Application of Graph Learning Methods to Metabolic Networks

Metabolic networks are complex systems of chemical reactions that sustain life. These networks neutrally can be represented as graphs, making them suitable for analysis using graph learning methods. This thesis explores the use of two graph learning models, Deep Graph Convolutional Neural Networks (DGCNNs) and Graph Kernel Neural Networks (GKNNs), to analyze metabolic networks represented in three ways: Abstract Metabolic Networks (AMNs), Metabolic Directed Acyclic Graphs (mDAGs), and Reaction Graphs (RGs). The study uses data from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database to evaluate the performance of these models. The results show that the choice of representation affects model performance. AMNs, the simplest representation, achieve high accuracy, making them suitable for high-level comparisons. mDAGs, which balance detail and complexity, are more challenging but useful for identifying key pathways. RGs, the most detailed representation, are the hardest to model due to their complexity. DGCNNs performs well on simpler representations but struggles with overfitting on smaller datasets. GKNNs shows better generalization, especially with larger datasets and simpler configurations. This work highlights the importance of selecting the right representation and model for analyzing metabolic networks. It also suggests future research directions, such as improving model optimization. These advancements could lead to better tools for understanding complex biological systems.