Threshold determination

In ddPCR generated droplets are identified as positive or negative based on a threshold at a certain fluorescence level and this ratio is used to calculate target abundance using Poisson-statistics. Therefore, determining a correct threshold is crucial for reliable quantification (Fig. 1). Defining a threshold is complicated by droplets with an intermediate fluorescence, termed as rain, which are puzzling to assign to the either positive or negative population. For the frequently used Bio-Rad ddPCR system, the QuantaSoft software offers an undisclosed method for automated threshold assignment and manual threshold setting by the end-user. The automated analysis often assigns thresholds so strict that a cloud of droplets is appointed positive that based on their low fluorescence is expected to be negative [16]. Alternatively, user-defined thresholds may be applied but these are generally not advised as they impair an unbiased interpretation of digital PCR data. Threshold setting can be challenging since the separation between positive and negative droplets may depend on many factors, such as the quality and quantity of the input sample, melting temperature and length of primers and probe, mismatches between the assay and target sequences, time between droplet generation and readout, pipetting precision, type of fluorescent reporter and type of quencher. Several algorithms have been developed by end-users that aim to offer more data-driven approaches to set thresholds. First, clustering methods were developed by Strain et al. and Jones et al. based on k-nearest neighbor-joining [17, 18]. The method of Strain et al. defines the median and variance of the negative and positive clouds to assess the statistical likelihood that outliers should be included in either cloud (p < 0.1). Jones et al. developed “definetherain” that uses negative and positive control samples to identify the two clouds. Subsequently, the mean fluorescence minus or plus three times its standard deviation is used as thresholds that are applied to the samples. Both these clustering methods calculate a threshold for each cloud of droplets and exclude intermediate fluorescent droplets from further analysis. In contrast, Dreo et al. proposed a single threshold determination method since droplets with intermediate fluorescence intensity can hold true positive droplets [19, 20]. This global manual threshold is defined as the mean fluorescence signal in the NTCs (no template controls) plus a number of standard deviations until one positive droplet remains in the NTCs [19]. These described methods assume a normal (binomial) distribution of the negative and positive clouds and do not account for shifts in baseline fluorescence between droplet populations of different samples. However, distribution fitting experiments and normality testing shows that droplet clouds do not follow a normal distribution and cannot be described by a single family of distributions. Furthermore, baseline fluorescence of the negative cloud has been shown to vary between samples and influence quantification [16]. Therefore, an alternative thresholding method was devised by Trypsteen et al. [16] that assigns a threshold regardless of the many factors that may affect the intensity and distribution of droplet fluorescence. This method, ddpcRquant, feeds data from negative controls to a generalized extreme value model and applies this threshold to the samples. The algorithm does not make assumptions of the underlying distribution of the droplet populations and accounts for baseline shifts. Alternatively, Lievens et al. [20] determine the threshold based on the shape of the fluorescence density peaks but to account for the possibility that clouds are not normally distributed set the threshold above the uppermost limit of the negative cloud. Recently, a novel method, “Umbrella”, was published that does not apply hard thresholding, but applies a model-based clustering and takes partition-specific classification probabilities into account to produce a final quantification result [21]. Threshold setting remains a challenging but crucial task. It is difficult to establish whether or not intermediate droplets represent true targets that should be used for analysis, since the current generation of ddPCR is not fitted with a fluorescence intensity sorter to allow for target confirmation by for example sequencing. Recent evidence however suggests that intermediate droplets should be considered to contain target molecules, as decreased amplification efficiency may arise from a suboptimal annealing temperature [22] or mismatches between the assay and the target sequence [16]. Furthermore, several studies that investigated ddPCR sensitivity have used a user-defined threshold that allocates rain to the positive fraction of droplets, and doing so have found results that are on par with the input reference and qPCR results [11, 20, 23, 24].

False-positives

Regardless of the method that is used to assign a threshold, currently available digital PCR platforms including the QX200 suffer from the observation of false-positive partitions and therefore false-positive results [11, 16, 18, 23–25]. One out of three wells of negative controls with no template had 2 or 3 positive droplets (0.16–0.22 copies/reaction) for HIV-1 RNA assay described by Kiselinova et al. [23]. These droplets had a similar fluorescence level as positive droplets in patients samples. The origin of these errors remains unclear and various hypotheses have been proposed. False-positive droplets can arise from contaminations or disturbed droplets that merge together, their joint fluorescence leading to a droplet with a higher baseline fluorescence that is miscalled as positive.

