Abstract. The EM algorithm is considered for the problem of separating probability distribution mixtures with components described by Markov chains, together with the related weighted log likelihood maximization problem. Auxiliary algorithms to select initial approximation and optimal mixture size are proposed, as well as a method for approximating the mixture with given data using support vector machines. The results are applied to boost the quality of gene fragment classifiers.

Keywords: Markov chain, classification, gene, bioinformatics, nucleotide, exon, intron, likelihood.