2.1 Patients and samples

The study was designed as a proof-of-concept study by choosing a range of alterations that are observed in a routine clinical diagnosis laboratory (Figure S1). So, we conducted it on germline blood samples from 30 patients (Table 1) diagnosed at the Georges– Francois Leclerc Cancer Center (CGFL, Dijon, France) between 2017 and 2022. These samples have representative alterations we can observe in a routine activity with techniques used in a routine molecular diagnosis lab. All germline samples were sent to our lab for analyses as part of the routine clinical diagnosis procedure for predisposing syndrome to breast, ovary, prostate, pancreas, or digestive tract cancer. All patients gave their consent to use their samples for research after their use for molecular diagnosis. The study was conducted in accordance with the Declaration of Helsinki and approved (approval no. 00010311) by the Ethics Committee of the Georges‑Francois Leclerc Cancer Center (Dijon, France) and by the Consultative Committee of Burgundy (Dijon, France) for the Protection of Persons Participating in Biomedical Research (Comité Consultatif de Protection des Personnes en Recherche Biomédicale de Bourgogne). Written informed consent was provided by all patients.

2.2 DNA extraction

Three hundred microliters of whole blood were extracted with the Monarch Genomic DNA Purification Kit (New England Biolabs) by following the manufacturer's protocol. The quantity of the extracted genomic DNA was assessed with a fluorimetric method using a Qubit device (Fisher Scientific), and integrity was checked with a Tapestation 4200 device (Agilent Biotechnologies). The size of the DNA obtained was higher than 10 000 bp.

2.3 Library preparation

1.2 µg of gDNA extracted from white blood cells was slightly fragmented with g-TUBE (Covaris) by two spins of 1 min at 6010g. Then, libraries were prepared with the LSK114 kit (ONT) following the manufacturer's instructions.

2.4 Flow cell loading and sequencing set-up

R10.4.1 MinION flow cells were prepared following the manufacturer's instructions. When the final library amount was superior to 300 ng (15 µL at 20 ng/µL), half of the library was loaded. The second one was loaded 24 h later after a nuclease wash of the flow cell following the manufacturer's instructions. When the library yield is below 300 ng, the whole library is loaded at once. A bed file containing the chromosomic coordinates (±10 kb upstream and downstream of every complete gene) of 152 cancer-predisposing genes (Table S1) was uploaded to the MinKnow software, representing a size of 13 964 404 nucleotides corresponding to 0.47% of the human genome. The device used was a GridION containing a GPU Nvidia Quadro GV100.

2.5 Bioinformatics analysis

Sequenced reads were mapped using minimap2 (v2.24-r1122, -ax map-ont –MD)⁷ to the reference genome hg19. From the mapping, SNVs and indels of the 152 predisposing genes were identified using Clair3 (v0.1-r12, –platform = “ont” –model_path = r1041_e82_400bps_hac_g632 –enable_phasing)⁸ and decomposed/normalized with vt (v0.57721).⁹ Variants were then annotated with the following resources: VEP dbNSFP (v4.3a)¹⁰ and VEP dbscSNV (v1.1), ClinVar (2023-01),¹¹ gnomAD (v2.1),¹² OMIM (2021-12).¹³ Based on the annotations and ACMG guidelines,¹⁴ variants were automatically classified into five classes: pathogenic, likely pathogenic, uncertain significance, likely benign, and benign.

From the mapping, LSR were detected using Sniffles2 (v2.0.7, –phase –minsvlen 45 –long-del-coverage 10 –long-dup-coverage 0 –no-consensus).¹⁵ Variants were then annotated with the following resources: OMIM (2021-12), DGV (2016-03),¹⁶ MedGen (2022-10),¹⁷ ClinGen region and gene (2020-03),¹⁸ DECIPHER (HI_Predictions_Version3),¹⁹ and gnomAD pop frequency and pliscore (v2.1_sv.sites).²⁰ Based on the annotations and ACMG guidelines,²¹ LSR were automatically classified in the 5 ACMG classes described above.

2.6 Quality score calculation

The quality score, also named the Phred score, is used to define the basecalling quality. This score is an estimated error probability. The quality score is calculated as −10log(E) where E is the estimated error probability. For example, a quality score of 20 corresponds to an error of 1 in 100.

