36 Exon Prediction: A Neural Network Approach

Exon prediction, including the use of complex multilayer neural networks, is an emerging field that has impacts in medicine, drug development, and law enforcement (Asyali, et al. 2006). However, few research projects have considered simple neural networks for human exon prediction. It is anticipated that the simplicity of two-layer neural networks can greatly reduce the complexity of existing exon prediction algorithms. A data set was compiled based on quantitative data collected on TATA boxes, TG-dinucleotides, CpG island qualities, poly-A tails, and codon bias. Results show that the proposed method hasprediction accuracy greater than 63% and a false positive rate less than 25% for all but one network. These results are an improvement from previous neural network gene prediction studies (Uberbacher and Mural, 1991). A simple Gaussian classifier, the nearest mean classifier (NMC), and principal components analysis (PCA) were used to explore the structure of the gene data set. The PCA results indicate that CpG islands have little effect on the existence of exons. This provides evidence supporting the lack of correlation between CpG islands and coding regions, a previously ambiguous relationship.