Cross-Platform Evaluation of Long-Read Sequencing for DNA Methylation and Portability of Methylation Predictors

Background DNA methylation is a central epigenetic mark with broad applications in aging research, epigenome-wide association studies (EWAS), and biomarker development. While the Illumina EPIC array remains the standard platform for large-scale methylation profiling, long-read sequencing technologies, including Oxford Nanopore Technology (ONT) and PacBio HiFi, enable native and bisulfite-free methylation detection at near-genome-wide scale and uniquely offer simultaneous characterization of genetic variation and DNA methylation from a single experiment. Despite their promise, systematic evaluations of measurement reproducibility and cross-platform agreement, particularly for methylation-based prediction models, remain limited. Methods We present a cross-platform benchmarking study integrating two population-based datasets: peripheral blood samples from the Personalized Environment and Genes Study (PEGS), assayed using ONT, PacBio HiFi, and the Illumina EPIC array; and 1,787 unique participants from the All of Us Research Program (AoU), represented by 1,815 platform-specific long-read datasets, including 28 participants assayed on two platforms. We assessed CpG coverage, cross-platform agreement, the contribution of sequencing depth in simulation, and the performance of 173 methylation-based predictors. Results Both ONT and PacBio measured approximately 28 million CpGs at read depth ≥ 10, including 0.78–0.82 million EPIC-array sites per PEGS sample. The ordinary Pearson correlations were 0.85–0.87 between paired ONT-PacBio measures, whereas mean-centered correlations were much lower, indicating limited agreement for methylation measures at individual CpGs. In simulations calibrated to PEGS EPIC data, depths of approximately 50–60× approached a reproducibility comparable to published Illumina array raw data for technical duplicates. Methylation predictor portability increased with model size: 19 of 21 predictors containing > 1,000 CpGs had r > 0.8 between ONT-PacBio duplicate pairs, whereas only four of 137 predictors containing < 500 CpGs reached the same threshold. We further evaluated pooled sample sequencing as a cost-reduction strategy but found that the resulting read depths were insufficient for accurate individual CpG methylation estimates. Conclusions Long-read sequencing provides broad methylome coverage and supports methylation predictors built on larger CpG sets, but current cross-platform agreement does not support direct substitution at individual CpG without additional depth, calibration, and platform-specific validation.