Telling mutualistic and antagonistic ecological networks apart by learning their multiscale structure

Characterizing and understanding the processes that shape the structure of ecological networks, which represent who interacts with whom in a community, has many implications in ecology, evolutionary biology, and conservation. A highly debated question is whether and how the structure of a bipartite ecological network differs between antagonistic (e.g., herbivory) and mutualistic (e.g., pollination) interaction types. Here, we tackle this question by using a multiscale characterization of network structure, machine learning tools, and a large database of empirical and simulated bipartite networks. Contrary to previous studies focusing on global structural metrics, such as nestedness and modularity, which concluded that antagonistic and mutualistic networks cannot be told apart from only their structure, we find that they can be told apart by combining a meso-scale characterization of their structure and supervised machine learning. Motif frequencies appear particularly informative, with an over-representation of densely connected motifs in antagonistic networks and of motifs with asymmetrical specialization in mutualistic networks. These characteristics can be used to predict interaction types with relatively good confidence. Beyond this classical mutualism/antagonism dichotomy, we also find significant structural uniqueness linked to specific ecologies (e.g., pollination, parasitism). Our results clarify structural differences between antagonistic and mutualistic networks and suggest the investigation of the structural uniqueness of specific ecologies as a promising approach for characterizing interactions beyond the coarse antagonistic/mutualistic dichotomy.