Choice of DNA conversion and qPCR assessment approaches

For BC, the EZ DNA methylation kit (Zymo Research) was chosen as the representative kit as it currently is one of the most popular bisulfite conversion kits judged from the literature as well as recommended by the company itself for use with Illumina Infinium MethylationEPIC BeadChip array that is the gold standard in methylation discovery studies. Additionally, it was recently ranked as the highest-performing kit among more than ten commercial BC kits in our previous investigation [18]. For EC, the NEBNext® Enzymatic Methyl-seq Conversion Module (New England Biolabs) was chosen as the only commercially available kit of its kind thus far. Given that both methods have different requirements and include diverse experimental steps, we slightly adjusted their protocols to allow for a fair comparison (Table 1). Particularly, for EC, and after consultation with the manufacturer to test a beta protocol more targeted for difficult, low input and already degraded DNA samples, no fragmentation prior to conversion was performed (see Methods for details). Importantly, that way we ensure that we solely measure the effect of the conversion method on the observed fragmentation levels. For BC, the elution volume was increased to 20 \(\upmu\)l to match the EC protocol.

To evaluate the two DNA conversion approaches, we employed the qPCR-based tool qBiCo (v2) that we have recently developed [18]. In brief, qBiCo is a 5-plex qPCR assay targeting a set of single-copy and repetitive sequences of the human DNA in order to assess several BC performance parameters: genome-wide conversion efficiency, converted DNA concentration and converted DNA fragmentation. First, to calculate global conversion efficiency, we employ two assays (Genomic/Converted) targeting the genomic and converted version of the multi-copy human L1 repetitive element (LINE-1) (\(\sim\)200 copies across the genome). Second, to calculate converted DNA concentration, we employ an assay (Short) targeting the converted version of the single-copy hTERT gene, previously used in commercial genomic DNA quantification assays. Third, to calculate converted DNA fragmentation, we additionally employ an assay (Long) targeting the converted but longer version of another single-copy TPT1 gene and compare it with the observed copies of the Short assay. More details on the methodology can be found in the Methods section. In our previous efforts, qBiCo was thoroughly optimized, validated and tested by using different bisulfite conversion kits, genomic DNA inputs and samples of varying quality [18], which also guided our experiments here.

Here, the performance of each separate qBiCo assay was based on PCR efficiency and linear fit of the standard curves. The lowest obtained R2 value for an assay was 0.95 with a mean value of 0.99 ± 0.01. The mean PCR efficiency was 90.3 ± 9.8% (Table 2). All qBiCo standard curves and performance parameters for each assay can be found in Supplementary file 1.

Table 2 Average performance of individual qBiCo assaysEvaluation of global conversion efficiency

Our study goal was to compare the global sample conversion efficiencies of both BC and EC methods when the same DNA samples were treated under different experimental conditions (Fig. 2). We considered conversion efficiency the most important parameter as it can determine whether a sample should be considered for downstream analysis, or not. qBiCo provides the conversion efficiency as one outcome parameter, which is estimated by comparing the detection of two fluorescently labeled fragments: the converted and non-converted (genomic) version of a DNA sequence within the human L1 repetitive element, found hundreds of times across the genome [18].

Fig. 2

Developmental validation of BC and EC DNA conversion methods in terms of conversion efficiency based on the recently developed qPCR QC tool qBiCo [18]. Several parameters were assessed including: A repeatability for intra-experimental variation; B reproducibility and sensitivity for inter-experimental variation; effects of C elution and D incubation time in the conversion methods protocol; E artificial methylation of commercial gDNA standards; robustness in terms of F UV treatment and G sonication prior to conversion; inhibition by H hematin and I proteinase; stability of converted gDNA by J storage time and K freeze-thaw cycles. The genomic DNA input for all conditions (C–K) was 100 ng

