After examining several combinations of input parameters, we found results to be relatively selleck chemicals consistent across input parameters and selected results from c 3 and m 50 for further analysis of the irradiated data, where c indicates units of change and m, the number of candidate profiles. This run significantly clustered 174 out of the 238 cases. Figure 2 shows gene expression pro files for the six clusters found to be significant out of 50 possible clusters. The Rand Index to the manually curated clustering was 0. 64, indicating good similarity. The cardinality of each cluster was relatively uniform, ranging from 18 genes in cluster 6 to 37 genes in Cluster 1. Visual examination of the clusters suggested that biphasic responding genes were distributed across the first four clusters and that Cluster 3 also included genes that showed the more gradual increase, which peaked at 4 to 6 hours.

STEM also clustered down regulated genes into a separate cluster, Cluster 5. Gene expression of the 238 genes differentially expressed after irradiation was also clustered using FBPA on gene expression data features. To determine the optimal number of clusters, we used the gap statistic. Where k is the number of clusters, we examined k 4, 8, and 11, which all showed near zero inequalities. The average homogeneity was 3. 026 and the average silhouette was 0. 558 for k 4. For k 8, the average homogeneity was 2. 098 and average silhouette was 0. 434. With k 11, average homogeneity was 1. 764 and average silhouette was 0. 371. Because good homogeneity and strong separation and structure were found with k 4, we chose this clustering.

We note here that we tended towards parsimonious clustering as much as possible to avoid over fitting the data and to group information that may be biologically relevant. The Rand Index to the manually curated standard was 0. 623 also indicating good similarity, equivalent to that of STEM clustering on the microarray data after irradiation. Figure 4 shows the gene expression profiles clustered using FBPA. The within method metrics gave interesting information. Because the method chose a small number of clusters, homogeneity was not strong, with the average homogeneity being close to 3. How ever, all but Cluster 3 showed good separation. The average silhouette over all clusters was 0. 558 indicating that strong structure was found.

We also Entinostat noted that genes were not uniformly distributed across all clusters. In irradiated samples, 61% of the total number of genes clustered belonged to Cluster 1, 24% to Cluster 2, 13% to Cluster 3 and 2% belonged to Cluster 4. Given that these genes were pre selected on the basis of response at 4 hours, the clustering of a large proportion of genes together in one cluster in directly irradiated cells is not unexpected, because cells respond robustly to irradiation and transcripts of many of the genes included in this study could be affected in concert.