Incorporation of splice site probability models for non-canonical introns improves gene structure prediction in plants

Abstract

Motivation: The vast majority of introns in protein-coding genes of higher eukaryotes have a GT dinucleotide at their 5′-terminus and an AG dinucleotide at their 3′ end. About 1-2% of introns are non-canonical, with the most abundant subtype of non-canonical introns being characterized by GC and AG dinucleotides at their 5′- and 3′-termini, respectively. Most current gene prediction software, whether based on ab initio or spliced alignment approaches, does not include explicit models for non-canonical introns or may exclude their prediction altogether. With present amounts of genome and transcript data, it is now possible to apply statistical methodology to non-canonical splice site prediction. We pursued one such approach and describe the training and implementation of GC-donor splice site models for Arabidopsis and rice, with the goal of exploring whether specific modeling of non-canonical introns can enhance gene structure prediction accuracy. Results: Our results indicate that the incorporation of non- canonical splice site models yields dramatic improvements in annotating genes containing GC-AG and AT-AC non-canonical introns. Comparison of models shows differences between monocot and dicot species, but also suggests GC intron-specific biases independent of taxonomic clade. We also present evidence that GC-AG introns occur preferentially in genes with atypically high exon counts. © The Author 2005. Published by Oxford University Press. All rights reserved.