3.1 Technical performances of adaptive sampling sequencing

We first analyzed some technical parameters to provide the best overview of this new chemistry. The first step was the library preparation yield. Indeed, even if the starting material was germline genomic DNA, some samples had been stored at −20°C for several years (mean = 1.5 years, min < 1 year, max = 5 years). The median library preparation yield was 36.98% (quantity of libraries obtained in relation to the initial quantity of gDNA). This yield was not different for samples stored for more than 2 years (median = 34.06%) or less than 2 years (38.77%; Figure 1A). Then, we tested the library preparation yield in relation to the initial quantity of gDNA. When less than 1 µg (min: 525 ng, max: 970 ng) of gDNA is used, the yield was not different from an initial quantity of 1.2 µg recommended by the provider (median = 33.09% and 38.77%, respectively). However, a higher quantity of gDNA (1.5 µg) was linked to a decrease in the library preparation yield (median = 21.35%; Figure 1B). Based on these observations, we can conclude the age of the samples did not influence the yield of the library preparation, and the optimal amount of gDNA to initiate is 1.2 µg, with an average yield of 40.86% (min: 8.86%, max: 83.08%). Consequently, even if we used the same starting amount of gDNA, we were not able to load the same amount of final library on flow cells. We therefore analyzed the sequencing throughput in relation to the amount of library loaded. The increase of the amount in fmoles but not in ng loaded on flow cells tended to improve the sequencing throughput as well as the reload of the flow cells 24 h after the first load (Figure 1C; Figure S2A). Adding up to 100 fmoles is beneficial for throughput, but more than 100 fmoles decreases throughput. Then, we correlated the number of available pores before each run with the sequencing throughput. The number of available pores at the beginning of sequencing is directly related to the sequencing throughput (Figure 1D). Finally, we also observed target gene coverage increases with the throughput (Figure 1E).

One of the specifics of ONT sequencing is the basecalling performed throughout sequencing. Three levels of basecalling accuracy were available: fast accuracy, high accuracy, and super-high accuracy. As adaptive sampling needs high informatics resources, the super-high accuracy basecalling cannot be performed simultaneously. We first tested fast-accuracy basecalling on three samples and observed only 48, 50, or 65% of the reads passed the filters. When we performed super-high accuracy basecalling on the same fast5 files (at the end of the sequencing runs), we obtained more than 80% of reads passing filters (Figure 1F). Next, we compared high-accuracy basecalling (performed during the sequencing) and super-high accuracy basecalling (performed separately from the sequencing). While the fast basecalling did not give a high percentage of good quality reads, high accuracy and super-high accuracy basecalling gave similar results with a median percentage of pass reads of about 75% (Figure 1G). Even though we observed the same results with high accuracy and super-high accuracy basecalling, we analyzed our samples with super-high accuracy once the sequencing runs had been completed. We also observed the percentage of pass reads was not influenced by the throughput (Figure S2B). Similarly, neither the percentage of reads in the target (Figure S2C) nor the enrichment (Figure S2D) were influenced by the sequencing throughput. The enrichment is directly linked to the percentage of reads in the target (Figure S2E). Moreover, enrichment had an important influence on the coverage of target genes as we obtained higher coverages with higher enrichments (Figure S2F). Finally, by using adaptive sampling, we obtained a mean coverage of shallow whole genome sequencing from the rejected gene of 2.04× (0.76–4.29), whereas the mean coverage on the genes contained in the gene selection file (also called manifest) was 14.9× (4.71–35.08), corresponding to a mean enrichment of 7.26× (3.74–9.42; Figure 1H; Figure S2G).

3.2 ONT adaptive sampling improves the characterization of LSR

Among our 30 samples, 11 carried an LSR on different genes detected by MLPA (Multiplex Ligation-dependent Probe Amplification; Table 2). As ONT sequencing allows the analysis of long DNA fragments, we wondered whether the sequencing of complete genes from a bed file could detect the same LSR as MLPA. As expected, all LSR were also detected with ONT adaptive sampling (Table 2). Moreover, with the sequencing of introns, we were also able to identify the exact start and stop coordinates of each LSR (Table 2), and the exact size of deletions or amplifications. For example, sample #5 harbored a deletion of exons 11 and 12 of BRCA1 by MLPA (Figure S3A). With adaptive sampling, we identified a deletion of 4483 bp from intron 10 to intron 12 (Figure 2A, Table 2). We also perfectly characterized two other deletions of exon 4 in BRCA2 (Figure S3B) and exons 4 and 5 in MLH1 (Figure S3C). Indeed, we observed a 2393 bp deletion, from intron 3 to intron 4, carrying exon 4 of BRCA2 (Figure 2B, Table 2), and a deletion of 4789 bp deletion from intron 3 to intron 5, carrying exons 4 and 5 of MLH1 (Figure 2C, Table 2). In addition to identifying some well-known LSR, we also better characterized some others. Indeed, a sample originally identified as having an exon 13 duplication in BRCA1 (NM_7294; Figure S3D up), was characterized as a carrier of duplication of both exons 12 and 13 of BRCA1 (NM_7300; Figure 2D up, Table 2). We confirmed this observation on another sample carrying the same alteration (Figure 2D down, Figure S3D down). Another sample was identified with a partial deletion of promoter and a complete deletion of exon 1a and exon 2 of BRCA1 (Figure S3E). ONT sequencing showed a large deletion of 4967 bp carrying whole exons 1 and 2 of BRCA1 but also exon 1 of the non-coding protein NBR2 gene (Figure 2E). Finally, we considered another sample harbouring a complete deletion of BRCA1 (Supplemental_Fig_S3F). We confirmed the complete deletion of the BRCA1 but observed this deletion was part of a larger deletion of 139 603 bp carrying over five complete genes NBR2, BRCA1, RND2, VAT, and IFI35 (Figure 2F).

