Markov model variants for appraisal of coding potential in plant DNA

Markov chain models are commonly used for content-based appraisal of coding potential in genomic DNA. The ability of these models to distinguish coding from non-coding sequences depends on the method of parameter estimation, the validity of the estimated parameters for the species of interest, and the extent to which oligomer usage characterizes coding potential. We assessed performances of Markov chain models in two model plant species, Arabidopsis and rice, comparing canonical fixed-order, χ2-interpolated, and top-down and bottom-up deleted interpolated Markov models. All methods achieved comparable identification accuracies, with differences usually within statistical error. Because classification performance is related to G+C composition, we also considered a strategy where training and test data are first partitioned by G+C content. All methods demonstrated considerable gains in accuracy under this approach, especially in rice. The methods studied were implemented in the C programming language and organized into a library, IMMpractical, distributed under the GNU LGPL.