The parameters of our model are learnt from the data using an EM algorithm (27) and, given these parameters, the maximum likelihood sequence of hidden states are inferred using the Viterbi algorithm (28). In our analysis, we apply the EM and Viterbi algorithms to one chromosome at a time. Identifiability of the states is maintained via our non-symmetric prior distribution structure for the B allele frequencies. The number of mixture components is conditioned on the hidden states and therefore arbitrary re-labelling is not possible. We assign a Bayes Factor to each region of copy number variation detected. This provides a probability measure of the strength of evidence from the data for the presence of a copy number variant in a region versus the null hypothesis that there is no variant. The greater the value of the Bayes Factor, the stronger the evidence for the existence of a copy number variant.
