Coverage-aware evaluation of Oxford nanopore methylation callers using whole-genome data

Comparison of four nanopore methylation callers against an array-derived consensus reference profile, assessed at five coverage thresholds and under three analytical strategies.

Oxford Nanopore sequencing detects methylation directly on native DNA, without bisulphite conversion, but the available methylation callers differ in their models, default parameters and genomic target space, and most published comparisons were limited to a single sequencing depth. The authors benchmark four widely used tools, Nanopolish, DeepSignal2, Megalodon and Remora, on whole-genome nanopore data from 26 human saliva samples, against a consensus reference methylation profile derived from an Illumina Infinium MethylationEPIC array saliva dataset (approximately 850,000 CpG sites). Performance is assessed at five coverage thresholds (1×, 5×, 10×, 30× and 100×) under three complementary strategies, using Pearson correlation, MAE, RMSE and Wilcoxon signed-rank tests. Single-read (1×) estimates prove unstable and strongly method-dependent, whereas coverage of at least 5× substantially improves accuracy and inter-method agreement. Increasing to 10× often gives the best compromise between accuracy and robustness, but comes with a marked loss of shared CpG sites, highlighting a critical trade-off between depth and site retention.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The usable take-home fits in one sentence: below 5×, a nanopore methylation call is worth little, and above 10× you gain in accuracy what you lose in the number of evaluable sites. This depth-versus-site-retention trade-off is rarely made explicit in protocols, although it has to be settled before sizing an experiment. The caveat is the setting, saliva and a forensic aim: the authors themselves state that MAE and RMSE are technical deviation measures rather than classification error rates, and transposition to constitutional diagnostic use still has to be done.

Analysis by Dr Thibaut Benquey