Prediction of Mycobacterium Tuberculosis Lineages from Annotated Whole Genome Sequences: An Evolutionary Approach

In this paper, we present an original evolutionary approach for Mycobacterium tuberculosis lineages classification. We use a protein frequency analysis in order to transform the genomic source data into more usable forms. First we consider the number of occurrences of each protein in each tuberculosis genome, then we only consider the presence or absence of proteins in a genome and generate binary sequences. Depending on their frequency of apparition, some proteins are discarded in order to reduce the number of predictive attributes. Classifiers composed of multiple binary masks are evolved through a dedicated evolutionary algorithm in order to reach optimal classification performances. We show that the performances obtained by such classifiers on binary presence sequences are quite good in comparison to the performances of classical machine learning techniques such as decision trees on protein occurrence data. All models investigated in this work are easily interpretable as it is important to get some insight about which proteins are important for prediction.