Benchmarking nanopore methylation analysis by comparison to publicly available bisulphite datasets

‘Epigenetics’ refers to heritable alterations of DNA that do not change the nucleotide sequence, such as cytosine methylation. 5-methyl cytosine (5mC) most frequently occurs in mammalian cells in CpG dinucleotides and can alter gene expression by suppressing transcription. During nanopore sequencing of native DNA, the signature of modified bases is present in the raw signal. To benchmark Nanopore methylation calling, we ran two replicates of the human genome NA12878 on one PromethION^TM Flow Cell, obtaining 20x per sample. We analysed both samples using the pipeline shown (Fig. 1a) and compared calls to two publicly available 50x bisulphite datasets. Nanopore read depth is more uniform than the bisulphite data (Fig. 1b), is not influenced by GC content (Fig. 1c) and fully maps to the genome (Fig. 1d). Taken together this results in a higher percentage of successfully called CpGs (>94%) at moderate overall read depth (Fig. 1e). Nanopore 5mC calls correlate well with bisulphite (>0.9) are highly reproducible (>0.95) (Fig. 1f) and require far less analysis time than bisulphite data (Fig. 1g). To examine the methylation status of larger features, we computed average methylation frequency of all CpGs in CpG islands. We found high agreement between 20x Oxford Nanopore Technologies and 50x bisulphite datasets (Fig. 1h), with a correlation >0.98. For comparison, two nanopore sequencing datasets showed higher reproducibility than two bisulphite runs (Figs. 1i-j). Both 20x nanopore sequencing replicates allow calling of >25,000 CpG islands (Fig 1k).

A common downstream analysis of methylation calls is the detection of differentially methylated regions (DMRs) between multiple samples/conditions, such as between cancer samples and their matched normals. Compared to targeted approaches like reduced-representation bisulphite sequencing (RRBS), nanopore sequencing allows us to investigate methylation changes on a genome-wide level. To illustrate this, we sequenced a skin cancer cell line (COLO829) and a matched normal sample (COLO829BL) to 100X, called methylation in both datasets and detected differentially methylated regions (Fig. 2a). As no WGS bisulphite dataset was available, we compared our results to RRBS data. The Oxford Nanopore dataset covered 28.8 million CpGs (representing 99% of all CpGs), whereas RRBS covered ~500,000 CpGs at this depth. Virtually all CpGs covered by RRBS were covered by nanopore sequencing (Fig. 2b). Using the nanopore dataset we identified differentially methylated regions between the tumour and normal across the whole genome. Figure 2c shows all detected DMRs stratified by their length and average methylation frequency difference between tumour and normal. The longest contiguous DMRs are regions of hypomethylation in the tumour sample. Figure 2d shows an example of such a region on chromosome 15. This locus is imprinted in healthy individuals, i.e. one of the alleles is epigenetically silenced. However, in the tumour, a complete lack of methylation can be observed over more than 1.5 Mb. Promoters, gene bodies, ALU and LINE-2 elements show higher methylation in the tumour while LINE-1 elements show hypomethylation (Fig. 2e). Finally, to illustrate smaller-scale methylation changes we chose to look at the known tumour suppressor PTEN, which shares a bidirectional promoter with its upstream neighbour KLLN. Here we found a 1.4 kb hypermethylated region overlapping with its promoter and 12 kb deletion covering exons 6-7. Nanopore cDNA reads confirm the missing exons and lower PTEN expression in the tumour.