Beyond Linear Programming: Physics-Informed Graph Neural Networks for Multi-Omics-Driven Metabolic Flux Prediction in Plants
Kat Powell
Department of Computer Science
Faculty Supervisor: Sara El Alaoui
Nitrogen availability significantly influences the growth and yield of bioenergy crops, yet its downstream effects on primary metabolic flux remain poorly characterized. Constraint-based modeling offers a useful starting point, but linear programming assumptions restrict its ability to capture nonlinear regulatory dynamics that emerge from context-specific multi-omics data.
Here we introduce a physics-informed Graph Neural Network (PIGNN) framework to predict flux distributions in plant metabolic networks. The metabolic network is encoded as a graph whose topology is derived from PlantSEED reconstructions, with stoichiometric properties and transcriptomic features defining condition-specific node and edge embeddings. Nodes represent metabolic components, while edges represent flux transfer between them conditioned on omics data.
We use Sorghum as a model plant and integrate a resource-rich, longitudinal multi-tissue dataset from a project studying sorghum’s growth across varying nitrogen levels. By integrating transcriptomics and metabolomics, the model learns how gene expression and metabolite profiles under different nitrogen levels translate into shifts in flux through pathways where amino acid, nucleic acid, and fatty acid biosynthesis converge.
A key contribution is a biochemically grounded PIGNN model that penalizes violations of stoichiometric mass balance and nutrient uptake bounds, embedding these constraints into the learning process to yield accurate, biologically consistent flux predictions.