Two-dimensional hierarchical heatmaps showing classifier discovered from FFPET samples accurately predicts class labels in agreement with FF. Top panel: samples in FFPET are ranked by classifier score. Leave-one-out cross validation error rate is ~14% (p-value of 3.7 × 10-7). Bottom panel: FF samples are clustered based on the "prognostic" genes derived from FFPET. The clustering analysis can group all samples correctly except one. The right panel indicates the relative truth as shown in Figure 7.