The effect of representative training dataset selection on the classification performance of the promoter sequences

Promoter prediction is an important task for genome annotation. The aim of this study is to build a classification method for promoter prediction. Base-stacking energy values of dinucleotides are used for feature extraction and Support Vector Machines (SVMs) are used for classification. Human genome promoter sequences are used as the positive training data and three types of datasets are prepared as the negative data including intergenic and transcribed sequences. Best results are achieved by selecting equal number of random sequences from intergenic and transcribed sequences while preparing the negative datasets.