Firstly, we tested the effect of the initial DNA input on the conversion efficiency and how repeatable/reproducible our measurement is. When comparing a single BC and EC conversion experiment, no significant difference in conversion efficiency was detected with the DNA input as covariate (p = 0.999). However, at 100 ng, the conversion efficiencies were significantly different, namely, 98.9 ± 0.3% and 99.4 ± 0.2% for BC and EC, respectively. (Fig. 2A: ****: p < 0.0001), probably driven by the very low variance at this high amount. Using 100 ng of input DNA which falls within the optimal range for both methods, the conversion efficiency was 98.5 ± 1.0% (BC) and 98.7 ± 2.1% (EC), respectively (Fig. 2B). Subsequently, when decreasing the DNA input five (20 ng) and ten times (10 ng), but still within the suggested limits from the manufacturers for both methods, the conversion was detected as 96.7 ± 2.3% (10 ng: 97.8 ± 0.7%) and 98.7 ± 0.7% (10 ng: 97.9 ± 0.9%), respectively (Fig. 2B). Interestingly, when further reducing the input and treating sub-optimal DNA amounts (5 and 1 ng), the conversion efficiency decreased to 97.5 ± 1.4% (1 ng: 94.6 ± 4.4%) and 96.6 ± 1.8% (1 ng: 87.2 ± 7.6%) for BC and EC, respectively (Fig. 2B). For all amounts, no statistically significant difference between the two conversion methods was observed while including the DNA input as a covariate (p = 0.291), although we critically acknowledge the small sample size (n = 3). Per DNA input amount, no statistically significant differences in conversion efficiency were detectable when taking into account the inter-experimental variation between the conversion methods. Additionally, an increasing variance was detected at DNA input amounts of 1 ng, indicating that conversion was not reproducible at this level op input DNA.

Secondly, we tested the effect of various other experimental conditions on the conversion efficiency, in terms of either protocol variations or sample characteristics. The effect of the changes to the conversion method protocols were not statistically significant, resulting in conversion efficiencies of 98.2 ± 0.3% and 99.4 ± 0.4% for the different elution protocols and 98.9 ± 0.7% and 99.3 ± 0.3% at the various incubation protocols for BC and EC, respectively (Fig. 2 C, D). Overall, the conversion efficiency decreased with a decreasing methylation status when treating artificially methylated DNA going down. Especially at a methylation status of 50% or lower, the conversion efficiency decreased to 85.0 ± 1.7% and 87.8 ± 2.2% for BC and EC, respectively (Fig. 2E). This effect was measured earlier when using qBiCo [18]. Both robustness tests showed a decrease in conversion efficiency after either UV treatment or sonication of the gDNA prior to conversion. Namely, at a fragment length of 1000 or 500 bp, the conversion efficiencies had a minor decrease (BC: 97.4 ± 0.7%;EC: 97.6 ± 0.4%), whereas at a median fragment length of 150 bp, these values lowered to 95.5 ± 0.4% and 93.3 ± 0.6% for BC and EC, respectively (Fig. 2F). BC was significantly more prone to the effect of pre-UV-treated DNA. At 120 s of UV treatment of the input DNA, conversion efficiencies decreased to 54.3 ± 7.7% for BC and 83.9 ± 6.0% for EC (Fig. 2G). Both common contaminants of gDNA, hematin and proteinase, and both storage conditions did not affect the conversion efficiency (Fig. 2 H–K). No statistical evaluations were made for these parameters due to the small set of data points.

Evaluation of converted DNA recovery

In order to compare the recovery of converted DNA between both BC and EC methods, the same DNA samples were treated under different experimental conditions (Fig. 3). The amount of recovered converted DNA was determined by quantification of a single-locus converted DNA sequence (hTERT gene) in the samples after conversion by the qPCR assay qBiCo [18]. This ensures enough DNA is recovered for follow-up analyses, such as for whole-genome or targeted bisulfite sequencing. We calculated converted DNA recovery dividing the amount of recovered converted DNA by the initial DNA input.

