Harmonization of DNA methylation datasets

This repository contains scripts for harmonizing DNA methylation (DNAm) datasets from four studies: ADNI, FHS, MESA, and HRS. The harmonized DNAm data are intended for evaluating methylation-based predictors and currently include baseline DNAm from cognitively unimpaired participants only.

Overview of processing steps

1. Sample-level QC

   • Removed samples with discordant predicted vs. recorded sex
   • Removed samples with bisulfite conversion rate < 85%
   • Removed samples with detection P-values > 0.01 for > 5% of autosomal probes

2. Probe-level QC

   • Removed probes that did not start with "cg"
   • Removed probes on the mitochondrial chromosome (chrM)
   • Removed probes missing from the annotation file

3. Imputation

   • A moderate number of probes had missing values.
   • Missing probe values (and values with detection P > 0.01) were imputed using the methyLImp2 R package.

4. Platform harmonization

   • All datasets (ADNI, FHS, MESA, HRS) used Illumina EPIC arrays; therefore, platform harmonization was not required.

5. Normalization

   • Applied BMIQ normalization within each dataset.

6. Batch effect correction

   • Performed using the harman R package.

Computational environment

This pipeline was executed on a workstation running Ubuntu 24.04.3 LTS with 16 CPU cores and 1.5 TB RAM. Peak observed memory usage was approximately 102 GB (for the FHS dataset).