False-positive droplets can pose a threat to reliable HIV DNA quantification in settings with low HIV DNA concentrations such as mother-to-child transmission, early treatment initiation and allogeneic stem cell transplantation (alloSCT). AlloSCT is currently the only known approach by which the HIV reservoir can be drastically reduced. Following a successful stem cell transplantation, patients are kept on ART and are monitored for HIV DNA levels, but reliable ascertainment of remainder HIV DNA is a challenge, especially when the interpretation of true-positive droplets is obscured by false-positive ones. Same holds true for ART-treated children, which may have initiated ART early after birth based on the HIV-status of their mother whereas uncertainty may exist if the infection was transmitted from mother to child. In these seronegative children, HIV DNA is the only proof of HIV infection and therefore the only justification for treatment with ART. However, confirmation of the presence of HIV DNA is challenging since patients who initiated ART early after infection are known to have small reservoirs and sample volumes are restricted in case of young children, which reduces the statistical power to assess the presence of HIV DNA. Therefore, false-positives can unrightfully lead to confirmation of HIV infection and continuation of ART and it is not advised to use digital PCR if the question is to discriminate between presence or absence of HIV DNA [11]. Since only a minor fraction of all potential CD4 positive target cells carry HIV DNA, a large number of cells need to be tested in order to be able to reliably quantify HIV DNA concentrations. High concentrations of total DNA however affect the viscosity of the aqueous phase and complicate the formation of droplets. The amount of DNA that can be loaded into a single reaction is therefore restricted [18, 26]. Researchers who aim to report an HIV DNA concentration in a million CD4-cells are required to split the target DNA among a number of reactions, thereby increasing the risk of detecting false-positive droplets and influencing final HIV DNA concentration outcome. This effect is even greater when samples are used in which HIV DNA is even less abundant, such as PBMCs, whole blood, dried blood spots or tissue biopsies.

Advantages

Apart from the issue of false-positives, digital PCR has shown to be equal or superior to qPCR in several aspects. One major advantage is that digital PCR produces direct absolute quantification. The absolute quantification results produced by digital PCR eliminate the need for a standard curve in case of DNA quantifications and comparisons of RNA quantifications. Of note, RNA quantification represents cDNA molecules and should therefore be corrected for cDNA synthesis efficiency [27]. Accurate quantification by qPCR is based on the quality of the standard curve: instability of the standard curve can lead to inaccurate HIV DNA quantification [28]. Additionally, Cq values in qPCR that arise from the standard and the samples are based on amplification efficiencies, and several factors may confound their correct interpretation. Amplification efficiency may be affected by inhibitors, amount of total DNA that is loaded as well as variation between the primer/probe and the patient’s viral sequence, and these factors may unrightfully elevate the Cq values. In qPCR, such mismatches would increase the Cq and in turn present a target abundance that is lower than the actual input. In ddPCR however, a reduced amplification efficiency leads to less fluorescence at end-point. As long as end-point fluorescence remains above the assigned threshold and the ratio between positive and negative fraction of droplets is unaltered, mismatches between assay and target are allowed as they do not affect the quantification outcome [16, 29, 30]. Tolerance to target sequence variation is especially crucial for HIV quantification as a higher chance of mismatches with the primer–probe set is to be expected due to high heterogeneity of the virus [31]. Besides the robustness of ddPCR with respect to inhibition and reduced amplification efficiency, a higher precision and reproducibility was observed for ddPCR in comparison to qPCR [18, 32]. This is especially crucial in HIV cure efforts where the aim is to detect potential effects of the interventions on the HIV reservoir. However, it is important to note here that contradicting findings have been published that observed a higher sensitivity of the qPCR platform [23, 33]. In duplex digital PCR experiments on linked targets, a minority of partitions was observed in which only one out of two assays demonstrated amplification [34]. It remains however unclear whether this observation is artificial due to DNA shearing and physical separation of supposed linked targets, or genuine failure to amplify due to assay-specific inhibitors, DNA degradation or tertiary structures. Furthermore, in case of genuine failure to amplify it is currently unclear whether this potential mode of target underestimation pertains to digital PCR alone or if similar mechanisms are at play in case of (q)PCR.