Comparison of Batch Effect Correction Methods in DNA Methylation Data

• Activity Number: 433
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Xinhui Wang* and Susan Groshen and Juan Pablo Lewinger and Kimberly Siegmund
• Companies: and University of Southern California and University of Southern California and University of Southern California
• Keywords: DNA methylation ; Batch effect ; SVA

Abstract

Technical effects can introduce noise into Infinium DNA methylation data that can cause bias and decrease power in statistical analysis if not appropriately accounted for. A variety of methods propose to correct for batch effects (BE) in microarray gene expression data. However, in DNA methylation studies the quantity of interest is a proportion of signal due to methylation, which is bounded between 0 and 1. Using simulation we compare the performance of gene expression BE correction methods on DNA methylation data. These include: Empirical Bayes (ComBat), Surrogate Variable Analysis (SVA), independent surrogate variable analysis (ISVA), and Remove Unwanted Variance (RUV-2). Sensitivity and specificity for identifying differentially methylated probes using linear regression of the phenotype of Interest on methylation value are summarized under different scenarios for the above methods and compared to a simple linear adjustment for batch effect. We found that in all scenarios, the SVA method outperforms all others with highest mean sensitivity and specificity and smallest variability. We recommend the SVA approach for general practice.