Blood-borne tumour biomarkers may be highly skewed and vulnerable to biases when determining performance characteristics


Session type:

Andrew Renehan1, George Bouliotis2, Cong Zhou1, Alastair Greystokes1, Caroline Dive1, Lucinda Billingham2
1University of Manchester, Manchester, United Kingdom,2University of Birmingham, Birmingham, United Kingdom


Blood-borne biomarkers are increasingly used for ‘go/ no go’ decisions in anti-cancer drug development. During the biomarker qualification process, there is a need to develop clinically meaningful performance characteristics, such as accuracy and cut-off points, but these have not been empirically evaluated in relation to biomarker distributions in cancer patients.


We pooled biomarker data from several studies in a variety of cancers – colorectal, small-cell and non-small cell lung cancer, and lymphoma. The distributions of 10 biomarkers – CEA, Ca125, Ca19-9, LDH, m30, m65, nDNA, circulating tumour cells (CTC), CYFRA21-1, IGFBP-2 – were assessed for skewness (defined by value < -1 or > +1). The appropriateness of log, Box-Cox and Johnson transformations was judged by the lowest kurtosis value. Monte Carlo simulations were derived based on the predominant patterns of distributions seen, using a 1:1 case-control design.


Twenty clinical settings using the various biomarkers were assessed. Skewness was a dominant feature over 18 scenarios. Three patterns of distributions emerged: normal; lognormal; extreme lognormal. Box-Cox and Johnson transformation did not always increase the ‘normality’ of the distribution. CTC distributions had particularly skewed patterns, often complicated by presence of zero values. Simulations showed that in a typical phase II trial scenario of a sample size less than 100, data-driven cut-off points overestimated the true cut-off by 5%, 20% and 30%, respectively, for the normal, lognormal, and extreme lognormal patterns.


Blood-borne tumour biomarkers may be highly skewed and vulnerable to biases when determining performance characteristics. A panel of transformations should be explored. Future approaches, using dynamic parameterizations (e.g. change in biomarker levels) will equally require considerations to appropriate transformations.