(120e) Next-Flux (Neural-net EXtracellular Trained Flux)

Genome-scale modelling has emerged as an effective tool for the modelling of mammalian cell metabolism. Despite the acceleration of advancements and applications in the field of genome-scale modelling, there are many fundamental challenges to be addressed. Current mammalian genome-scale models (GeMs) are heavily underdetermined, and methods used to constrain the solution space are limited in depth and applicability. These constraints fail to guide flux in a biologically consistent manner, require expensive/unavailable data and are often based on a multitude of assumptions.

In this work we present a novel method for constraining Chinese hamster ovary (CHO) cell GeMs that exploits the link between extracellular and intracellular metabolomics to create biologically relevant bounds on intracellular fluxes. We train an artificial neural network (ANN) with a variety of published CHO cell extracellular metabolomics and corresponding ¹³C-labelled intracellular fluxomic data to quantify the relationship between easily measured metabolites and difficult to predict fluxomics. The resulting model is used to predict intracellular flux distributions for unseen conditions, given a set of process-level experimental data (metabolite uptake rates, cell growth rate, and recombinant protein production rate). The model predicts the expected upper and lower bound of up to 47 intracellular fluxes across glycolysis, pentose phosphate pathway, TCA cycle and amino acid metabolism, which are then used to constrain a CHO cell GeM.

We show that this hybrid approach that combines stoichiometric modelling with data-driven constraints accurately predicts physiologically relevant metabolic states across a variety of validation conditions. Flux of key reactions is constrained within tight bounds, preventing infeasible solutions and guiding flux down appropriate pathways. The resulting intracellular flux distribution is more biologically consistent, improving the reliability of CHO cell GeMs in the application to industrial systems.

We present this method in an easy-to-implement package for the systems biology community to employ.