Wed. Oct. 30, Notes on differential methylation stats

Methods:

• Methylpy DMRfind:
  1. root mean square test for each CpG within an individual to determine if mCG counts:CG counts is statistically different than what would be expected from “genetic origins of methylation variation” alone. If the CpG site is significant after FDR correction, it is considered a DMS (differentially methylated site)
  2. DMS are combined into blocks (“region”) if they fall within a predefined window and have methylation changes in the same direction.
  3. Need to run group statistics
     • Examples of group statistics:
       • clustering with BackSPIN
       • random effects model? “CpG cluster differential methylation was determined by pooling CpG level methylation differences using a random effects model”

Reference: Schultz et al. (2015) Nature. doi:10.1038/nature14464 See pages 4-8 of Supp. Material

• Methylkit:
  1. calculateDiffMeth uses a logistic regression model to test if treatment has any effect on CpG (loci) methylation or region methylation.
     • is log(πi/(1−πi))=β0+β1Treatmenti a “better” model than log(πi/(1−πi))=β0 ?
     • a q-value is calculated and a cutoff can be defined to select significant DMLs or DMRs
       • this method is only capable of comparing two groups at a time.

Reference: methylkit documentation, see section 3.6

• DMGs:
  1. bionmial glm comparison between methylated and non-methylated CpGs
     • glm (methylated,non_methylated ~ pH * position, family = “binomial”)
     • each treatment replicate is compared to a control replicate in a combinatorial fashion to assign equal weightage to each replicate
  2. Liew also does glm on mean % methylation with gaussian method

Reference Liew et al. (2018) Science Advances