BClass: A Bayesian Approach Based on Mixture Models for Clustering and Classification of Heterogeneous Biological Data

Abstract

Based on mixture models, we present a Bayesian method (called BClass) to classify biological entities (e.g. genes) when variables of quite heterogeneous nature are analyzed. Various statistical distributions are used to model the continuous/categorical data commonly produced by genetic experiments and large-scale genomic projects. We calculate the posterior probability of each entry to belong to each element (group) in the mixture. In this way, an original set of heterogeneous variables is transformed into a set of purely homogeneous characteristics represented by the probabilities of each entry to belong to the groups. The number of groups in the analysis is controlled dynamically by rendering the groups as 'alive' and 'dormant' depending upon the number of entities classified within them. Using standard Metropolis-Hastings and Gibbs sampling algorithms, we constructed a sampler to approximate posterior moments and grouping probabilities. Since this method does not require the definition of similarity measures, it is especially suitable for data mining and knowledge discovery in biological databases. We applied BClass to classify genes in RegulonDB, a database specialized in information about the transcriptional regulation of gene expression in the bacterium Escherichia coli. The classification obtained is consistent with current knowledge and allowed prediction of missing values for a number of genes.

BClass is object-oriented and fully programmed in Lisp-Stat. The output grouping probabilities are analyzed and interpreted using graphical (dynamically linked plots) and query-based approaches. We discuss the advantages of using Lisp-Stat as a programming language as well as the problems we faced when the data volume increased exponentially due to the ever-growing number of genomic projects.