Members of some phyla were used for training of all tools, other phyla were not represented by a training species.
The models underlying all metagenomic gene prediction tools were built on the basis of genomes from selected training species.
We find that the numbers of predicted interactions are highly similar between the training and prediction species, although the training species do have an enrichment of genes at the high confidence levels (Table 1).
We show that gene prediction accuracy varies over reads from different test species but we believe that this variation is independent from the degree of relatedness to training species.
Generalization capabilities of those models can only be analyzed if training species and their close relatives (we define close relative as species from the same genus) are excluded from the evaluation setup.
In this study we find that our random forest-based classifier, ENTS, can predict biologically meaningful PPIs both within the species on which we trained the classifier and within species sharing a relatively distant common ancestor with the training species.
