G B Hutchinson1. 1. Department of Medical Genetics, University of British Columbia, Vancouver, Canada. hutch@netshop.bc.ca
Abstract
MOTIVATION: To develop an algorithm utilizing differential hexamer frequency analysis to discriminate promoter from non-promoter regions in vertebrate DNA sequence, without relying upon an extensive database of known transcriptional elements. RESULTS: By determining hexamer frequencies derived from known promoter regions, coding regions and non-coding regions in vertebrates' DNA sequence, and a formula first applied by Claverie and Bougueleret (1986), a discriminant measure was created that compares promoter regions with coding (D1) and non-coding (D2) sequence. The algorithm is able to identify correctly the promoter regions in 18 of 29 loci (62.1%) from an independent test data set. With program options set to identify only one promoter region in the forward strand, there are 11 false-positive predictions in 208 714 nucleotides (one false positive in 18 974 single-stranded bp). With options set to analyze sequence in discrete segments, there is no appreciable improvement in sensitivity, whereas the specificity falls off predictably. It is of particular interest than a search for a peak score (independent of an absolute threshold) is more accurate that a search based upon a fixed scoring threshold. This suggests that the selection of promoter sites may be influenced by the global properties of an entire sequence domain, rather than exclusively upon local characteristics.
MOTIVATION: To develop an algorithm utilizing differential hexamer frequency analysis to discriminate promoter from non-promoter regions in vertebrate DNA sequence, without relying upon an extensive database of known transcriptional elements. RESULTS: By determining hexamer frequencies derived from known promoter regions, coding regions and non-coding regions in vertebrates' DNA sequence, and a formula first applied by Claverie and Bougueleret (1986), a discriminant measure was created that compares promoter regions with coding (D1) and non-coding (D2) sequence. The algorithm is able to identify correctly the promoter regions in 18 of 29 loci (62.1%) from an independent test data set. With program options set to identify only one promoter region in the forward strand, there are 11 false-positive predictions in 208 714 nucleotides (one false positive in 18 974 single-stranded bp). With options set to analyze sequence in discrete segments, there is no appreciable improvement in sensitivity, whereas the specificity falls off predictably. It is of particular interest than a search for a peak score (independent of an absolute threshold) is more accurate that a search based upon a fixed scoring threshold. This suggests that the selection of promoter sites may be influenced by the global properties of an entire sequence domain, rather than exclusively upon local characteristics.