Cancer prediction with gene expression profiling and differential evolution

In the field of bioinformatics, the classification of tumors is a difficult and time-consuming task. When diagnosing cancer, gene expression levels are typically one of the most useful tools. However, the biological noise present in microarray data leads to unsatisfactory precision and accuracy. The utilization of thousands of genes in the process of diagnosing tumors is an important task. The two levels of feature selection have been proposed in order to determine the genes that are the most informative to diagnose cancer. Using three different statistical methods, the first level of selection reveals the prognostic genes. In the second level, the differential evolution algorithm considers the prognostic genes that were obtained from statistical measures as initial members to identify the most relevant features. The scaling factor in the modified differential evolution algorithm was made to vary in a dynamic manner in order to evolve the mutant member of the population. The proposed model is a hybrid of statistical approach and evolutionary computation with modified differential evolution algorithm that identifies the candidate genes from thousands of genes from gene expression data. The findings obtained through this hybrid approach upon testing five gene expression datasets provide evidence that it has outperformed when compared to the existing systems for DLBCL outcome, prostate outcome, prostate, and colon tumor datasets with improved classification accuracies of 14%, 4%, 0.62%, and 0.13%, respectively.