Integration of ab initio Gene Finding with Mapping of Homologous Proteins

"Spliced alignments" of proteins to eukaryotic genome sequences provide important external evidence that is frequently integrated into algorithms of structural genome annotation.  However, the accuracy of "spliced alignments" in identification of gene elements, exons and introns, depends critically on the evolutionary distance between the species from which the protein and the genome originated from, i.e. the accuracy drops down fast if remotely related species are considered. Still some of the gene structure elements can be mapped correctly even if the protein and the genome come from a remotely related species.

We present a new approach to integration of evidence data derived from protein to genome alignment into the training process of an ab initio gene prediction algorithm.  This approach filters out presumably incorrect elements of gene structure from the whole set of elements mapped by a large number of spliced alignments of potential homologous proteins. The algorithm, called GeneMark-EP, can incorporate existing tools of spliced alignment, e.g. ProSplign or Exonerate. This new algorithm with semi-supervised training was tested on several model and non-model genomes and demonstrated improved accuracy on the non-model genomes where pure unsupervised training met difficulties due to complexity of genome organization.