Fig. 4.

Two-way clustering of winning and losing interactions. Samples were clustered with Ward’s hierarchical clustering method using the top random forest predictors for winning and losing corals (A, D, and G), interfaces (B, E, and H) and algae (C, F, and I) using functional genes (A–C), bacterial taxa (D–F), and metabolites (G–I) as inputs for the random forests. Abbreviations are as follows: A, algae; C, coral; D, D. strigosa; I, interface; L, losing; M, O. faveolata; W, winning. For metabolites, all of the significant predictors of the coral–turf algal competition outcome were unannotated in the GNPS database. Thus, the first number is the GNPS cluster index, the following number is the mass-to-charge ratio, the third number is the retention time, and finally, the network subcluster ID where −1 indicates a single looped compound. For all heat maps, redder indicates relatively higher abundances and bluer indicates relatively lower abundances.