Neural networks for genome signature analysis

Neural networks show promise for mitigating the combinatorial explosion in genomic data. Researchers are interested in the applicability of neural networks for the design of automatic genomic analysis tools. This: paper describes the application of a variety of neural network models, including back-propagation, radial basis function networks, self-organizing maps, and committee machines, to the problem of gene classification using genome signatures. Results shows that in a two-way classification problem, average accuracies of 97% can be attained with these models, while for a more difficult four-way classification task average accuracy was more than 83%. Methods for developing the training and test data for the signature problem are discussed, as well as modifications to the general algorithms of the neural network models.