Fig. 3

Developmental validation of BC and EC DNA conversion methods in terms of converted DNA recovery based on the recently developed qPCR QC tool qBiCo [18]. Several parameters were assessed including: A repeatability for intra-experimental variation; B reproducibility and sensitivity for inter-experimental variation; effects of C elution and D incubation time in the conversion methods protocol; E artificial methylation of commercial gDNA standards; robustness in terms of F UV treatment and G sonication prior to conversion; inhibition by H hematin and I proteinase; stability of converted gDNA by J storage time and K freeze-thaw cycles. The genomic DNA input for all conditions (C–K) was 100 ng. Missing values for A are displayed at y = 0 in black

Firstly, the repeatability/reproducibility of the recovery was compared for both conversion methods. Generally, the recovery index for BC was significantly higher than for EC (Fig. 3) (p = 2.16e\(-\)26). When compared in a single experiment, BC shows recoveries of 1.9 ± 0.2, 2.1 ± 0.2 and 2.0 ± 0.4 for 100, 20 and 10 ng of input DNA, respectively. In contrast, for EC these values were 0.6 ± 0.2, 0.4 ± 0.1 and 0.4 ± 0.2 (Fig. 3A). The recovery was overestimated for all BC samples shown by their values exceeding 1. At the sub-optimal DNA input amounts of 5 and 1 ng, the recoveries were systematically overestimated when above the limit of detection, showing values of 2.4 ± 0.6 (1ng: 3.6 ± 1.9) and 1.2 ± 0.7 (1ng: 2.7 ± 1.0) for BC and EC, respectively (Fig. 3A). Additionally, the number of missing values increased greatly at these low amounts of input DNA. When taking into account inter-experimental differences at the representative DNA input amount of 100 ng, the recoveries were 2.5 ± 0.7 and 0.7 ± 0.2 for BC and EC, respectively (Fig. 3B), showing to be significantly different between BC and EC with the DNA input as a covariate (p = 9.5e\(-\)37). Also, when reaching the limit of detection of this index at 1 ng input DNA, the recovery is overestimated, as shown by the increase in recovery from 2.5 to 4.3 for BC and 0.7 to 2.7 for EC (Fig. 3B).

Secondly, no clear effects were found of the changes in elution strategy and incubation time in the conversion protocols (Fig. 3 C,D). Interestingly, low artificially methylated DNA mixtures were harder to recover after conversion by both BC and EC. The decrease in recovery from 100% methylated DNA down to 0% methylated DNA appeared stronger after BC than EC, with decreasing linearly from 3.0 ± 0.4 and 0.7 ± 0.1 to 0.18 ± 0.04 and 0.17 ± 0.03 for BC and EC, respectively (Fig. 3E). EC showed to be more robust for gDNA samples that underwent UV treatment or sonication prior to conversion, shown by a lower relative decrease of recovery. After conversion of UV-treated samples, the recovery decreased 2.5-fold for BC (3.5 ± 0.4 to 1.4 ± 0.2) and 3-fold for EC (0.9 ± 0.2 to 0.3 ± 0.1) at 120 s of UV treatment. A similar decrease from 3.6 ± 0.9 to 1.4 ± 0.2 for BC and 0.6 ± 0.1 to 0.2 ± 0.04 for EC was seen at sonication to a 150 bp fragments (Fig. 3 F,G). The addition of the solute, NaOH, in which hematin was dissolved, decreased the recovery of the EC DNA samples from 0.8 ± 0.1 to 0.3 ± 0.1, but no decrease was seen for BC: 2.0 ± 0.3 to 2.3 ± 0.3 (Fig. 3H). Afterward, no further decrease in recovery was detected when adding hematin. Both common contaminants of gDNA, hematin and proteinase, and both storage conditions did not affect the recovery in BC and EC (Fig. 3 I–K).

