Fatih Tiras, Christian Cole, Alexander Gray
This work provides, to our knowledge, the first coverage-aware, multi-metric and multi-strategy evaluation of ONT methylation callers on human saliva. Our results show that both the choice of caller and the sequencing depth have a major impact on methylation estimates, and that very high coverage thresholds can reduce the number of evaluable CpG sites to a degree that may limit downstream analyses. The study offers practical guidance for planning ONT-based methylation experiments in epigenetic and forensic contexts, particularly for selecting coverage thresholds and callers before downstream marker validation, age-prediction modelling or tissue-specific forensic interpretation.
BACKGROUND: Oxford Nanopore Technologies (ONT) enables direct detection of native DNA methylation without bisulphite conversion, offering long reads and field-deployable workflows. However, multiple ONT-compatible methylation callers differ in their underlying models, default parameters and genomic target space, and most comparative studies have evaluated them at a single sequencing depth using limited performance metrics. This makes it difficult to design ONT experiments and to interpret methylation calls in applications such as epigenetic biomarker discovery and forensic age estimation.
RESULTS: We systematically benchmarked four widely used ONT methylation callers-Nanopolish, DeepSignal2, Megalodon and Remora-using whole-genome nanopore sequencing data from 26 human saliva samples and a consensus reference methylation profile derived from an Illumina Infinium MethylationEPIC BeadChip array saliva dataset (GSE111631; approximately 850,000 CpG sites). Performance was assessed at five coverage thresholds (1×, 5×, 10×, 30× and 100×) under three complementary analytical strategies: individual-level comparisons, within-sample consensus across tools and pooled method-level analyses. Agreement with the reference was quantified using Pearson correlation, methylation-level deviation metrics (MAE and RMSE), Wilcoxon signed-rank tests and corresponding effect size estimates, supported by visual summaries. MAE and RMSE were interpreted as technical measures of methylation-level deviation from the reference profile, not as forensic classification error rates, false-positive/false-negative rates or age-prediction error rates. Across all tools, single-read (1×) estimates were unstable and strongly method-dependent, whereas coverage ≥5× substantially improved accuracy and inter-method agreement. Increasing coverage further to 10× often yielded the best compromise between accuracy and robustness, but also led to a marked loss of shared CpG sites, highlighting a critical trade-off between depth and site retention.
CONCLUSIONS: This work provides, to our knowledge, the first coverage-aware, multi-metric and multi-strategy evaluation of ONT methylation callers on human saliva. Our results show that both the choice of caller and the sequencing depth have a major impact on methylation estimates, and that very high coverage thresholds can reduce the number of evaluable CpG sites to a degree that may limit downstream analyses. The study offers practical guidance for planning ONT-based methylation experiments in epigenetic and forensic contexts, particularly for selecting coverage thresholds and callers before downstream marker validation, age-prediction modelling or tissue-specific forensic interpretation.