Locating CNV candidates in WGS data using wavelet-compressed Bayesian HMM
John Wiedenhoeft, Alex Cagan, R. V. Kozhemjakina, Rimma G. Gulevich, Alexander Schliep · 2017
The avalanche of NGS data and the growing demand for Bayesian methods pose huge algorithmic challenges in the case of whole-genome CNV inference. At the same time, fast, accurate and efficient computation is crucial in both clinical and fundamental research settings. Recently, Wiedenhoeft, Brugel, and Schliep (2016) presented a method to drastically improve computation of full latent state marginals of Bayesian HMM in terms of speed and convergence behavior, but handling the memory requirements due to the sheer size of the input remained challenging. We present an improved implementation of HaMMLET, a wavelet-compressed Forward-Backward Gibbs sampler for Bayesian HMM. We present a new data structure for dynamic compression, which can be constructed in-place and in linear time. We demonstrate its application for CNV inference on rat populations divergently selected for tame and aggressive behavior.