Following two consensus heatmaps (Figure S3.1) are for \(k\) = 2 and 3 (2 and 3 groups) on GDS1875 GDS dataset under top-value method ATC and partitioning method hclust (cola report is here). Figure S3.1 illustrates that when the number of subgroups increases form 2 to 3, a tiny subgroup only with one single sample is separated from one of the big subgroups.
Figure S3.1. Consensus heatmaps for k = 2 and 3.
Next we plot the 1-PAC scores as well Jaccard indices for the partitions under each \(k\) (Figure S3.2). We can see the 1-PAC scores are both very high at \(k\) = 2 and 3, which means the partitions are both very stable for the two \(k\). However, the Jaccard index is extremely high at \(k\) = 3 (0.97), which quantitatively shows partitions in \(k\) = 3 is highly similar as \(k\) = 2, and on the other hand, an extra subgroup only with one single sample does not provide any useful output for the analysis. Thus, we decline \(k\) = 3 is a better choice while still take \(k\) = 2 as the best number of subgroups.
Figure S3.2. The 1-PAC scores as well as Jaccard indices for the partitions under each k.