Evaluation of converted DNA fragmentation

To determine the effect of both conversion methods on the fragmentation of input DNA, we measured the fragmentation level of the converted DNA in various experimental conditions. This level of fragmentation was determined by the quantification of a short (85 bp) and a long (235 bp) fragment, targeting the hTERT and TPT1 genes in the samples after conversion by the qPCR assay qBiCo [18]. This follows the same principle as in current commercially available quantification kits for genomic DNA with an incorporated degradation level indication, such as the Quantifiler Trio (Thermo Fisher Scientific) or PowerQuant (Promega) kits; however, now targeting the converted DNA sequence. It is important to emphasize here that the fragmentation level is a qualitative index, showing degradation of the DNA sample after conversion.

Firstly, the fragmentation indices for both conversion methods were compared (Fig. 4). In a comparison of both conversion methods performed in a single experiment, BC shows a 2.7-fold higher fragmentation index than EC with values of 2.2 ± 0.4 and 0.8 ± 0.1, respectively (Fig. 4A; p = 3.96e\(-\)14). When reproduced over various experiments, this significant difference between the two methods was holding (p = 5.81e\(-\)15), with a value of 1.4 ± 0.8 for EC and 3.4 ± 0.9 for BC at the reference amount of 100 ng gDNA (Fig. 4B). Generally, BC fragments the DNA to a higher degree than EC, as expected due to the harshness of sodium bisulfite [25]. The detected fragmentation index was stable for both methods over 100, 20 and 10 ng of input DNA. At 5 ng of input DNA for BC, however, the fragmentation index increased from 0.9\(-\)1.3 to 3.0, but was under the limit of detection for EC. A point to note here is that the amplified large fragments of the hTERT gene are harder to detect than the other qBiCo assays due to its increased length, often resulting in missing values at lower converted DNA inputs as also seen here (Fig. 4A).

Fig. 4

Developmental validation of BC and EC DNA conversion methods in terms of fragmentation index based on the recently developed qPCR QC tool qBiCo [18]. Several parameters were assessed including: A repeatability for intra-experimental variation; B reproducibility and sensitivity for inter-experimental variation; effects of C elution and D incubation time in the conversion methods protocol; E artificial methylation of commercial gDNA standards; robustness in terms of F UV treatment and G sonication prior to conversion; inhibition by H hematin and I proteinase; stability of converted gDNA by J storage time and K freeze-thaw cycles. The genomic DNA input for all conditions (C–K) was 100 ng. Missing values for A are displayed at y = 0 in black

Secondly, two parameters which are expected to influence the fragmentation index are UV treatment and sonication prior to conversion, showing the importance of the quality of the initial gDNA samples. EC was more robust to both of these treatments as the fragmentation index for the BC samples was higher for all conditions. After 60 s, a 5-fold increase in the fragmentation index happened at EC from 1.6 ± 0.2 to 8.4 ± 1.8, whereas a 10-fold increase was measured at BC from 3.8 ± 0.2 to 36.2 ± 3.9 (Fig. 4F). These values showed an exponential increase of the fragmentation index in case of UV treatment prior to conversion, indicating that already damaged DNA fragments faster. A similar increase in fragmentation index was seen at the sheared DNA of 150 bp fragment length. Here, compared to the non-sheared samples, an 11-fold increase was seen at BC from 4.0 ± 0.3 to 42.8 ± 14.0 and an 8-fold increase at EC from 2.7 ± 0.2 to 21.1 ± 6.3 (Fig. 4G). Additionally, at 0% methylated standards the fragmentation index was detected at a 2.7-fold decreased fragmentation level for the BC samples at 1.6 ± 0.3 as opposed to the 4.5 ± 0.4 index for the other methylation levels. None of the other parameters seemed to have an effect on the fragmentation index for both conversion methods.

Overall DNA conversion method performance based on qBiCo

While comparing individual conversion performance-related indexes can give us valuable insights into each method’s strengths and weaknesses, we also aimed to holistically evaluate each conversion method and how the measurement of each index could influence one another.

