Combining deep learning and knowledge graphs in PPI network prediction tasks

Balbi, Laura de Oliveira Lopes Henchman

http://hdl.handle.net/10451/59200

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
TM_Laura_Balbi.pdf		2.64 MB	Adobe PDF	Download

Send Feedback

Authors

Balbi, Laura de Oliveira Lopes Henchman

Advisor(s)

Pesquita, Cátia Luísa Santana Calisto

Abstract(s)

Many bioinformatics problems pertain to large, highly complex amounts of biological data that are often modelled in a graph-like arrangement to allow for a systemic-level analysis of data. Graphs provide a means of encoding biological knowledge into a formal structure that is useful for modelling and analysing relationships in biological systems. Several machine learning approaches have been developed to deal with data represented as graphs, namely by using graph neural networks, end-to-end representation-learning methods that can directly learn from graph-structured data. A form of knowledge representation that allows for the conceptualization and specification of domains of interest is the use of ontologies. These can have a biological application into representing and structuring existing knowledge by its meaning and relationships, allowing for the organisation of large volumes of biological entities into knowledge graphs. With both biological data and biological knowledge represented as graphs, an opportunity to directly enrich the data graph with knowledge about its entities arises. Thus, the aim of this work was to explore different approaches into combining a PPI network with information pertaining to the Gene Ontology to leverage the additional biological knowledge in a protein function prediction setting. Two methodologies were proposed for the development of said approaches. The first comprehends the creation of approaches to merging a PPI network with protein function information and combination of said approaches with different ML models to test if they benefit from the additional information during protein function prediction. The second methodology sees the construction of a GNN -based method to learning PPI and GO representations in separate and combining them for a global learning of protein function -related information. The evaluation of the different approaches and different experimental conditions with a benchmark dataset showed an overall performance increase.