BS-Seq data analysis with more than two (2+) conditions

[OHahn]

Dear all,

I could need some help/recommendations as I am about to analyse BS-Seq data for more than two conditions.

We conducted BS-Seq on mice samples, which were held under four different feeding conditions. For each conditions the dataset contains three replicates. The experiment went well. (So the data is already generated!)

Now my question: After conducting the alignment, how would you suggest to proceed in order to analyse differential methylation on ALL four conditions. I know there are enough tools and pipelines out there to conduct pairwise analyses but that does not help me in this case as multiplie pairwise comparisons are probably not really comparable to each other. (and if so, I could imagine that would be probably only on a qualitative but not statistically proper quantitative level).
Also, the handling of the information of the replicates should not be considered.
As I am only collecting ideas at the moment, I do not care too much whether the analysis would be on the scale of single Cs or whole regions.

An ideal solution would probably be an equivalent to an ANOVA in Transcriptomics (which is also used by DESeq2 or edgeR) where one can conduct the test with multiple factor-levels and even multiple factors.

Is there any magical tool out there being able to handle this kind of task and I have not heard about it? Or has anyone came across a paper/pipeline where they tackled that question? (I didnt)

If not maybe someone of you has a good sugestion how such a test should be conducted? We are currently thinking about converting somehow the information from a BS-Seq mapper (like Bismark) into sth that could be read by an ANOVA, Deseq2 etc. OR adapting the statistical framework of an ANOVA to mapped BS-reads.

Thank you very much for reading this long text and I am happy about any idea/suggestion you might have!

Best,

Oliver

[dpryan]

It's surprising that no one has written something for this yet. I started to do something like this but never got around to completing it (e.g., most of the documentation is missing or wrong). If nothing else, the base code should already be there to fit models of arbitrary complexity with a beta-binomial using prior distributions (i.e., it's like DESeq2 and edgeR, but with a beta-binomial rather than a negative-binomial). I should note that we were specifically interested in looking at changes in variance, so none of that is profiled out.

There's no direct way to convert BS-seq counts for use with DESeq2 or a linear model. If you can find a convenient linear model function (e.g. limma) that allows per-row weights then you should be able to use that after logit transforming ratios. That would be a reasonable approach, but everything I know of like limma only allows per-column/sample weights, which wouldn't be sufficient for you.

[dariober]

It's surprising that no one has written something for this yet.

Indeed!!

I started to do something like this but never got around to completing it (e.g., most of the documentation is missing or wrong). If nothing else, the base code should already be there to fit models of arbitrary complexity with a beta-binomial using prior distributions

Interesting, thanks for sharing! I came up with a similar idea of using beta-binomial model using the glm functionality in the R package aod.

If you can find a convenient linear model function (e.g. limma) that allows per-row weights then you should be able to use that after logit transforming ratios.

An idea I wanted to try is to linearise the % methylation by means of e.g. logit, as you suggest. Then use the total counts per position as weight for limma/voom (i.e. total counts ~ counts per gene). Then combine the weights from voom with the logit data for lmFit. (Not sure it makes sense...)

[dpryan]

Regarding aod, I recall trying that and pretty much anything else that could fit betabinomials. As I recall, if you don't need to fit both mu and phi, possibly with different model matrices, at the same time then all of those work well enough (just keep in mind that beta binomials are a royal pain to fit well).

I don't know about the limma/voom part of that, but the rest is going in the right direction. One would need to put a lot of thought into how voom works and converting that to be relevant for beta-binomials. I suspect that feeding a matrix of per-position/per-sample weights into lmFit along with the logit transformed methylation percentages would produce reasonable results. That'd also be vastly simpler than dealing with beta-binomials, though that should all be tested before anyone actually uses it.