Gene expression patterns can distinguish risk groups in training cohort. (A) Heat maps of the 70 genes illustrate remarkably similar expression patterns among 351 newly diagnosed patients used to identify the 70 genes. Red bars above the patient columns denote patients with disease-related deaths. The 51 genes in rows designated by the red bar on the left (top rows; up-regulated) identified patients in the upper quartile of expression at high risk for early disease-related death. The 19 gene rows designated by the green bar (down-regulated), identified patients in the lower quartile of expression at high risk of early disease-related death. (B) Training cohort frequencies for sample differences between ratios of the mean of log₂ expression of the 51 up-regulated genes/19 down-regulated genes. This self-normalizing expression ratio has a marked bimodal distribution, consistent with the upper/lower quartile log-rank differential expression analysis, which was designed to detect genes that define a single high-risk group (13.1%) with an extreme expression distribution. Interpreted as an up/down-regulation ratio on the log₂ scale, higher values are associated with poor outcome. The vertical line shows the high-risk versus low-risk cutoff for the log₂-scale ratio determined by K-means clustering: the percentage of samples below and above the cutoff is also shown. Kaplan-Meier estimates of EFS (C) and OS (D) in low-risk myeloma (green) and high-risk myeloma (red) showed inferior 5-year actuarial probabilities of EFS (18% vs 60%, P < .001; HR = 4.51) and OS (28% vs 78%, P < .001; HR = 5.16) in the 13.1% patients with a high-risk signature.