Background: Microarray technologies provide huge amount of information
and is particularly helpful in the prediction and diagnosis of cancer. To accurately
classify cancers, genes related to cancer have to be selected, as genes mined from
microarrays possess too much noise.
Method: In the current work, new binary modifications of the Grey Wolf Optimization
(GWO) is suggested for choosing optimal features subsets for classification. In the
proposed approach, GWO is modified by binarizing only the initial three optimal
solutions and updation of the wolf position using stochastic crossover. Modification
was also carried out using sigmoidal functions to compress the continuous updated
positions. Multilayer Perceptron – Neural Network (MLP-NN) classifier is used for
classifying the selected features.
Results: The ALL/AML Leukemia dataset is used for evaluating Markov Blanket filter, minimum
Redundancy Maximum Relevance (mRMR), Binary GWO and Mutated Binary GWO (MBGWO) with
regard to classification accuracy, sensitivity and specificity. The proposed MBGWO achieved
classification accuracy of 95.45% and also has better sensitivity and specificity.
Conclusion: Experiments reveal the capabilities of the proposed MBGWO to explore features space for
the optimal features combination for gene selection from microarray data.