Here, gene fusions are the most powerful type of <
i>
in silico<
/i>
-derived functional associations. However, many fusion compilations were made when <
100 genomes were available, and algorithms for identifying fusions need updating to handle the current avalanche of sequenced genomes. The availability of a large fusion dataset would help probe functional associations and enable systematic analysis of where and why fusion events occur. As a result, here we present a systematic analysis of fusions in prokaryotes. We manually generated two training sets: (i) 121 fusions in the model organism <
i>
Escherichia coli<
/i>
(ii) 131 fusions found in B vitamin metabolism. These sets were used to develop a fusion prediction algorithm that captured the training set fusions with only 7 % false negatives and 50 % false positives, a substantial improvement over existing approaches. This algorithm was then applied to identify 3.8 million potential fusions across 11,473 genomes. The results of the analysis are available in a searchable database. A functional analysis identified 3,000 reactions associated with frequent fusion events and revealed areas of metabolism where fusions are particularly prevalent. In conclusion, customary definitions of fusions were shown to be ambiguous, and a stricter one was proposed. Exploring the genes participating in fusion events showed that they most commonly encode transporters, regulators, and metabolic enzymes. The major rationales for fusions between metabolic genes appear to be overcoming pathway bottlenecks, avoiding toxicity, controlling competing pathways, and facilitating expression and assembly of protein complexes. Finally, our fusion dataset provides powerful clues to decipher the biological activities of domains of unknown function.