Computational Prediction of Replication Origins in Herpesviruses

Computational methods for replication origin prediction in individual herpesvirus genomes have been previously devised based on the locations of high concentrations of palindromes. In order to make use of similarities in genome composition and organization of related herpesviruses, an artificial neural network approach is explored. We implement feed-forward artificial neural networks trained by 17 input variables comprising the positions of known replication origins relative to the genome lengths and the dinucleotide scores. The overall prediction accuracy of the neural network approach for our data set is better than that of the palindrome based approach. Furthermore, suitable combinations of the prediction results given by the two approaches substantially increase the prediction accuracy achieved by either method applied individually.