(591d) Modeling Metabolic Networks Using Graph Neural Networks

Mathematical modeling of metabolic networks has applications ranging from health-related applications such as disease progression and drug discovery to metabolic engineering and bioremediation. Commonly used frameworks such as genome-scale stoichiometric models (GEMs) or kinetic models suffer from drawbacks such as the inability to make dynamic predictions or uncertainty in parameter estimation.

Neural networks have been shown to be effective in capturing trends and behavior in nonlinear datasets. Despite a high number of parameters, neural networks can be suitably trained on the available training datasets to serve as predictive models. In this work, we utilized graph neural networks (GNNs) to model the performance of metabolic networks that are characterized under various network perturbations. Three scenarios were explored in the study to cover different types of datasets and network coverage.

In the first scenario, we set up a synthetic network modeled using ordinary differential equations (ODEs), and trained the GNN on different configurations of simulated ODE data. We showed that even on a moderate number of 5 time points of data from the synthetic dataset, the neural network was able to recapitulate the metabolic behavior of the synthetic network.

In the second scenario, a GNN was used to model the ¹³C metabolic flux datasets of Escherichia coli mutants. Modeling network behavior using flux datasets helped formulate empirical design rules on choosing between concentration and flux measurement datasets for training a GNN. We show that flux values can serve as a better training dataset than metabolite concentrations in certain cases with appropriately designed gene knockout mutants.

Finally, we trained a GNN on the lipid metabolism of Saccharomyces cerevisiae using lipidome concentration measurements of various gene knockout mutants. The neural network was able to accurately predict the trends of increase or decrease in specific lipid classes within the testing dataset, suggesting future applications in strain design for metabolic engineering or health interventions in neurodegenerative disease-related lipid imbalance.

We thus demonstrated that a GNN can serve as a substitute for mechanistic frameworks such as GEMs or kinetic models in metabolic network modeling. We suggest that by employing previously developed problem formulations in GNNs such as link prediction, these models can also be used to formulate hypotheses on discovering new regulatory interactions within a network.