Because of the use of different thermocyclers during data generation process, a strong batch effect was observed, and we applied a series of strategies of quality control and data analysis to counter such batch effect. On the probe level QC, at first, we selected good quality probes according to the detection P value<0.01 across all samples. We further removed those probes predicted to cross-hybridize with the sex chromosomes35 and those having overlaps with known SNP with MAF ≥0.01 (±10 bp) based on the 1000 Genomes database. On the subject level QC, we at first used principal component analysis (PCA) based on 50 000 randomly selected probes to select subjects that were within ±3 s.d. from the mean of a principal component (PC) for PC1, PC2, and PC3. Secondly, we filtered out those subjects with poor bisulfite conversion efficiency. We have compared data normalization strategy of COMBAT33 and independent component analysis (ICA) (http://cran.r-project.org/web/packages/fastICA/index.html)with the adjustment of batch variable in the analysis, and we found that the adjustment of the batch variable outperforms the other two strategies.
