Evaluation of qPCR curve analysis methods for reliable biomarker discovery: Bias, resolution, precision, and implications

Jan M. Ruijter, Michael W. Pfaffl, Sheng Zhao, Andrej N. Spiess, Gregory Boggy, Jochen Blom, Robert G. Rutledge, Davide Sisti, Antoon Lievens, Katleen De Preter, Stefaan Derveaux, Jan Hellemans, Jo Vandesompele

RNA transcripts such as mRNA or microRNA are frequently used as biomarkers to determine disease state or response to therapy. Reverse transcription (RT) in combination with quantitative PCR (qPCR) has become the method of choice to quantify small amounts of such RNA molecules. In parallel with the democratization of RT-qPCR and its increasing use in biomedical research or biomarker discovery, we witnessed a growth in the number of gene expression data analysis methods. Most of these methods are based on the principle that the position of the amplification curve with respect to the cycle-axis is a measure for the initial target quantity: the later the curve, the lower the target quantity. However, most methods differ in the mathematical algorithms used to determine this position, as well as in the way the efficiency of the PCR reaction (the fold increase of product per cycle) is determined and applied in the calculations. Moreover, there is dispute about whether the PCR efficiency is constant or continuously decreasing. Together this has lead to the development of different methods to analyze amplification curves. In published comparisons of these methods, available algorithms were typically applied in a restricted or outdated way, which does not do them justice. Therefore, we aimed at development of a framework for robust and unbiased assessment of curve analysis performance whereby various publicly available curve analysis methods were thoroughly compared using a previously published large clinical data set (Vermeulen et al., 2009) [11]. The original developers of these methods applied their algorithms and are co-author on this study. We assessed the curve analysis methods' impact on transcriptional biomarker identification in terms of expression level, statistical significance, and patient-classification accuracy. The concentration series per gene, together with data sets from unpublished technical performance experiments, were analyzed in order to assess the algorithms' precision, bias, and resolution. While large differences exist between methods when considering the technical performance experiments, most methods perform relatively well on the biomarker data. The data and the analysis results per method are made available to serve as benchmark for further development and evaluation of qPCR curve analysis methods (http://qPCRDataMethods.hfrc.nl).