Aplicação de aprendizado de máquina para descobertas de farmacogenômica no tratamento do câncer de esôfago

• Ano de defesa: 2022
• Autor(a) principal: Leal, Gabriel Fernandes
• Orientador(a): Bordini, Rafael Heitor
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Escola Politécnica
• País: Brasil
• Palavras-chave em Português: Aprendizado de Máquina; Farmacogenômica; Oncologia; Câncer de Esôfago
• Palavras-chave em Inglês: Machine Learning; Pharmacogenomics; Oncology; Esophageal Cancer
• Área do conhecimento CNPq: CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede2.pucrs.br/tede2/handle/tede/10261

Pharmacogenomics is the area that studies how genomic variations influence drug response. Through it studies, it is possible to explore and define the most suitable drugs for different patients and their genetic profiles, in order to make treatments more personalized. Recent studies map the response of cancer-related cell lines to a wide collection of drugs used in treatments, applying machine learning techniques for prediction tasks. Our goal is to develop deep neural network models seeking to predict the response of different profiles to 174 drugs used for the treatment of esophageal cancer. Deep learning models were built to estimate the response of different compounds, based on its IC50 values, by integrating expression, mutation and clinical data. Autoencoders were developed to extract the representation of the training data, combined with a deep neural network. The initial model obtained positive results compared to previous work and based on these we explored new approaches to improve the neural network. We introduced a new architecture with the integration of clinical data due to the importance of risk factors related to esophageal cancer cases. Furthermore, another motivation to explore these data is that they are still more common to be obtained in clinical practice. The models presented results of 0.74 and 0.72 respectively, considering the mean squared error evaluation metric. Despite the positive results, implementation limitations were identified, especially regarding clinical data in terms of quantity and quality of information. The experimental results show that the research topic is promising and can lead to innovations capable of improving the quality of life of patients.