Neural-Mechanistic Hybrid Modeling for Metabolism (BioFlux)

BioFlux is my line of work on physics-informed machine learning for metabolism: neural-network solvers (no trained weights) that make genome-scale metabolic models predictive from multi-omics data while respecting hard mechanistic constraints (mass balance), so predictions stay biochemically valid.

• First result: a flexible neural-mechanistic hybrid that simulates the iron-deficiency response in plant plastidial metabolism (senior author, bioRxiv, 2025).
• Method: biochemistry-informed mechanistic gradient optimization with mass-balance constraints, run as a PINN in solver mode (no trained weights); convergence analysis across data-fit vs. mass-balance loss formulations.
• Why it matters: predict how a cell, strain, or plant will behave under perturbation — before it’s measured.