Integrating Metabolic Networks into Hybrid Bioprocess Models

FBA-Hyb, a hybrid bioprocess model integrating flux balance analysis for Escherichia coli, improves predictive accuracy by 42% (R2) and preserves strict stoichiometric feasibility.

• Why it matters: Hybrid machine-learning bioprocess models can need large datasets and may produce inconsistent predictions that violate metabolic stoichiometry, limiting reliable optimization and control.
• What they did: The framework uses neural networks to predict FBA inputs (substrate uptake rates and FBA objectives) and a surrogate FBA module from symbolic regression to generate fluxes with backpropagation through the neural controlled differential equation model.
• The result: In leave-one-process-out cross validation on two Escherichia coli fed-batch case studies, FBA-Hyb achieved 42% higher R2 and stayed stoichiometrically feasible during extrapolation, while Std-Hyb was inconsistent in 22% of cases.