Prediction of Protein Interdomain Linker Regions by a Nonstationary Hidden Markov Model

Abstract

Proteins have two types of regions: linker regions that connect active domains and nonlinker regions, which include domains and terminal residues. It is important to try to distinguish these regions using information in the sequence of amino acids that, when folded, form the protein. In this article we develop a nonstationary hidden Markov model to infer region boundaries. The approach is based on a latent variable defined on the sequence; this variable is a continuous index of the region type and affects the probabilities that specific amino acids will appear. We develop an efficient Bayesian estimate of the model using Markov chain Monte Carlo methods, particularly Gibbs sampling, to simulate the parameters from the posteriors. We apply our method to protein sequences with domains and interdomain linkers delineated using the Pfam-A database. The prediction results are superior to simpler methods. Importantly, our method estimates the probability that each amino acid belongs to a linker region, giving insight into the properties of interdomain linkers.