Motivation: Gene fusion is an important evolutionary process. It can yield valuable information to infer the interactions and functions of proteins. Fused genes have been identified as non-transitive patterns of similarity in triplets of genes. To be computationally tractable, this approach usually imposes an a priori distinction between a dataset in which fused genes are searched for, and a dataset that may have provided genetic material for fusion. This reduces the 'genetic space' in which fusion can be discovered, as only a subset of triplets of genes is investigated. Moreover, this approach may have a highfalse-positive rate, and it does not identify gene families descending from a common fusion event.Results: We represent similarities between sequences as a network. This leads to an efficient formulation of previous methods of fused gene identification, which we implemented in the Python program FusedTriplets. Furthermore, we propose a new characterization of families of fused genes, as clique minimal separators of the sequence similarity network. This well-studied graph topology provides a robust and fast method of detection, well suited for automatic analyses of big datasets. We implemented this method in the Cþþ program MosaicFinder, which additionally uses local alignments to discard false-positive candidates and indicates potential fusion points. The grouping into families will help distinguish sequencing or prediction errors from real biological fusions, and it will yield additional insight into the function and history of fused genes. Availability: FusedTriplets and MosaicFinder are published under the GPL license and are freely available with their source code at this address: http://sourceforge.net/projects/mosaicfinder.