Validation of statistically reliable biomarkers
Marc Buyse
IDDI, Belgium
Definitions
Biomarkers play an increasing role in the development of new cancer treatments. A biomarker is defined as "a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacologic responses to a therapeutic intervention". The term biomarker covers characteristics measured at baseline as well as those measured repeatedly over time, before, during or after treatment. Clinical data, laboratory data, imaging data, gene expression and proteomic data can all be considered potential biomarkers.
Uses of biomarkers
Biomarkers can be useful as:
- prognostic factors that predict the outcome of individual patients in terms of a clinical endpoint
- predictive factors that predict the effect of a specific treatment on a clinical endpoint in groups of patients
- surrogate endpoints that replace a clinical endpoint of interest
Biomarkers can be used to stratify the patients at entry in clinical trials, to select the patients eligible for clinical trials, to monitor patients and guide treatment decisions, or to substitute for a clinical endpoint in the evaluation of the effects of new treatments.
Requirements for biomarkers
The clinical literature is replete with examples of the use of biomarkers, but in many cases these have not been properly identified and validated, resulting in a large number of false claims, inappropriate trial designs, and ultimately sub-optimal use of biomarkers for patient management.
- Requirements for prognostic biomarkers: The baseline value of the biomarker, or changes in the biomarker over time, should be correlated with the clinical endpoint in untreated or in treated patients
- Requirements for predictive biomarkers: The baseline value of the biomarker, or changes in the biomarker over time, should be correlated with the effect of treatment on the clinical endpoint
- Requirements for surrogate biomarkers: The difference in the biomarker values between two randomized treatments should be correlated with the difference in the clinical endpoint
Study designs
Different study designs are required for the identification of biomarkers. Case-control or cohort studies are sufficient to identify prognostic biomarkers, large randomized trials are needed to identify predictive biomarkers, and multiple randomized trials are needed to identify surrogate biomarkers. In all cases, the biomarker should be validated either through cross-validation (internal validation in the discovery set) or in different trials (external validation in a confirmatory set).
Validation criteria
The criteria used to validate biomarkers include classification measures (sensitivity and specificity, ROC curves), treatment effect measures (odds ratios, hazard ratios) association measures (correlation coefficients, information theory based measures), and prediction measures (the surrogate threshold effect). These various criteria all have advantages and limitations that will be illustrated with several examples in the field of cancer.