Evaluation of pre-processing on the meta-analysis of DNA methylation data from the Illumina HumanMethylation450 BeadChip platform

Abstract

Introduction Meta-analysis is a powerful means for leveraging the hundreds of experiments being run worldwide into more statistically powerful analyses. This is also true for the analysis of omic data, including genome-wide DNA methylation. In particular, thousands of DNA methylation profiles generated using the Illumina 450k are stored in the publicly accessible Gene Expression Omnibus (GEO) repository. Often, however, the intensity values produced by the BeadChip (raw data) are not deposited, therefore only pre-processed values-obtained after computational manipulation-are available. Pre-processing is possibly different among studies and may then affect meta-analysis by introducing non-biological sources of variability. Material and methods To systematically investigate the effect of pre-processing on meta-analysis, we analysed four different collections of DNA methylation samples (datasets), each composed of two subsets, for which raw data from controls (i.e. healthy subjects) and cases (i.e. patients) are available. We pre-processed the data from each dataset with nine among the most common pipelines found in literature. Moreover, we evaluated the performance of regRCPqn, a modification of the RCP algorithm that aims to improve data consistency. For each combination of pre-processing (9 × 9), we first evaluated the between-sample variability among control subjects and, then, we identified genomic positions that are differentially methylated between cases and controls (differential analysis).

Disease classification for whole blood DNA methylation: meta-analysis, missing values imputation, and XAI

“…We have obtained that harmonization can increase the classification accuracy by up to 20%. This is fully consistent with the original paper, which proposed the harmonization method regRCPqn [44]. When the preprocessing of training and test data is the same, harmonization has almost no effect on the final classification accuracy.…”

“…Ref. [44] developed an approach to systematically assess the impact of different preprocessing methods on meta-analysis. Its main advantage is the possibility of harmonization of the newly introduced datasets that does not require corrections to the previously analyzed datasets, employed for training the machine learning model.…”

Methylation data imputation performances under different representations and missingness patterns

“…Further, methylation data are stable and reproducible and offer a large amount of publicly available data, thanks to the cost-effectiveness of methylation arrays (Illumina Human Infinium Beadchips 27k, 450k and now 850k). This publicly available abundance of data, in turn, enables meta-analyses to advance discovery, thanks to numerous (ad hoc) preprocessing approaches [10].…”

Tissue- and ethnicity-independent hypervariable DNA methylation states show evidence of establishment in the early human embryo

“…We tested this using reliability metrics derived from analysis of technical replicates and found no evidence that hvCpGs are driven by technically unreliable probes. However, we note that better adjustment for technical artifacts within datasets 111 and the addition of further datasets would likely lead to the identification of more hvCpGs.…”