a, Depiction of a standard pairwise analysis that considers a sequence relative abundance matrix for s samples (s × OTUs (operational taxonomic units)) and its corresponding environmental matrix (s × p (parameters)). sPLS results emphasize OTU(s) that are the most correlated to environmental parameters. b, Depiction of a graph-based approach. Using only a relative abundance matrix (s × OTUs), WGCNA builds a graph where nodes are OTUs and edges represent significant co-occurrence. Co-occurrence scores between nodes are weights allocated to corresponding edges. These weights are magnified by a power-law function until the graph becomes scale-free. The graph is then decomposed within subnetworks (groups of OTUs) that are analysed separately. One subnetwork (group of OTUs) is considered of interest when its topology is related to the trait of interest; in the current case carbon export. For each subnetwork (for instance the subnetwork related to carbon export), each OTU is spread within a feature space that plots each OTU based on its membership to the subnetwork (x axis) and its correlation to the environmental trait of interest (that is, carbon export). A good regression of all OTUs emphasizes the putative relation of the subnetwork topology and the carbon export trait (that is, the more a given OTU defines the subnetwork topology, the more it is correlated to carbon export). c, Depiction of the machine learning (PLS) approach that was applied following subnetwork identification and selection. Greater VIP scores (that is, larger circles) emphasized most important OTUs. VIP refers to variable importance in projection and reflects the relative predictive power of a given OTU. OTUs with a VIP score greater than 1 are considered as important in the predictive model and their selection does not alter the overall predictive power.