Probabilistic Methods for Computational Annotation of Genomic Sequences

Abstract

New sequencing techniques have increased enormously the speed that new genomic sequences are produced. As manual inspection is impossible for this amount of data, there is an ongoing need for computational tools that can annotate this data efficiently and accurately. Essential parts of the annotation process of genomes are the prediction of protein-coding genes, and the classification of the obtained protein sequences according to their function. Currently, computational predictions are not accurate enough to be considered overall reliable. At the same time that new data is produced that needs to be analysed, the amount of available data that can be used to guide the prediction is growing as well. In particular, databases containing annotated proteins and functional descriptions of protein families, are widespread and easily accessible, and can provide additional input to gene prediction programs. In the focus of this thesis is the introduction of a new method that uses protein profiles that can be generated from a set of related proteins to improve the accuracy of present gene prediction methods. It was implemented as an extension to the gene prediction program AUGUSTUS, called the ``Protein Profile Extension'' (PPX). Since a correct classification of protein sequences relies on accurate gene predictions especially of regions typical for a class or family, this method can be viewed a combination of gene prediction and protein classification that is designed to improve classification rates. Both gene prediction and protein classification commonly evaluate sequences based on probabilistic models, identifying sequences that have a high probability under the model. All these models have in common the Markov property, stating that the direct neighbourhood determines the sequence composition at specific location, without long-distance dependencies. The thesis describes the specific models used in the presented methods.