Artificial neural networks methods for identification of the most relevant genes from gene expression array data

Gene array studies can assess the global expression patterns of thousands of genes under multiple conditions. The paper demonstrates the application of large-scale artificial neural networks (ANNs) for gene array analysis and cancer cell identification by training an ANN model on data generated by gene array experiments involving four different small round blue-cell tumors. Recursive input pruning of the ANN model and re-training techniques were used to identify the more relevant genes. Out of the original list of 2308 genes, an ANN model with 9 gene inputs and 5 neurons in the hidden layer correctly classified the four cancer cell types in the training set with only two possible misclassifications in the validation set. These misclassifications were identified as questionable by pattern analysis of the outputs of the ANN hidden neurons. Causal index evaluation shows the influence of each of the identified most relevant genes on the cancer cell classification; this constitutes important new knowledge extracted from the ANN model.