3.3 ONT adaptive sampling accurately detects single nucleotide variants

With the previous chemistry of ONT sequencing, the detection of SNV showed a high error rate. In order to test the accuracy of the new chemistry, we analyzed SNV detection on 15 samples with likely pathogenic/pathogenic (n = 15) or unknown (n = 2) variants previously detected by short-read sequencing and 4 samples without any variants detected by short-read sequencing (Table 3). All pathogenic variants except two were listed in the vcf files obtained after the alignment and variant calling. By focusing on the alignment files (bam format) for the two undetected variants, we observed them directly with an allelic frequency between 25 and 40% (Table 3). The two variants were not bioinformatically detected, probably due to their nucleotide environment and the sequence complexity created by the mutations. Indeed, homopolymer regions and highly repeated regions are difficult to sequence. Despite important advances in chemistry, sequencing could create artefacts, whatever the technology. For that, it is important to be able to discriminate true from false SNV. With short read sequencing, minimum coverage, and quality scores are now well established to validate variants.²² With ONT sequencing, especially with adaptive sampling and sequencing of native DNA, these quality criteria are not established yet. That is why we performed some analyses to find the minimum coverage and quality score allowing confidence in SNV detection. Ten samples were sequenced multiple times (2 or 3 times) due to low sequencing throughput. We performed variant analysis on the 21 runs and the 14 concatenated files obtained from the different runs. For each sample, we selected a specific SNV and followed its coverage and quality score for each run and each concatenation, representing a total of 35 SNV. We know the improvement of sequencing throughput increases coverage. Herein, we observed increase in coverage tended to improve the quality score of SNV for a majority of samples (Figure 3A). Moreover, we observed a main the part of the dots had a quality higher than 14. Then, we applied this quality threshold to the variants detected with SeqOne's bioinformatics pipeline and listed in Table 3. Eleven of 13 (85%) variants presented a quality higher than 14 (Figure 3B). The SNV of the 2 variants with a lower quality score were insertions. Nucleotide insertion or deletion-induced sequencing or bioinformatics issues as both variations had quality scores below 14. These variants, not detected with our bioinformatics pipeline, were all frameshift mutations.

Quality score threshold is not sufficient to validate a mutation and guidelines need to apply a minimum read depth. With short read sequencing, the minimum read depth is 30×. With ONT adaptive sequencing, native gDNA is sequenced without any PCR amplification. This specificity completely abolishes the appearance of PCR artefacts. Based on this specificity, we thought a read depth lower than 30× could be applied. We observed a majority of good-quality mutations had a coverage higher than 10× (Figure 3A). By applying these filters (quality > 14 and coverage > 10×), we were able to detect 82.9% (Figure 3B) of already known variants present in Table 3.

Then, we tested if we could confirm, with Sanger sequencing, variants selected by applying both thresholds on genes that were not analyzed by short-read sequencing. We randomly selected six variants in six different genes (Table S2) representing different types of alterations (small deletion, transition, transversion, double point mutations) for which we designed specific PCR primers for Sanger sequencing. All variants were confirmed by the gold standard: the double mutation c.703_704delinsAA in the GPT gene (Figure 3C), the splicing mutation c.27+1G > T in the CFAP126 gene (Figure 3D), the deletion c.162_179del in the MSH3 gene (Figure 3E), and three other point mutations in ANKRD26, TFKC, and BLM genes (Figure S4).

Finally, the use of adaptive sampling allows the analysis of whole gene sequences and therefore introns. In a sample with a family cancer history but without any alteration in coding sequences of cancer-predisposing genes analyzed by short-read sequencing, we observed some deep intron variants meeting our quality and coverage criteria. After checking in silico the splicing impact of these variants, we observed that two of them could create new alternative splicing sites in either BRCA1 with a new 5′ splicing site (Figure 3F) or BRCA2 with a new 3′ splicing site (Figure 3G). These variations need to be characterized with specific methods as minigene experiments. Unfortunately, we were not able to study their impact as it was impossible to obtain new blood samples dedicated to RNA analysis from these patients.