Differential methylation with 1 sample only

Question

[Pin.Bioinf]

Hello, I have a smoothed, filtered methylation matrix that has :

• 8 samples for TAB sequencing data

  • each sample is a different time point

• 8 samples for BS sequencing data

  • each sample is a different time point corresponding to the TAB data (paired)

I need to do a differential methylation analysis but it has only 1 sample per condition: 1 sample per time point-method of sequencing (TAB or BS).

What should I do having no replicates?

Thanks.

[Kevin Blighe]

Hello, why was the experiment designed in that way? Was a statistician or bioinformatician not included in the discussions?

I presume that you want to see differences between all 8 time-points, with just 1 sample per time-point. The standard statistical test applied to such a setup would be unreliable. Your best bet may be to normalise all samples together and then just literally compare the Beta values via division

[Pin.Bioinf]

Hello, I guess there was none of them involved in the experiment. Also, they did more samples for rnaseq, I guess they added the methylation analysis to try luck and see if they can integrate both omics and get something added to the rnaseq experiment.

What do you mean by comparing the beta values via division? Could you give me an example? I also heard some people compute absolute difference between beta or m values.

[Kevin Blighe]

Yes, the difference in mean Beta value is typically used, but first they calculate the mean across replicates in each group, e.g.:

Mean TimePoint2 - Mean TimePoint1 = Difference in means

However, these are typically used in conjunction with a P value, such as that derived from the Wilcoxon Signed Rank test. You could technically just subtract the single Beta value that you have at each time-point.

By 'division', I just meant dividing the values (instead of subtracting) - this would give Beta ratios.

If this is just a pilot study, then okay. You may struggle to publish it though.

Answer 1

[oneillkza]

This is years old, but it came up for me in a Google search, so I thought I should answer. The DSS package in Bioconductor is designed for exactly this purpose (differential methylation with only one sample per condition). It uses some fancy genome-wide sampling (dispersion shrinkage) to be able to come up with reasonable P-values.

https://academic.oup.com/bioinformatics/article/32/10/1446/1743267 https://link.springer.com/article/10.1007/s40484-019-0183-8

https://bioconductor.org/packages/release/bioc/html/DSS.html https://bioconductor.org/packages/release/bioc/vignettes/DSS/inst/doc/DSS.html

It's also worth noting that it does its own smoothing, so no need to do that in preprocessing.