Data processing choices can affect findings in differential methylation analyses: an investigation using data from the LIMIT RCT

Abstract

Objective: A wide array of methods exist for processing and analysing DNA methylation data. We aimed to perform a systematic comparison of the behaviour of these methods, using cord blood DNAm from the LIMIT RCT, in relation to detecting hypothesised effects of interest (intervention and pre-pregnancy maternal BMI) as well as effects known to be spurious, and known to be present.

Methods: DNAm data, from 645 cord blood samples analysed using Illumina 450K BeadChip arrays, were normalised using three different methods (with probe filtering undertaken pre- or post- normalisation). Batch effects were handled with a supervised algorithm, an unsupervised algorithm, or adjustment in the analysis model. Analysis was undertaken with and without adjustment for estimated cell type proportions. The effects estimated included intervention and BMI (effects of interest in the original study), infant sex and randomly assigned groups. Data processing and analysis methods were compared in relation to number and identity of differentially methylated probes, rankings of probes by p value and log-fold-change, and distributions of p values and log-fold-change estimates.

Results: There were differences corresponding to each of the processing and analysis choices. Importantly, some combinations of data processing choices resulted in a substantial number of spurious 'significant' findings. We recommend greater emphasis on replication and greater use of sensitivity analyses.

Keywords: Bioinformatics; DNA methylation; Differential methylation; Reproducibility.

Figure 1. Flowchart of data processing and analysis.

Combinations of data-processing and analysis choices, consisting of six normalised datasets (SQN, BMIQ or SWAN, with probe filering before or afterwards), use or non-use of ComBat processing (supervised or unsupervised), and analysis with either an unadjusted model, a model adjusted for batch, or a model adjusted for batch and cell type proportion.

Figure 2. Probes ranked in top 10 by p-value in batch+cell adjusted model, for (A) infant sex, (B) BMI in standard care, (C) short-haired in Tabby.

For each probe the rank is given by pre- vs post-filtering, normalisation method, and batch-handling method. The model is one adjusting for batch (either explicitly in the model or via batch-correction algorithm) and cell type proportion. Adjust = adjusted for batch in the model; SCB = Supervised ComBat; UCB = Unsupervised ComBat.

Figure 3. Top 10 probes by LogFC: infant sex.

Largest LogFC for Infant Sex (female), by normalisation and batch correction method.

Figure 4. Top 10 probes by LogFC: BMI in standard care.

Largest LogFC for effect of BMI in standard care group, by normalisation and batch correction method.

Figure 5. Top 10 probes by LogFC: ‘short-haired’ in ‘Tabby’.

Largest LogFC for effect of ‘Short-Haired’ in ‘Tabby’ group, by normalisation and batch correction method.

Figure 6. Distribution of unadjusted P values by normalisation and batch correction method, for batch and cell adjusted models.

Only models from data where probe filtering was performed post-normalisation are included. The model is one adjusting for batch (either explicitly in the model or via batch-correction algorithm) and cell type proportion.

Figure 7. Distribution of log-fold-change estimates by normalisation and batch correction method, for batch and cell adjusted models.

Only models from data where probe filtering was performed post-normalisation are included. The model is one adjusting for batch (either explicitly in the model or via batch-correction algorithm) and cell type proportion.