Biomarker discovery studies may fail to validate because the clinical population does not represent the intended clinical use or because hidden preanalytic variability in the discovery samples contaminates the apparent disease specific information in the biomarkers. This preanalytic variability can arise from differences in blood sample processing between study sites, or worse, introduce case/control bias in samples collected differently at the same study site. To better understand the effect of different blood sample processing procedures, we evaluated protein measurement bias in a large multi-center lung cancer study. These analyses revealed that perturbations in serum protocols result in changes to many proteins in a coordinated fashion. We subsequently developed protein biomarker signatures of processes such as cell lysis, platelet activation and complement activation and assembled these preanalytic signatures into quantitative multidimensional Sample Mapping Vectors (SMV) scores. The SMV score provides critical evaluation of both the quality of every blood sample used in discovery, and also enables the evaluation of candidate protein biomarkers for resistance to preanalytic variability.

The underlying platform technology uses SOMAmers (Slow Off-rate Modified Aptamers) as affinity reagents to quantify proteins. Following analytic validation of the assay, which simultaneously measured approximately 850 proteins with sub-pM limits of detection and intra and inter-assay CV of <5%, we initiated a lung cancer discovery program. The intended use was defined as early detection of lung cancer in individuals with indeterminate pulmonary nodules and in high-risk smokers. This case/control study included almost 1000 samples from 4 different centers, with blinded verification in ∼400 samples. Although the AUC of ∼0.9 in both training and blinded verification was promising, we had to eliminate several markers due to preanalytic bias leading to site-to-site differences. Even worse, when the SMV scores of preanalytic effects were applied retrospectively, we found there were substantial case-control biases within centers, and that a number of cancer markers were influenced by these biases. As a result, our initial diagnostic performance was partly dependent on markers that not only related to cancer biology, but that were also contaminated by preanalytic case/control bias.

To eliminate this effect, we used the SMV score to define an unbiased fraction of the original sample set, including samples from the Pittsburgh Lung Screening Study (PLuSS, supported by NCI SPORE in Lung Cancer grant P50 090440), and re-assayed the unaffected samples using a new version of the assay with 1033 analytes. The new 7-marker classifier had an AUC of 0.86 and contained several new markers that were not available in the prior version of the assay, which in turn enabled the elimination of some markers that were partly contaminated with case/control bias. The modest loss in performance was an acceptable price for avoiding dependence on hidden case/control information.

Applying quantitative measures of preanalytic variability, we identified preanalytic sample bias across 4 large study centers, revealing unintentional differences inherent in how biological samples are obtained, processed and stored for different study populations. This study highlights the consistency of cohort studies and provides a tool for evaluating bias in case/control studies before proceeding to biomarker discovery. Choosing biomarkers with not just the best case/control discrimination but that are also resistant to sample processing bias increases the likelihood that a robust biomarker panel will perform well in the clinic.

This content is only available via PDF.
©2013 American Association for Cancer Research.
2013