Since DSS can handle missingness better than methylKit, its worth at least checking how it handles the low-coverage sample data that I’ve been excluding from runs.
I just commented out the following two lines from the 10-diff-methyl-DSS.Rmd file and re-ran it using sbatch 10-diff-methyl-DSS.sh.
drop <- c("CF01-CM01-Zygote", "CF08-CM04-Larvae", "CF08-CM05-Larvae", "EF04-EM04-Zygote", "EF05-EM05-Zygote")
file.list <- file.list[!str_detect(basename(file.list), str_c(drop, collapse = "|"))]Results
New n for offspring analyses is the full 32 samples.
Only slightly more loci present in the samples (31596324 vs. 31450950 in the subsampled run), but fewer loci retained after filtering (6337469 vs. 7594836; not that surprising, given the low coverage).
During multifactor design tests, more DMLs and DMRs identified, for both the additive and interaction models(Additive: 11356 DMLs/1893 DMRs vs. 8404/1788; Interaction: 20 stage-dependent loci vs. 14 in the subsampled run)
However, during single-stage smoothed runs, fewer DMLs and DMRs identified (Zygote: 5357 DMLs/4145 DMRs vs. 7287/5208; Larvae: 4183 DMLs/3963 DMRs vs. 5330/4661).
The increase in multifactor-identified DMLs/DMRs, but decrease in per-stage tests isn’t unexpected. Having fewer loci post-filter to test on would naturally reduce the number of identified DMLs. The multifactor test, however, gained power via the increase in sample size, which helped outweigh the accompanying reduction in loci count.