According to the manufacturers of conversion kits, global conversion efficiency should be > 99% However, no minimum threshold has been decided among the scientific community as it has not been possible so far to measure conversion efficiency prior to DNA sequencing. While we were able to detect converted DNA when lower amounts are converted, combining the results of all three indices showed that the reproducible limit of detection for both BC and EC was 10 ng. At this amount, which corresponds to 0.5 ng of converted DNA input for each of the two technical replicates into qBiCo, we could confidently obtain measurements without statistically significant differences from the optical amount (100 ng). Not only was the variation of our measurements between replicates high, but we also obtained missing values. We believe this variation is caused by the conversion itself, rather that the qPCR detection, as we have previously shown that qBiCo can deliver reliable measurements down to 310 pg [18].

Next, as the fragmentation level increased, the reported conversion efficiency decreased, which was most evidently seen at samples treated with UV. It has been reported previously that for highly converted DNA samples the conversion efficiency cannot be accurately reported by qBiCo [18]. Additionally, both BC and EC seemed to exhibit a lower performance at artificially induced global (genome-wide) methylation percentages of 0–25%. Especially, the conversion efficiency decreased from 98–99% to 85–88%, while the recovery decreased to 0.14\(-\)0.22 for both conversion methods. In other words, with more degraded and less available DNA template, the detected conversion efficiency is less or less accurate. We also suspect these observations to be due to preferential degradation of unmethylated DNA [17]. Yet, it is important to reflect that very low methylated samples at the genome-wide scale are not encountered naturally, and are used here only for testing the limits of the technology.

Application and variability at low-level DNA conversion (10 ng)

Finally, we aimed to apply both conversion methods to real-life settings and compared their performance side-by-side when the minimum DNA amount was converted (10 ng), which is often the case when dealing with difficult templates like forensic-type or cell-free DNA. We decided to employ the same experimental conditions as in the previous conversion performance validation experiments, as no large differences in performance were observed when changing the various experimental parameters, as presented above. To achieve this, we converted a small set of whole blood samples under the same experimental conditions that allowed us to further assess the variability and reproducibility of BC and EC on a larger scale. For this experiment, we converted 22 gDNA samples, whose qBiCo performance indices resulted in a similar trend as previously seen during validation (Fig. 5.) It is important to mention that to circumvent possible qBiCo batch effects, BC and EC samples were randomly mixed over two qBiCo assays. Sample conversion efficiencies were detected at 97.3 ± 1.4% for BC and 97.8 ± 1.0% for EC (Fig. 5A), matching the previous measurements during validation (97.8 ± 0.7% for BC and 97.9 ± 0.9% for EC) (Fig. 2B). Once again, a clear overestimation of the converted DNA recovery index was detected for bisulfite-converted samples with a mean recovery of 1.3 ± 0.3, which indicates systematic bias, while similarly, the converted DNA recovery for enzymatic converted samples was again low with a mean recovery of 0.4 ± 0.2 (Fig. 5B). Lastly, the mean fragmentation levels were measured at 3.3 ± 1.7 and 2.1 ± 0.9, for BC and EC, respectively (Fig. 5C). For various samples, one of the technical qPCR replicates resulted in an index below the limit of detection of qBiCo as seen previously at the DNA input amount of 10 ng. More specifically, there was one (5%) EC sample for which it was not possible to determine the conversion efficiency. Additionally, for 7 (32%) of the qBiCo measurements, the fragmentation indices were below the limit of detection. Only for one other sample, the fragmentation index could not be determined, due to missing data of both technical replicates.

Fig. 5

Comparison between BC and EC conversion performance using 22 whole blood gDNA samples (10 ng). Indices obtained by qBiCo [18] are shown on the y-axis: A conversion efficiency, B converted DNA recovery and C converted DNA fragmentation. Missing values are displayed at the bottom of the graphs in black