Estimation of the fluorescence coefficient α
The fluorescence coefficient α, also known as an optical calibration factor, relates the number of target molecules to fluorescence intensity:
Fi = αXi + N(0, σ2)
Lalam, 2007 treats α as a known constant and do not specify a method to determine its value. We found that variability in α can induce significant amounts of uncertainty in our estimates of X0. The explanation behind this effect is that α and X0 are co-dependent when the fluorescence intensity is fixed. In other words, X0 can have any value if α is unrestricted. Hence taking variability of α into account is important in order to provide a more reliable estimate in our uncertainty in X0.
The choice of model
The fluorescence coefficient α can be estimated using samples where the initial copy number X0 is known. The experimental data for such samples is in the form of dilution series, with initial copy numbers varying between 1,000 and 100,000 molecules. A binomial branching process, such as the one used to describe the qPCR experiment above, behaves deterministically for molecule numbers as large as the ones in the dilution series. The pseudo-marginal Metropolis Hasting inference algorithm fails for deterministic processes (detailed explanation in Inference).
As it was found that the experiment behaves deterministically for the dilution series used in estimating α, a sigmoid curve model was considered more appropriate to represent the data.
Figure 15. A number of 1000 qPCR amplification simulations with efficiency r = 0.7 were performed in silico for X0=50 and X0=100,000, respectively. The amplification curve segment between cycles 32-35 is shown, with the 5-95% confidence interval, showing the variability in amplification between different experiments with the same initial copy number X0. The plot shows how the stochasiticity has a larger effect for small numbers of molecules.
Sigmoid model
Goll et al., 2006 propose a sigmoidal model to represent the qPCR process. In the paper, the model is used to perform absolute quantification on qPCR data and obtain the initial copy number X0 by using non-linear regression. The model does not capture the stochasticity of the amplification process, which is why it was not considered for single-cell data. However, the amplification process behaves deterministically in the case of dilution series with large initial copy numbers, that are used to obtain the fluorescence coefficient α, making the sigmoid model an appropriate choice.
The fluorescence curve is represented by the following equation:
where Fi represents the fluorescence at cycle i, Fmax the maximal fluorescence intensity, C1/2 the cycle with half of the maximal fluorescence intensity, k is a constant related to efficiency and Fb represents the background fluorescence. Non-linear regression is applied to the data to fit the four sigmoid parameters (Fmax, C1/2, k, Fb), using the Python function scipy.optimize.curve_fit. The fluorescence at cycle 0 is estimated by using these parameters in the following equation:
Finally, the fluorescence coefficient α is obtained:
Figure 16. Example of sigmoid fitting on real fluorescence data where X0 = 1,000,000 is known. Sigmoid fitting is used to estimate F0, the initial fluorescence intensity. The estimate for α = 2.97817284227e-09 is computed as the ratio between F0 and X0.
Variability in the coefficient α
Since literature papers assume that the parameter α is a fixed constant known in advance, it is interesting to see what the variability of α is depending on different factors such as the qPCR primer or experimental setting.
The estimation of the fluorescence coefficient α was done on experimental data from 3 different days (16/12/2016, 02/03/2017, 10/03/2017) and with 3 different primers (BJCOII, BJ-CYB C57, BJmmMUP) used to amplify standards with known initial copy number. The estimation showed both inter- and intra-experimental variability. The adjoining figure shows the estimate value of α plotted on the log scale, with visible inter-experimental variability present on orders of magnitude of α.
Figure 17. The α estimates show significant inter-experimental variability when different primers are used to amplify the standards (plotted on the log scale)..
Reassuringly, it can be observed from the plot above that the distributions of α estimates from samples with the same primer (BJmmMUP) are comparable, although the data comes from very different experimental settings, on different days (16/12/2017 vs. 10/03/2017). More work is needed to establish whether these two distributions differ significantly in their means. However, these results indicate that the variability in α between experiments is explained by the type of primer. This supports using an α estimate in the inference for a new sample if the same primer was used on the standards for α-estimation and on the new sample.
Figure 18. When the same primer is used, the distribution of α does not show significat inter-experimental variability. While further work is needed to establish this, it supports the practice of using an α estimate from standards that use the same primer as the sample of interest for which X0 is inferred.