Overall design of the study is presented in Figure 1. A retrospective cohort (n = 25) was intended for the protocol development and its analytical characterization on formalin fixed/paraffin embedded (FFPE) tumor tissues. For a prospective cohort (n = 18), FFPE and CSF samples were matched for diagnostic characterization and clinical validation of the protocol. All patients of the prospective cohort, presenting with newly diagnosed or recurrent radiographically-confirmed CNS tumors, were enrolled in the Molecular Diagnostics in Pediatric Neurooncology and Liquid Biopsies Project at the Dmitry Rogachev National Medical Research Center of Pediatric Hematology, Oncology and Immunology. An exclusion criterion was complete surgical resection (R0) determined on the basis of postoperative MRI data, unless CSF sampling has occurred during or before surgery. The Informed Consent for participation in the study was obtained from parents/guardians of all patients. The study was approved by the Institutional Review Board at the Dmitry Rogachev Center.

Clinical and analytical data for CSF donors (prospective cohort) are summarized in Table 1. CSF samples (1.0–13.3 ml, n = 19), collected either intraoperatively, or through a ventricular access device, or by lumbar puncture, were clarified by centrifugation at 1,100 g for 10 min and stored in 1 ml aliquots at −80°C. The samples were thawed immediately before cfDNA isolation. Cell-free DNA was isolated with the use of QIAmp Circulating Nucleic Acid Kit (Qiagen, the Netherlands). Concentrations of amplifiable fragments were measured by DNA Fragmentation Quantification Assay (EntroGen, United States). The cfDNA profiles were analyzed with the use of Cell-free DNA ScreenTape Assay Reagent Kit (Agilent Technologies).

The protocol was adapted and optimized using the Droplet Digital™ PCR platform (Bio-Rad) for FFPE tumor tissues specimens of the retrospective cohort (n = 25, Figure 1) with the known H3F3A K27M status determined by Sanger sequencing (Supplementary Table S1). Genomic DNA was isolated from the tumor tissues using FFPE RNA/DNA Purification Plus Kit (Norgen Biotek, Canada). The degree of fragmentation corresponding to cfDNA was modeled by shearing in a Covaris M220 Focused-Ultrasonicator (Covaris, United States) in a standard 150 bp mode. The DNA was subsequently purified and concentrated with AMPure XP beads (Beckman Coulter, United States) and quantified on a Qubit 3.0 fluorometer (Thermo Fisher Scientific, United States) with Qubit dsDNA BR Assay Kit (Thermo Fisher Scientific). Representation of different fragment lengths (37 bp, 150 and 300 bp) was evaluated using DNA Fragmentation Quantification Assay (EntroGen, United States) and additionally by automated electrophoresis in a TapeStation 2200 system (Agilent Technologies, United States) with the use of High Sensitivity DNA ScreenTape Analysis Reagent Kit (Agilent Technologies).

Due to the expected low concentrations of the H3F3A target in cfDNA, the preamplification step was introduced to increase the sensitivity of dPCR assay. The reactions were set in 25 µl of 1× PCR buffer B (Syntol LLC, Russia) containing 50 nM of primers, 0.25 mM dNTPs, 2.5 mM MgCl₂ and 0.1 U/µl Taq polymerase (Syntol LLC) with 50–100 copies of 150 bp DNA fragments per reaction. Preamplification was run in a T100 thermal cycler (Bio-Rad) at 98°C for 1 min, followed by eight cycles of 98°C for 10 s, 58°C for 30 s and 72°C for 30 s, followed by 72°C for 2 min. The product was diluted 1:1 with TE buffer and used in dPCR with LNA-modified allele-specific probes. The reactions were set in duplicates, each comprising 20 µl of 1× ddPCR Supermix for Probes (No dUTP) (Bio-Rad, United States) with the primers and probes added to final concentrations of 1,800 nM and 500 nM, respectively. The process was run in a C1000 Touch Thermal Cycler (Bio-Rad) equipped with a QX200 Droplet Digital PCR (ddPCR) system (Bio-Rad) at 95°C for 10 min, followed by 40 cycles of 94°С for 30 s (ramp 2°C/s) and 57°C for 60 s (ramp 2 °C/s), followed by 98°C for 10 min. The data were processed and analyzed in QuantaSoft Version 1.6.6.0320 (Bio-Rad). Fluorescence thresholds (respectively, 2,000 and 1,500 RFU for FAM and HEX) were set manually.

