BACKGROUND: Gene fusion detection - also known as the 'Rosetta Stone' method - involves the identification of fused composite genes in a set of reference genomes, which indicates potential interactions between its un-fused counterpart genes in query genomes. The precision of this method typically improves with an ever-increasing number of reference genomes. RESULTS: In order to explore the usefulness and scope of this approach for protein interaction prediction and generate a high-quality, non-redundant set of interacting pairs of proteins across a wide taxonomic range, we have exhaustively performed gene fusion analysis for 184 genomes using an efficient variant of a previously developed protocol. By analyzing interaction graphs and applying a threshold that limits the maximum number of possible interactions within the largest graph components, we show that we can reduce the number of implausible interactions due to the detection of promiscuous domains. With this generally applicable approach, we generate a robust set of over 2 million distinct and testable interactions encompassing 696,894 proteins in 184 species or strains, most of which have never been the subject of high-throughput experimental proteomics. We investigate the cumulative effect of increasing numbers of genomes on the fidelity and quantity of predictions, and show that, for large numbers of genomes, predictions do not become saturated but continue to grow linearly, for the majority of the species. We also examine the percentage of component (and composite) proteins with relation to the number of genes and further validate the functional categories that are highly represented in this robust set of detected genome-wide interactions. CONCLUSION: We illustrate the phylogenetic and functional diversity of gene fusion events across genomes, and their usefulness for accurate prediction of protein interaction and function.