Coarse-graining metabolic networks via feature learning reveals cross-species growth predictors

Bacterial growth and the underlying metabolic networks are highly dissimilar across species, posing a fundamental challenge for various bioengineering tasks. For a given species across nutrient environments, growth is regulated via proteome allocation, which gives rise to linear relationships between growth and the fractions of coarse-grained proteome sectors. However, it remains unclear whether there are coarse-grained predictors that correlate with growth across species monocultures in a given environment. Here, simulating genome-scale metabolic models, we discover two conserved growth predictors: the number of nutrients utilized, and a set of metabolic sectors that we identify by coarse-graining metabolic networks via feature learning. We find evidence supporting both predictors in recent experiments; notably, the proteome fractions of the sectors correlate with growth across species in a gene expression dataset. We further analyze coarse-grained networks to show that interactions between sectors improve growth predictions. Our work provides a framework to investigate how coarse-grained growth predictors emerge across diverse metabolic networks.