Biomedical knowledge discovery through multi-relational graph embeddings

• Autor(a) principal: Pereira, Rodrigo Amaral Ribeiro
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/31293

Knowledge graphs are multi-relational graph structures that allow to organize data in a way that is not only query able but that also allows the inference of implicit knowledge by both humans and, particularly, machines. In recent years new methods have been developed in order to maximize the knowledge that can be extracted from these structures, especially in the machine learning field. Knowledge graph embedding (KGE) strategies allow to map the data of these graphs to a lower dimensional space to facilitate the application of downstream tasks such as link prediction or node classification. In this work the capabilities and limitations of using these techniques to derive new knowledge from pre-existing biomedical networks was explored, since this is a field that not only has seen efforts towards converting its large knowledge bases into knowledge graphs, but that also can make use of the predictive capabilities of these models in order to accelerate research in the field. In order to do so, several KGE models were studied and a pipeline was created in order to obtain and train such models on different biomedical datasets. The results show that these models can make accurate predictions on some datasets, but that their performance can be hampered by some inherent characteristics of the networks. Additionally, with the knowledge acquired during this research a notebook was created that aims to be an entry point to other researchers interested in exploring this field.