Integration of gene predictions using artificial neural networks.

摘要

In eukaryotes, especially in human beings, only a small proportion of genomic DNA sequence consists of functional fragments that are called exons. Exons are separated by non-functional fragments called introns. Many whole-genome sequences are available in the public domain and accessible over the World Wide Web. The challenge to scientists today is to correctly identify genes and their functions from these genomic sequences. Although many prediction engines are available over the Web to facilitate such identification, their scope of application and their capacity for prediction are limited. This thesis integrates three of the most prominent such engines, GrailExp, GenScan, and MZEF using a Multilayer Perceptron and a Mixture of Experts neural networks, in order to improve the capability and confidence of prediction.; The system was trained using 575 predictions that are mapped to known target values from 33 human genomic sequences, and tested using the prediction results from another unrelated set of 28 human genomic sequences. This thesis has identified two major drawbacks to the accuracy of prediction by individual engines, (i) contradictory predictions even within an engine, and (ii) inconsistency of prediction between the engines. Analysis of variance was performed over the result, and demonstrates that the integration system has significantly better recovery, by 25% on average, than individual prediction engines. The system based on a multilayers perceptron is available for exon prediction of human genomic DNA at http://www.cbr.nrc.ca/pany/integ.html.