Oligonucleotide sequences of primers and probes previously published in [16] were as follows: forward H3 K27M: 5′-GGT​AAA​GCA​CCC​AGG​AAG-3′; reverse H3 K27M: 5′-CAA​GAG​AGA​CTT​TGT​CCC-3′; wild-type probe: 5′-HEX-TC⁺GC⁺A⁺A⁺GA⁺GT⁺GC-RTQ1-3′; K27M probe: 5′-6-FAM-TC⁺GC⁺A⁺T⁺GA⁺GTGC-RTQ1-3′, with “locked” nucleic acid-modified (LNA) bases indicated by⁺. LNA-modified allele-specific probes are known to increase dPCR specificity. Distinctive structural feature of LNA residues is the methylene bridge between atomic positions O2′ and C4′ of furanose, which locks the sugar in 3′-endo conformation thus enhancing hybridization of the probe with the template. Introducing LNA-monomers into oligonucleotide sequences significantly improves their discriminating capacity in the recognition of single-nucleotide substitutions [17].

The optimized protocol was tested for compatibility with the nanoplate-based digital PCR system (Qiagen) technologically different from the water-oil emulsion droplet-based PCR platform (Bio-Rad). The mixtures were set up in duplicates with QIAcuity Probe PCR Kit in Nanoplate 8.5K 96-well plates (Qiagen). The reactions, containing 3 µl 4× Probe PCR Master Mix, 1.2 µl primer–probe mix, 5 µl DNA template, and 2.8 µl RNase-free water in 12 µl total volume each, were run in a QIAcuity One instrument (Qiagen). The analysis was carried out with the use of QIAcuity Software Suite (Qiagen). Fluorescence thresholds for optimal resolution and cloud clustering, selected manually to fit all reactions, constituted 40 and 20 RFU for FAM and HEX, respectively. To enhance the accuracy of concentration measurements, the Volume Precision Factor (VPF), adjusting for tiny variations in nanoplate geometry, was applied as recommended by the manufacturer.

To quantitatively describe the dependence of dPCR outputs on reaction parameters (preamplification cycle number and template copy number), regression lines and determination coefficients were calculated in Microsoft Excel 2016. Bland–Altman test was used to assess quantitative agreement between the Droplet Digital™ PCR (Bio-Rad) and the QIAcuity™ Digital PCR System (Qiagen) based on the mean difference and standard deviation. The estimated concentrations of the target in starting samples were expressed in log₁₀ copies/µl. Non-paired measurements (i.e., with missing value for an alternative platform) were excluded from the analysis.

The H3F3A target preamplification protocol, as well as ddPCR settings and conditions, were optimized in order to increase the analytical sensitivity and specificity of the H3 K27M variant determination. The optimality was defined as a combination of high number of positive droplets with robust cloud separation and minimal “rain” effects (a fraction of droplets showing sub-threshold fluorescence).

The optimization was carried out by varying preamplification cycle number (4, 6, 8, or 10), annealing temperatures and concentrations of the oligos, as well as the number of DNA copies at the start of the reaction. The exponential dependence of the ddPCR-detectable amount of DNA target on the preamplification cycle number (Figures 2A,B) allows significant boost in representation of the target. Figure 2C demonstrates the dependence of ddPCR outcomes on the number of DNA template copies at the start of the reaction when using eight cycles of preamplification of the H3F3A fragment. The analyses were set in duplicates for each sample. The data indicate that eight cycles of preamplification of the H3F3A gene fragment are necessary and sufficient for optimal cluster density and accurate interpretation of the results of ddPCR with 50–100 copies of the template at the start of preamplification. Representative examples of primary ddPCR data for different reaction conditions are shown in Figure 3.

The protocol was further optimized by adjusting annealing temperatures of ddPCR primers (within the range of 53–61°C) and final concentrations of the primers and probes in the reactions. Oligonucleotide concentrations were tested extensively within the range of 900–2,700 nM with a step of 300 nM for the primers and 250–750 nM with a step of 125 nM for the probes. Reducing the probe concentration to values below 500 nM resulted in the reporter fluorescence in HEX channel too low for reliable cloud separation. Elevated concentrations of oligonucleotides in the mixture increased the overall intensity of fluorescence but negatively affected specificity of the reaction. High concentration of the FAM-labeled fluorescence probe specific for the H3 K27M allele led to the non-specific annealing of the probe to the normal allele. This resulted in the ambiguous appearance of the small amount (less than 50 in two technical replicates) of positive droplets with fluorescence above 2,000 RFU in the FAM channel in samples with absence of the H3 K27M substitution. Such results were considered as false positive (ambiguous sample). A sample containing at least 50 droplets with above-the-threshold fluorescence in the FAM channel and reproducible in two technical replicates was considered positive. In contrast, a sample containing more than 50 droplets with above-the-threshold fluorescence in the HEX channel and less than 50 droplets with above-the-threshold fluorescence in the FAM channel in the total number of duplicates and reproducible in two technical replicates was considered negative.

