Unexpected System-specific Periodicity In Quantitative Real-Time Polymerase Chain Reaction Data And Its Impact On Quantification

Abstract
The “baseline noise” of quantitative real-time PCR (qPCR) data is a feature of every qPCR curve and has substantial impact on quantitation. In principle, two different forms of baseline noise can be encountered: (i) the dispersion of fluorescence values in the first few cycles of a qPCR curve around their mean (within-sample noise) and (ii) the dispersion of fluorescence values between different qPCR curves at the same cycle (between-sample noise). The most predominant effect that results in between-sample noise is an overall shifting of the qPCR curve on the y-axis (“baseline shift”), which is frequently compensated by “baselining” qPCR data. Common approaches are to subtract an averaged (Lievens et al., 2012; Rutledge, 2011; Goll et al., 2006), iteratively estimated (Ramakers et al., 2003; Ruijter et al., 2009) or lower asymptote derived (Tichopad et al, 2003; Peirson et al., 2003; Spiess et al., 2008) baseline value from all fluorescence values prior to quantitation (compare Table 1 in Ruijter et al., 2013).
Recently, we showed preliminary results on a published large scale technical replicate dataset (Ruijter et al., 2013) that indicated between-sample periodicity for fluorescence values at early and late cycle numbers (Tellinghuisen & Spiess, 2014). A more detailed interrogation of the between-run noise periodicity revealed that this effect occurs at all cycle numbers and constitutes a second and completely independent noise component that adds to the overall baseline shift. Most importantly, periodic noise persists even after classical “baselining” and results in a propagation of periodicity into estimated Cq values when using fixed threshold methods (LinReg, FPKM, DART, FPLM), hence resulting in periodic Cq values. In contrast, Cq values obtained from variable threshold methods based on first- or second-derivative maxima (Cy0, Miner, 5PSM) or from normalization of fluorescence data are completely devoid of periodic noise, corroborating the feasibility of these approaches.
The origin of periodic noise in qPCR data remains elusive. By employing a larger cohort of published and also self-generated high-replicate qPCR data from different platforms, we used classical algorithms of time series/signal analysis (i.e. autocorrelation analysis) to characterize the periodicity in more detail. Interestingly, we generally observed a periodicity of 24/12 for 384/96-well plate systems, respectively. These findings strongly suggest an effect of uneven temperature profiles in peltier block systems or variable liquid deposition of manual/automated multichannel pipetting systems, manifesting themselves as periodic qPCR data. We will present ways to eliminate periodic noise from qPCR data that results in a more reliable estimation of Cq values.