Abstract

Compressed representations of the input are a standard approach to accelerate bioinformatics algorithms. We will discuss how compression of discrete emission sequences can be used to accelerate computations with Hidden Markov Models.

For continuous-valued observation sequences, such as data obtained from SNP arrays, exome or whole-genome sequencing we introduce our compression approach based on Haar Wavelets. This makes it possible to use a full Bayesian approach for identifying Copy Number Variants (CNV) on standard computers by accelerating the Gibbs-sampling scheme to the point of exceeding the speed of maximum-likelihood methods, maintaining the higher fidelity of the Bayesian approach. Current work focuses on extending the method to joint analysis of CNV data in trios, pedigrees, or across multiple conditions, for example yielding posterior probabilities of de novo CNVs in children affected with autism.