Flexible Methods for Standard Error Calculation in digital PCR Experiments

Digital PCR (dPCR) is a highly accurate and precise technique for the quantification of target nucleic acid(s) in a biological sample. This digital quantification relies on the binomial or Poisson distribution to estimate the amount of target molecules based on positive and negative partitions. However, the implementation of these distributions require adherence to underlying assumptions that are often neglected, leading to a suboptimal (too optimistic) variance estimation of the target concentration, especially when considering the multiple sources of variation in experimental dPCR setups. Moreover, these parametric methods cannot be easily used for downstream statistical inference when more advanced analysis are required, such as for copy number variation. We evaluated the performance of three new statistical methods (BootsVar, NonPVar, BinomVar) in both simulations and real-life datasets for target and variance estimation in dPCR setups while taking into account a combination of commonly observed sources of experimental variability that can interfere with the underlying assumptions of the current parametric methods. The results demonstrate the capability of the new methods for variance estimation and present a more accurate reflection of the true variability over the classical binomial approach. In addition, these statistical methods are flexible and generic in the way that they work well for the variance estimation of non-linear statistics that work with ratios (e.g. CNV) and for multiplex dPCR setups. In this study, we provide guidelines when to use the binomial-assumption based methods and when the non-parametric one is better to achieve more accurate variance estimates.