Neural ODEs for CHO Hybrid Modeling

Literature models for metabolism in Chinese hamster ovary (CHO) are not regular chemical kinetics--they include both kinetic expressions and stoichiometric constraints that the kinetics must obey.

We formulate several alternative realizations of these (conflicting) rules, and then use the neural ordinary differential equations framework to fit unknown terms in the resulting model to both real and simulated data.

A key wrinkle of the process is that the several literature forms of the model imply the evaluation of an inner constrained optimization problem on every forward pass of the ODE. This needs to be accounted for in the neural network form as well. For this, we rely on the work of Zico Kolter et al. for embedding constrained optimization problems in neural networks.