To demonstrate that a biomarker is predictive of treatment benefit, the study requires biomarker status on all patients as well as patients who were treated with the agent of interest and patients not so treated, preferably in the context of a randomized study. A formal statistical test of the treatment-by-biomarker interaction should be significant. To establish whether a marker is purely prognostic, it needs to be demonstrated that there is a significant association between the biomarker and outcome, regardless of treatment, and that treatment effects do not depend on the biomarker. Finally, a biomarker may have both predictive and prognostic implications. These concepts are summarized in Figure 2.