Unexpected System-Specific Periodicity in qpcr Data and its Impact on Quantitation

Transcription

1 Unexpected System-Specific Periodicity in qpcr Data and its Impact on Quantitation Andrej-Nikolai Spiess Department of Andrology University Hospital Hamburg-Eppendorf

2 The infamous Ruijter et al. (2013) 384-replicate Dataset (Raw data) Fluorescence

3 The infamous Ruijter et al. (2013) 384-replicate Dataset (Linear model baselined data) Fluorescence

4 Plots of Cycle 10 Fluorescence Values of all 379 Samples Fluorescence Boxplot Point-Cloud

5 Plot of Cycle 10 Fluorescence Values Throughout all 379 Samples Fluorescence Runs Test p-value = 0.4

6 Plot of Cycle 20 Fluorescence Values Throughout all 379 Samples Runs Test p-value = 2E-16!!

7 Plot of Cycle 40 Fluorescence Values Throughout all 379 Samples Runs Test p-value = 2E-16!!

8 Plot of Cq values of all 379 Samples at Fq = 1000 Runs Test p-value = 2E-16!!

9 How can we reveal structure in qpcr data? We have seen that there seems to be some sort of pattern in Fluorescence values as well as Cq values in a technical replicate dataset. qpcr data is the result of a time-dependent process (Cycling!) Hence, methods of time series analysis should be feasible for analysing qpcr data and for revealing inherent structural features. One of these methods is: Autocorrelation analysis

10 Autocorrelation: Correlation against a shifted itself k = lag k =

11 Autocorrelation analysis of Cq values 1. Take Cq values of all samples 2. Fit a linear/quadratic model 3. Subtract trend and use residuals for autocorrelation analysis CQ

12 Autocorrelation analysis of Cq values => There is systemic pattern! sample period 24 sample period

13 Lightcycler 96, GAPDH, 96 technical replicates, Single Channel Pipette Thomas Volksdorf, Hamburg Index Runs test p-value:

14 Lightcycler 96, GAPDH, 96 technical replicates, 8-Channel Pipette Thomas Volksdorf, Hamburg Index Runs test p-value: 0.15

15 StepOne, GAPDH, 96 technical replicates, Single-Channel Pipette Thomas Volksdorf, Hamburg Index 4 e e e e e Runs test p-value: sample period

16 CFX96, VIM, 96 technical replicates, Single-Channel Pipette Stefan Rödiger, Cottbus Index Runs test p-value: 3E-7 16-sample period

17 Rotorgene, PRM2, 72 technical replicates, Single-Channel Pipette Andrej Spiess, Hamburg Index Runs test p-value: 9E-4

18 Mapping the CFX96 Cq value residuals to the MTP positions (Heatmap) A 0.5 B 0 C D E F Index G H

19 Not only periodicity in Cq, but also in E, when estimated at Fq! Why? Fq = Cq is periodic. F(Cq 1) is periodic. Fq/F(Cq 1) is periodic! Autocorrelation on Efficiency F Cq /F Fq = 1000

20 Good for Jan: LinReg removes periodicity in E ACF Runs test p-value: L a g

21 Ok, how do the mechanistic models perform? MAK2: Boggy & Wolff, 2010 Raw fluorescence CM3: Carr & Moore, Raw fluorescence Cycles Cycles ACF Lag ACF N 0 N Lag

22 Periodicities in Cq values taken from the different methods in Ruijter et al. (2013) (Supplements) Y RESID ACF PSM Y Y RESID ACF Y RESID RESID ACF LinReg FPKM Cy0 ACF ACF RESID RESID RESID Y Y Y DART ACF FPLM ACF Miner

23 Another way to destroy periodicity in Cq values: Normalization to [0, 1] (Larionov et al., 2005)

24 Scaling the plateau phase results in periodic Cq values, that can be compensated X qpcr curve Sinus scaling factor SDM F(thresh) = 1

25 We can create plateau dispersion by assuming error in E! Efficiency Raw fluorescence N = N0 * E1 * E2 * E3 *... σ ymax = 0 D σ Cq = Efficiency Raw fluorescence1200 σ ymax = σ Cq =

26 In HRM, periodicity of Tm values is even more extreme! Residuals Autocorrelation 1.0 Breusch-Godfrey test: p = 1.3e-16 Runs test: p = 5.1e iq Residuals Autocorrelation CFX Breusch-Godfrey test: p = 9e Runs test: p = 7.5e

27 When mapping TM residuals to their MTP position, interesting things appear... => Do we see uneven thermal block profiles?

28 Summary We can observe periodic Cq values in many qpcr systems. We can effectively apply time series analysis methods to make them visible. We see this in all cycles starting from the exponential region. It s unlikely a result of multichannel pipettors. Mapping of periodic data to MTP positions suggests block effects. The periodicity propagates from Cq to E, if estimated there. There is no Cq periodicity when using methods based on FDM, SDM. Cq periodicity is driven by plateau phase periodicity. If we remove this (normalization), we can remove Cq periodicity. Plateau phase periodicity may be a result of cycle-to-cycle noise in E. There is even more dramatic Tm periodicity in HRM technology. Solution? Fingerprinting a qpcr system, remove fingerprint from Cq/Tm data.