Table 3.

Properties of biomarker and phenotypic measurements.

Property	Explanation	Examples and impact
Dynamism	Measurements can be static or dynamic.	A static biomarker cannot correlate with a dynamic phenotype if the dynamic phenotype cannot predict itself on repeat measurement. Heritable genotype is a static measurement, tumour somatic mutation status is dynamic. Care should be taken to check that dynamic changes during the time elapsed between a biomarker measurement and clinical correlate does not invalidate any potential association. Genotype is a good example of a static (subject level) biomarker which is often proposed to impact on plasma pharmacokinetics a dynamic (subject level) phenotype. Intra‐patient versus inter‐patient reproducibility of pharmacokinetic data will therefore aid an assessment of the likelihood of success of this approach before a single DNA sample has been isolated.
Level	Measurements can be made at the level of a thin slice of tissue, single lesion or subject level.	Many candidate predictive tumour markers are measured on single sections of tumour tissue and the results automatically extrapolated as a patient level attribute. Intra‐patient (but inter‐sample and inter‐lesion) correlations of biomarker data can be used to eliminate candidate biomarkers.
Average or single molecule detection	Classical assay development methods were established on analytes present at the level of billions of molecules. Recent technologies, especially using nucleic acids, take us down orders of magnitude to levels where stochastic variation must be considered.	The Jak2 mutation assay from Ipsogen is suggested to be used at an input of 25ng of DNA. This corresponds to 10 000 copies of the genome (5000 cells). Therefore the assay may be useful for detecting the presence of high levels of mutant sequence but will not be useful for monitoring residual disease burden once mutant granulocytes drop to a level where the input recommendations preclude the presence of mutant sequence.
Univariate versus multivariate markers	Composite measurements should have clear rules to derive a final easy to interpret multivariate index.	The rules governing the combination of target, non‐target and new lesions for determining RECIST measurements are a good example of deriving a single patient level index for decision making. Similarly the development of the Oncotype Dx assay is a good example for molecular biomarkers (Paik et al., 2004).
Continuous/categorical	Measurements can be intrinsically continuous or discrete. Continuous measurements can be made discrete via cut‐offs.	Decisions are invariably discrete and therefore care must be taken to provide a clear message with output from continuous markers.