Analysis of variability and epigenetic age prediction across microarray and methylation sequencing technologies

The variability and accuracy of epigenetic aging clocks are dependent on the choice of clock and the underlying technology utilized for methylation measurement. Using 100 technical replicate samples from two adult buccal cohorts, we compared technical methylation variability and signal strength between the Infinium MethylationEPIC v2.0 array and the Twist Human Methylome Panel across 753,648 shared CpGs. Twist methylation sequencing showed skewed methylation distributions and fewer highly correlated CpGs than MethylationEPIC arrays. Variance analysis revealed a skew toward higher signal strength in MethylationEPIC datasets with a subset of CpGs showing high signal strength in both methylation sequencing and array datasets. Despite these biases, four principal component (PC) trained epigenetic clocks (pcHorvath1, pcHorvath2, pcHannum, and pcDNAm PhenoAge) were robust across both technologies, even with missing data. While pcHannum and pcDNAm PhenoAge had similar replicate errors across both technologies, pcHorvath1 was more reproducible on arrays while pcHorvath2 was more reproducible on methylation sequencing. Furthermore, original non-PC versions of these clocks were significantly less reproducible in Twist datasets and displayed obvious differences in age prediction compared to their PC-trained versions across both technologies. For example, mean replicate error for the original, non-PC DNAm PhenoAge clock jumped from 3.850 years in arrays to 6.156 years in methylation sequencing. This underscores the need for careful selection of epigenetic clocks and technology-specific adjustments when optimizing for accuracy and reproducibility. It further emphasizes the importance of correcting for technical noise which, if left unaddressed, is substantial.

Files include metadata sheets, methylation beta tables, and methylation bedGraph files.

表观遗传衰老时钟（epigenetic aging clocks）的变异性与准确性，取决于所选时钟类型与用于甲基化检测的底层技术。本研究采用来自两个成人颊部队列的100份技术重复样本，针对753,648个共有CpG位点（CpGs），对比了Infinium甲基化EPIC v2.0芯片（Infinium MethylationEPIC v2.0 array）与Twist人类甲基化组面板（Twist Human Methylome Panel）的技术甲基化变异性与信号强度。结果显示，相较于甲基化EPIC芯片，Twist甲基化测序呈现偏态甲基化分布，且高度相关的CpG位点更少。方差分析表明，甲基化EPIC数据集的信号强度整体偏高，且存在一类在两种甲基化检测技术中均表现出高信号强度的CpG位点。尽管存在上述偏倚，四种经主成分（principal component，PC）训练的表观遗传衰老时钟——pcHorvath1、pcHorvath2、pcHannum及pcDNAm PhenoAge——在两种技术平台下均表现出良好稳健性，即便存在数据缺失亦是如此。其中，pcHannum与pcDNAm PhenoAge在两种技术平台上的重复误差相近；pcHorvath1在芯片平台上的重现性更佳，而pcHorvath2则在甲基化测序平台上表现更优。此外，这些时钟的原始非PC版本在Twist数据集内的重现性显著更低，且在两种技术平台上的年龄预测结果与其PC训练版本均存在明显差异。例如，原始非PC版本的DNAm PhenoAge时钟的平均重复误差从芯片平台的3.850年跃升至甲基化测序平台的6.156年。这一结果凸显了在优化准确性与重现性时，需谨慎选择表观遗传衰老时钟并针对技术平台进行适配调整的必要性。本研究同时强调了校正技术噪声的重要性：若未加以处理，技术噪声的影响将十分显著。本研究附带的文件包括元数据表、甲基化β值表及甲基化bedGraph文件。