The CvN method improves classification of NK-AML patients. Side-by-side comparison of clustering performance of the CvC (A-C) and CvN (D-F) methods on a NK-AML data set (GSE15434). Heat maps (hierarchical clustering) of genes identified by the CvC method (A) and CvN methods (D), using a NK-AML patient data set. Differentially expressed genes identified by each method were selected by variance (1614 and 1383 probe sets in A and D, respectively) and rescaled gene wise. An initial hierarchical clustering was used to identify the optimal number of patient clusters (n = 6; supplemental Figure 9). This was followed by K-means clustering (K = 6), which distributed the samples into 6 patient clusters (color labeled). (B,E) 3-dimensional-PCA plots of the 6 K-means-derived patient clusters identified by the CvC (B) and CvN (E). (C,F) Kaplan-Meier plots depicting the OS curves for of the 6 NK-AML clusters assessed by (C) the CvC method and (F) the CvN method (P = .04 and P = .007, respectively, χ-square). (G) Median gene expression fold change of selected MsigDB gene signatures that overlap significantly (P < 1e−5, median, subclass-wise) with patient-specific NK-AML signatures.