Allele-specific methylation detection using nanopore sequencing and NanoMethPhase

Vahid Akbari opened his talk by discussing the current problems with methylation detection in human samples. The current gold-standard, he explained, is to treat DNA with bisulphite – but this has a number of limitations: incomplete conversion of residues, induced DNA damage, and difficulty mapping short reads against the genome, to name a few.

In contrast, nanopore sequencing is capable of generating ultra-long reads that map uniquely to the genome with methylation data contained within them too, removing the need for sequencing of paired samples as is necessary with bisulphite treatment. This, Vahid explained, opens up the opportunity to perform allele-specific methylation calling, as SNVs can be called from the reads alongside the methylation data, allowing clear phasing.

To decide on a tool for methylation calling, Vahid first benchmarked a number of available algorithms, choosing three of the most popular in literature to investigate: nanopolish, DeepSignal and Megalodon. The tools were used to call methylation in a publicly available NA12878 dataset, and the results compared to whole-genome bisulphite data and array data, both also publicly available.

All methods displayed a strong correlation to both whole-genome and array data, but Vahid decided to proceed with nanopolish, as overall it was more compute-resource-friendly. But, Vahid questioned, how does methylation calling behave with respect to coverage? Overall, an improvement was seen with increasing depth, but beyond approximately 20-25X, minimal further improvements were seen, indicating nanopore methylation calling can be performed to high accuracy at relatively low depth.

In addition, Vahid remarked, nanopore sequencing also yielded 5-10% more CpGs than whole-genome bisulphite sequencing, and these mostly mapped to repeat regions, highlighting the downfalls of performing short-read sequencing for methylation analysis.

With the methylation caller decided, Vahid moved onto SNV calling for the purposes of phasing. For this, the team chose Clair for SNV calling, and phased with Whatshap, achieving a haplotype block N50 of 1.8 Mb on public data. This provided the foundation for haplotype methylome detection with nanomethphase. Vahid used available parental SNV calls to assign reads to each haplotype, and compared this to phasing without parental information, giving similar results but with slightly weaker haplotype assignment in the latter. In both cases, differential methylation was clearly identified across well-known imprinting control regions and was consistent with expectations. Visualisation in IGV also showed clear allele-specific methylation patterns, supporting the findings.

Concluding, Vahid showed the full workflow – detection of methylation followed by SNV calling and phasing, before analysis of haplotype-specific methylomes and differential methylation patterns.