Improving the Prediction Accuracy of Gene Structures in Eukaryotic DNA with Low C+G Contents

have developed a gene prediction program GeneKey. When trained with the widely used dataset lected by Kulp and Reese, GeneKey can achieve high prediction accuracy for genes with derate and high C+G contents. However, the prediction accuracy is much lower for CG-poor es. To tackle this problem, we construct a new LCG316 dataset composed of gene sequences h low C+G contents. For CG-poor genes, the prediction accuracy of GeneKey when trained with G316 dataset has been improved prominently. Further statistical analysis demonstrates that some cture features, such as splicing signals and codon usage, of CG-poor genes are quite different m that of CG-rich ones. The combination of the two datasets enables GeneKey to get high and anced prediction accuracy for both CG-rich and CG-poor genes. The results of this work imply t careful construction of training dataset is very important for improving the performance of ious prediction tasks. The GeneKey program is available at http://infosci.hust.edu.cn. ywords: DNA sequence, prediction of gene structure, prediction of protein coding region