Statistical learning of bacterial growth in combinatorially constructed environments
Optimizing growth conditions and culture media is a major goal in microbiology. A challenge is that nutrients can have complex, non-additive effects on growth. The fact that resources interact with one another has long been known in ecology, but systematic maps of resource interactions at all orders have been lacking and it is not known whether interactions are primarily low order or fundamentally complex. To tackle this problem, we have followed a full factorial design approach and measured the growth of seven different bacterial species in all possible combinations of 8 carbon sources under carbon-limiting conditions. Our approach allows us to directly estimate interactions at all orders. Even though all C-sources stimulate growth on their own as well as in combination with other nutrients, most of them can also have negative effects on growth when they are added to at least some nutrient mixtures. We show that the switch from positive to negative fitness effects is governed by global epistasis among resources. An analysis of variance shows that additive effects and pairwise interactions explain most of the variation in fitness, allowing us to train simple regression models that accurately predict bacterial growth in novel environments. The generality of these findings across seven different bacterial strains belonging to two different families indicates that interactions between carbon sources under carbon-limiting conditions may be generally learnable from a relatively sparse set of constructed environments, enabling the rational optimization of growth conditions.