The optimal fluorescence levels and maximal ddPCR efficiency were achieved at eight cycles of preamplification with 50–100 copies of DNA template and annealing temperatures of 58°C and 57°C for preamplification and ddPCR, respectively. The clearest cluster separation efficiency was achieved with 1,800 nM of the primers and 500 nM of the probes. The uniform threshold values of the fluorescence intensity, above which the droplets were regarded as positive, were set manually. The thresholds for FAM and HEX signals constituted 2,000 and 1,500 RFU, respectively. Typical results of ddPCR obtained before and after the optimization experiments are shown in Figure 4. The optimized parameters of the developed ddPCR-based system are described in the Patients and Methods section.

Analytical sensitivity of the modified ddPCR protocol was tested using a series of 10 sequential two-fold dilutions of genomic DNA comprising the studied H3 K27M variant, with concentrations descending from 600 to 1 pg/μl of DNA at the start of preamplification. The reactions were carried out in two technical replicates. The sensitivity threshold was determined by minimal dilution affording at least 50 droplets with above-the-threshold fluorescence in at least one of the channels in total counts for the duplicates (Figure 5). The limit of detection for the H3 K27M variant determined with the use of optimized ddPCR protocol was 9 pg of DNA (corresponding to 3 haploid genomes).

Analytical specificity of the developed ddPCR-based approach was tested using a selection of genomic DNA samples isolated from tumors with known H3 K27M status determined by Sanger sequencing (the retrospective cohort, Figure 1). The lack of positive signals for H3 K27M-negative samples confirmed full specificity of the developed protocol (see Supplementary Table S1).

PCR-based assay of representation/integrity of amplifiable cfDNA in the studied samples revealed wide variation. For pts. 6, 15, 16, and 18, insufficient analytical quality of the isolates was indicated by the absence of ≥150 bp DNA fragments; these samples were excluded from further analysis as non-informative. In other samples, the concentrations ranged from 3 to 15,170 copies/µl with a median of 81 copies/µl for the working fragment length of 150 bp (Table 1).

The developed ddPCR-based system was applied for determination of the H3 K27M somatic variant in cfDNA isolated from CSF. The analysis of CSF-derived cfDNA was validated against matching samples of FFPE tumor tissue by Sanger sequencing. The studied variant was identified in specimens #9–12 implicating the diagnosis of DMG H3 K27-altered. FFPE specimens #1–8 and #14–16 were H3 K27M-negative, and no solid bioptates were available for pts. 13, 17, and 18 (Table 1).

In sum, 14 ventricular (6 collected intraoperatively and eight from a ventricular access device), and 5 lumbar CSF samples were analyzed, corresponding to 18 patients with CNS tumors grade I–IV. In three out of four H3 K27M-positive cases, the mutation was unequivocally identified in ventricular (pts. 11 and 12) or lumbar CSF (pt. 9). For pt. 11, both ventricular and lumbar CSF samples were available. Interpretation of the results for ventricular CSF was straightforward whereas analysis of the lumbar sample produced ambiguous results—solitary high-intensity signals in two channels (i.e., for both alleles) independently of dPCR platform. Comparative representation of these data is given in Figure 6. It is important to note that variant allele frequencies (VAFs) measured by these two platforms were identical and did not depend on the use of preamplification step. At the same time, VAFs in CSF were considerably lower compared with tumor tissue (37% vs. 51.3%).

For pt. 10, only lumbar CSF with extremely low yields of cfDNA was available for the analysis, producing the same uninterpretable result despite preamplification.

All CSF samples from patients with H3 K27M-negative tumors (n = 11) produced negative results in dPCR for H3 K27M substitution, which indicates high diagnostic specificity of the developed protocol.

The accuracy of quantitative analysis for samples containing small amounts of genetic material (10–100 copies of DNA template) is limited by stochastic sampling. For small DNA copy numbers, the variation in the number of mutant alleles taken from the extracted volume of cfDNA and introduced into the preamplification reaction is described by the classical Bernoulli test scheme, and for copy numbers of 100 and above, according to the Moivre-Laplace limit theorem, by the Gaussian (normal) distribution [18]. In ordinary terms, with increasing concentrations of cfDNA, the probability of capturing the H3 K27M mutant allele into the analyzed pool of molecules increases. To describe the influence of the initial abundance of the mutant allele on the ddPCR outcome, the H3F3A target was amplified from 65, 50, 25, or 12 copies of the template (sample #9) at the start of preamplification. Comparative representation of the outcomes for different amounts of the template is given in Figure 7.

