DOE Human Genome Program Contractor-Grantee Workshop VIII
February 27-March 2, 2000  Santa Fe, NM

91. Splice Site Recognition

Terry Speed and Simon Cawley

University of California at Berkeley, Berkeley, CA 94720-3860

With the increasing abundance of completely sequenced genomes the automation of genome annotation has become an important research goal. We focus on the classification of splice sites in eukaryotic genes, an integral sub-task in most successful genefinding programs. In particular we focus on probabilistic models for splice sites, since they can be readily incorporated into probabilistic genefinders without having to worry about how to weight the evidence of splice site classifiers. We make use of variable length Markov chains (also known as context models). VLMCs can capture long-range dependencies in splice sites without having the usual problem of exponential increase in the number of parameters encountered with regular Markov models. We compare these VLMCs with existing splice site recognition methods, both as a stand-alone problem and within PfParser, a hidden Markov model genefinding program for Plasmodium falciparum (a Malaria parasite).