within a machine learning framework, the model paves the way for robust, automated predictions, even for targets that were absent during the training process. This innovation signifies a breakthrough in drug discovery, presenting far-reaching implications for the future landscape of the pharmaceutical industry.engDescoberta de drogasModelação Relação Estrutura-AtividadeTriagem virtual baseada na estruturaMachine LearningEstrutura de proteínasTeses de mestrado - 2024master thesis203682998