(A, B) Principal component analysis (PCA) plots indicating different K-means cluster configurations, using k = 3 and k = 4 clusters, respectively, after performing bootstrapping. With k = 3, different starting points give different clusters. The two most common clusters (top row) are very similar, and they are obtained in 241 and 179 starts out of 1000, respectively. However, the clustering that best represents the data when k = 3 is the third one found in 168/1000 starting points as its withinss (wss) metric is lower (highlighted in red). Indeed, this configuration is more equivalent to those clustering configurations when k = 2. (B) Clusters seem more stable when k = 4. Accordingly, the best clustering appears to be the ones represented in the bottom row, which contains two main groups and two small groups with just two patients. (C, D) The second method used was the Monte Carlo reference-based consensus clustering (M3C), which also indicated that k = 2 is the optimal number of clusters, as indicated in (C) the flat line in the CDF plot and (D) in the highest relative cluster stability index (RCSI) plot. (E–G) Using spectral clusters, instead of elliptical K-means clusters, M3C analysis indicates that k = 3 gives the optimal number of clusters, as indicated in the (E) CDF plot, (F) RCSI plot, and (G) the NXN consensus matrix, where each element represents the fraction of times two samples clustered together.