Hybrid Algorithm for Clustering Gene Expression Data

Microarray gene expressions provide an insight into genomic biomarkers that aid in identifying cancerous cells and normal cells. In this study, functionally related genes are identified by partitioning gene data. Clustering is an unsupervised learning technique that partition gene data into groups based on the similarity between their expression profiles. This identifies functionally related genes. In this study, a hybrid framework is designed that uses adaptive pillar clustering algorithm and genetic algorithm. A first step towards, the proposed work is the utilization of clustering technique by adaptive pillar clustering algorithm that finds cluster centroids. The centroids and its clustering elements are calculated by average mean of pairwise inner distance. The output of adaptive pillar clustering algorithm results in number of clusters which is given as input to genetic algorithm. The microarray gene expression data set considered as input is given to adaptive pillar clustering algorithm that partitions gene data into given number of clusters so that the intra-cluster similarity is maximized and inter cluster similarity is minimized. Then for each combination of clustered gene expression, the optimum cluster is found out using genetic algorithm. The genetic algorithm initializes the population with set of clusters obtained from adaptive pillar clustering algorithm. Best chromosomes with maximum fitness are selected from the selection pool to perform genetic operations like crossover and mutation. The genetic algorithm is used to search optimum clusters based on its designed fitness function. The fitness function designed minimizes the intra cluster distance and maximizes the fitness value by tailoring a parameter that includes the weightage for diseased genes. The performance of adaptive pillar algorithm was compared with existing techniques such as k-means and pillar k-means algorithm. The clusters obtained from adaptive pillar clustering algorithm achieve a maximum cluster gain of 894.84, 812.4 and 756 for leukemia, lung and thyroid gene expression data, respectively. Further, the optimal cluster obtained by hybrid framework achieves cluster accuracy of 81.3, 80.2 and 78.2 for leukemia, lung and thyroid gene expression data respectively.