Learning the mechanism of collective microbial function via random community-media pairing

In microbial systems, many biochemical functions arise from pathways encoded and executed at the community level. The collective nature of these functions complicates bottom-up efforts to determine each species' contribution. Previous work has shown that regression over randomly sampled datasets of collective functions succeeds at predicting those functions. Building on this top-down idea, this paper asks whether regression can also reveal mechanistic insight into a cross-feeding relationship. For this, we propose extending the random sampling method to vary the growth environment, cultivating each community in the spent medium of another randomly constructed community. Using a model-based analysis, we illustrate that the mechanistic roles of species are structurally non-identifiable under the standard random sampling protocol, but become identifiable after the proposed modification, such that even a simple LASSO regression can recover them. More generally, our work suggests that the utility of machine-learning-based approaches can be greatly enhanced by a synergistic experimental design.