摘要
Gene finding, the accurate annotation of genomic DNA, has become one of the central topics in biological research. Although various computational methods (gene finders) have been proposed and developed, they all have their own limitations in gene findings. In this paper, we introduce an integrating gene finder, which combines the results of several existing gene finders together, to improve the accuracy of gene finding. Four integration schemes, based on majority voting, are developed for the analysis of two datasets – the basic dataset and the testing dataset. The basic dataset consists of 1500 DNA sequences and the testing dataset consists of 103 DNA sequences. It is demonstrated that a simple integration (a simple voting for each nucleotide) can significantly improve the finding performance, and removing confusing gene finders, caused by poor performance or redundant results, is important for a further improvement of the integration. The best prediction results are obtained using weighted majority voting, aided by the mRMR (Minimum Redundancy Maximum Relevance) (Peng, 2005) method for the gene finder selection. The prediction accuracies are 84.16% and 90.06% for the basic dataset and testing dataset respectively, which are better than any individual gene finding software in our research.
Gene finding, the accurate annotation of genomic DNA, has become one of the central topics in biological research. Although various computational methods (gene finders) have been proposed and developed, they all have their own limitations in gene findings. In this paper, we introduce an integrating gene finder, which combines the results of several existing gene finders together, to improve the accuracy of gene finding. Four integration schemes, based on majority voting, are developed for the analysis of two datasets – the basic dataset and the testing dataset. The basic dataset consists of 1500 DNA sequences and the testing dataset consists of 103 DNA sequences. It is demonstrated that a simple integration (a simple voting for each nucleotide) can significantly improve the finding performance, and removing confusing gene finders, caused by poor performance or redundant results, is important for a further improvement of the integration. The best prediction results are obtained using weighted majority voting, aided by the mRMR (Minimum Redundancy Maximum Relevance) (Peng, 2005) method for the gene finder selection. The prediction accuracies are 84.16% and 90.06% for the basic dataset and testing dataset respectively, which are better than any individual gene finding software in our research.
