A Probabilistic Genome-Wide Gene Reading Frame Sequence Model

We introduce a new type of probabilistic sequence model, that model the sequential composition of reading frames of genes in a genome. Our approach extends gene finders with a model of the sequential composition of genes at the genome-level – effectively producing a sequential genome annotation as output. The model can be used to obtain the most probable genome annotation based on a combination of i: a gene finder score of each gene candidate and ii: the sequence of the reading frames of gene candidates through a genome. The model — as well as a higher order variant — is developed and tested using the probabilistic logic programming language and machine learning system PRISM a fast and efficient model prototyping environment, using bacterial gene finding performance as a benchmark of signal strength. The model is used to prune a set of gene predictions from an underlying gene finder and are evaluated by the effect on prediction performance. Since bacterial gene finding to a large extent is a solved problem it forms an ideal proving ground for evaluating the explicit modeling of larger scale gene sequence composition of genomes. We conclude that the sequential composition of gene reading frames is a consistent signal present in bacterial genomes and that it can be effectively modeled with probabilistic sequence models.