A flocking based data mining algorithm for detecting outliers in cancer gene expression microarray data

The existence of outliers is a major factor of inaccuracy in cancer gene expression microarray-based experiments. Researchers confirm that in many cases outliers in one class in cancer microarray based-experiments are contaminated. As a result, outliers appear to have gene expression similar to samples of an existing class in the dataset. Hence, it is essential to analyze each class in the dataset independently from other classes. Existing outlier detection algorithms identify outliers with respect to the whole dataset. Our algorithm isolates detected classes and analyzes each class as a separate dataset. We propose a novel, simple and biologically inspired algorithm to detect outliers in cancer microarray data. This algorithm is inspired from the natural phenomena of bird flocking. We model microarray gene expression data as an artificial life where similar samples flock in a virtual space to form swarms and outliers' samples are being naturally repulsed by optimum subswarms. We demonstrate empirically that our algorithm detects biologically meaningful outlier samples. We analyze the performance of the algorithm using real colon cancer dataset widely used in the bioinformatics literature.