Fast Bayesian Haplotype Inference Via Context Tree Weighting
We present a new, Bayesian method for inferring haplotypes for unphased genotypes. The method can be viewed as a unification of some ideas of variable-order Markov chain modelling and ensemble learning that so far have been implemented only separately in some of the state-of-the-art methods. Specifically, we make use of the Context Tree Weighting algorithm to efficiently compute the posterior probability of any given haplotype assignment; we employ a simulated annealing scheme to rapidly find several local optima of the posterior; and we sketch a full Bayesian analogue, in which a weighted sample of haplotype assignments is drawn to summarize the posterior distribution. We also show that one can minimize in linear time the average switch distance, a popular measure of phasing accuracy, to a given (weighted) sample of haplotype assignments. We demonstrate empirically that the presented method typically performs as well as the leading fast haplotype inference methods, and sometimes better. The methods are freely available in a computer program BACH (Bayesian Context-based Haplotyping)
Unable to display preview. Download preview PDF.