The number of copies of the DNA fragment-of-interest produced in the reaction at time t can be calculated as follows: n = f ( k ) k ⁡ exp ( t v ) , where n is DNA copy number after PCR, k—DNA copy number at the start of the reaction, v—reaction rate, t—time, and f(k) is a threshold function of k, f(k) = {0 at k < h; 1 at k≥h}, with h regarded as the threshold value.

The known number of DNA copies at the start of the reaction enables estimation of the expected number of amplified DNA fragments. Considering that k = 65 DNA copies at the start of the reaction corresponds to n = 1,574 positive droplets in the FAM channel, we get exp(tv) = 24.2. Accordingly, at k = 50, the expected number of amplified molecules will be n = 1,210, but only if k is not below the threshold value h; otherwise, n = 0. Thus, the empirical value of n = 3 is an artifact and should be replaced with a zero value. Similarly, at k = 25, the expected number of amplified copies will be n = 605, and at k = 12 it will be n = 290. In accordance with this model, the numbers of positive droplets (measured in the FAM channel corresponding to the mutant allele) are valid only for the case of 65 copies of the template at the start of the reaction. However, this value could vary depending on VAF in the tumor and should not be considered as a threshold level, but only as the explanation for the appearance of single positive droplets. It’s important to note that in the described case #9 H3 K27M VAF was relatively low—14%.

The appearance of the low amount (<2) of the positive droplets has been seen after preamplification in several samples with confirmed H3 wild type status belonging to patients with the infant-type hemispheric glioma, ETV6-NTRK3-positive (sample #1), posterior fossa ependymoma, group A (sample #7), diffuse pediatric-type high-grade glioma, H3-wildtype, and IDH-wildtype (sample #8) and pilocytic astrocytoma, KIAA1549 (exon 15)—BRAF (exon 9)—positive (sample #14). The presence of such droplets confirms their artefactual origin. The explanations of the appearance of the solitary droplets are either spontaneous cleavage of the fluorescence probe or its unspecific binding to the wild-type allele.

At the same time, CSF derived cfDNA samples from patients with H3 K27M altered tumors provided adequate positive results beginning from significantly lower amounts of the input DNA (19 copies/µl, sample#11.1 and 31 copies/µL, sample#12.1) compared to 65 copies/µl. Remarkably, the VAF in both of this cases were higher (22 and 60%, respectively), comparing to the sample #9. Single mutant allele copies in the starting material could provide solitary positive droplets (samples#10 and 11.5), but extremely low overall cfDNA input prevents an unambiguous interpretation of the results.

The optimized protocol was tested using an alternative digital PCR technology platform (QIAcuity™) to evaluate its universality. Analytical consistency between the platforms was evaluated using Bland–Altman method. The analysis was performed on cfDNA with and without preamlification step. A total of 43 paired measurements were analyzed; four measurements, performed on a single dPCR platform, were excluded from the analysis given the lack of pairness. Due to high volume of CSF collected from patients 11 and 12, nucleic acids were isolated in quadruplicates and in duplicates, respectively (all included in the analysis); see Supplementary Table S2 for the raw data.

Bland-Altman analysis (Figure 8) showed that the bias (mean difference) was 0.17 log₁₀ copies/µl, with a standard deviation (SD) of 0.15 log₁₀ copies/µl. For the vast majority of measurements (39 of 43, 90.7%), the results showed marked pairwise consistency of concentration values for mutant and wild-type alleles obtained with the two platforms, falling within the ±1.96 SD limit of agreement (−0.12…0.46 log₁₀ copies/µl). The rest of the differences (4 of 43, 9.3%) were outside the limits of agreement. For three of these outliers (two wild-type and one comprising the mutation), concentration values obtained with QIAcuity exceeded the differences, ranging from 0.52 to 0.60 log₁₀ copies/µl. For one of the outliers, the concentration of mutant allele obtained with QX200 was higher (−0.22 log₁₀ copies/µl). It should be noted that all differences falling outside the limits of agreement corresponded to non-preamplified cfDNA. Concentration values obtained with the use of preamplification had lower discrepancy between the platforms (−0.01…0.37 vs. −0.22…0.6 log₁₀ copies/µl).

Overall, the results of H3F3A K27M mutant and H3F3A wild-type target quantification obtained with the QIAcuity dPCR system were comparable with those obtained with the QX200 ddPCR platform, which indicates universality of the developed protocol.