We described also a procedure for selecting a partition among a set of candidate ones and a measure of cluster reliability. Note that our definition of the likelihood of a partition is not an absolute measure of the “goodness” of the partition, but expresses how much a partition is similar to all the generated partitions and is of use for assessing cluster reliability. In principle, the TBC might be applied also to next-generation sequencing data alignments, and the PLX-4720 sequence length should not be a serious problem if Roche 454 Life Science technology is used to amplify specific regions. With shotgun sequencing the data need to be analysed via sliding windows, and then a problem of variant reconstruction arises. We do not know how the TBC would behave in presence of sequencing errors that should be -at least- corrected before running the clustering. However, another approach that uses the Chinese restaurant process as a prior to infer clusters via Gibbs’ sampling has been recently proposed, and performs both clustering and error correction at the same time. TBC was also evaluated in the problem of transmission event identification. Using a data set of patients followed up at the Catholic University of Sacred Heart in Rome, Italy, with known transmission history, TBC was able to identify transmission events in 25% of cases, whilst CTree assessed on 16.7%. The transmission event dataset was also evaluated using a previously published method, specifically tuned for HIV transmission cluster identification, and that method identified 25% of transmission events. With a human-visual evaluation of subtrees and node reliability of a maximum-likelihood phylogenetic tree, we were able to infer correctly 50% of transmission events. Thus, even a detailed phylogenetic analysis was not able to resolve all transmission events. In fact, for HIV it has been shown previously that many factors can limit the concordance of phylogenetic reconstruction and the reported epidemiological evidence. The transmission event data set of CUSH was composed by sequence samples of patients taken at different times and disease stage: some patients were sequenced multiple times either before treatment initiation or at treatment failures, whilst others had only one sequence sample taken. We recognise that a larger and less sparse data set would be desirable in order to assess better the TBC performance on this particular problem. A way to optimise the percentile threshold -without knowing a priori the sequence grouping is to run the TBC using different thresholds and then calculate for each partition a cluster validity measure, such as the Goodman-Kruskal index, the Dunn’s index, or the average silhouette value.