Stochastic Models for Discovery of Regulatory Modules in Biological Sequences
by
Mayetri Gupta
Harvard University
Coauthors: Jun S. Liu (Harvard University)

A better understanding of gene regulatory networks could lead to momentous advances in medical research and drug discovery. Accurate discovery of regulatory binding sites (motifs), conserved patterns in DNA sequences, is a first step in this direction. In eukaryotic genomes, motif detection is a challenging problem as motifs tend to be short, variable, and occur in multi-pattern clusters (regulatory modules). In this talk we introduce a Bayesian methodology for state-space model selection and parameter estimation in hidden Markov models, and apply it to the location of regulatory modules. Given a potentially large and diverse starting set of motif profile matrices we need to find motif classes comprising the module and the location of sites. The module framework assumes an underlying Markov structure for the position of sites on the sequence and also the order of pattern types. A fast algorithm based on evolutionary Monte Carlo (Liang and Wong, 2000) enables us to detect likely clusters comprising the module and obtain improved parameter estimates. On the inferential side, determining the level of significance of the final motif alignment can be formulated as a model selection problem. The calculation of the Bayes factor is analytically intractable, and computationally infeasible. We instead propose a MAP criterion, for which a set of necessary and sufficient conditions is demonstrated to ensure correct model determination in the asymptotic limit. The performance of the new method will be demonstrated by both simulation studies and applications to bacterial and human genomes.

Date received: September 2, 2003