Abstract: Genome-scale metabolic modeling enables omic data integration through mathematical simulation and has become an indispensable cornerstone for understanding cellular metabolism. Traditional analysis tools, such as mechanistic modeling and flux balance analysis, require deep domain expertise to specify the kinetic parameters or significant manual effort to acquire fluxomic data to formulate the constrained optimization problem. To circumvent the above limitations, we develop a novel metabolic dynamics modeling framework, which learns a structured neural ODE process (SNODEP) model to predict the time-varying flux and balance distributions by leveraging the more accessible single-cell RNA sequencing (scRNA-seq) technology. Compared with ML-based alternatives, our method achieves enhanced prediction performance, not only due to the intrinsic suitability of neural ODE for modeling dynamics-governed time series data but also because the design of SNODEP explicitly accounts for the destructive measurement process of scRNA-seq and the sequential dependence between context points. Comprehensive evaluations across $4$ metabolic pathways ($340$ experiments in total) show that our method can predict future gene expression, flux, and balance dynamics well, even generalizing to more challenging settings of irregularly sampled data and unseen gene knockout configurations. We hope our work can catalyze the development of more robust and scalable models for metabolic pathway analysis.