In addition to improving the analysis at the variable selection step, our approach differs from our previous studies (Chan et al., 2011; Gelernter et al., 2006) by replacing k-means cluster analysis, which uses several randomly chosen starting points, with a k-medoids method. Although repeating the k-means analyses with several starting points improves the stability of the resultant clusters, they may not be replicable at different runs of the clustering process. By using k-means analysis to create 50 clusters at each run and repeating it 10 times, 5010 cells have to be cross tagged to find stable clusters, requiring extensive computation. These 5010 cells may differ with different runs due to the randomness of starting points, leading to different cluster solutions. An information-theoretic criterion (Kaufman and Rousseeuw, 1990), such as the one used here, can select the initial points for the k-medoids analysis. Thus, the clusters derived using this approach do not vary when the analysis is run multiple times.
