Learning beyond-pairwise interactions enables the bottom-up prediction of microbial community structure

The way to deal with higher-order effects (i.e., modification of pairwise interactions by third-party species) has been a major consideration in community ecology. Ignoring these effects is not in line with reality, yet fully considering them make the situation overly complex. Here, we propose a simple framework incorporating higher-order effects into a bottom-up community modeling, and assessed its validity using a seven-member synthetic bacterial community on a host plant, duckweed. Our findings revealed that actual interspecies interactions in community could not be predicted from pairwise co-culturing results; however, using information from trio combinations allowed for acceptable prediction. In addition, inclusion of four-, five-, and six-member combinations did not markedly enhance the prediction accuracy from trio-based prediction, suggesting that trio combinations, the smallest unit of higher-order effects, provide a reasonable baseline to unravel complex interaction networks. Building on this finding, we developed a prediction rule to estimate the structure of 4 - 7 member communities based on information from [≤] 3-member combinations, which yielded significantly better accuracy (relative mean square percentage errors of 22.7% - 61.2%) than pairwise-based model (53.5% - 185.2%). This highlights the possibility of establishing a quantitative link between the interspecies interactions and community structure, by observing beyond-pairwise combinations.