differential methylation region analysis without replicated

[zinky]

hello everyone!
I got two samples bisufite sequencing data(each about 10G from illumina platform).after rowdata QC and alignment as well as methylation calling by Bismark,now i have methylation level in each sample's C (CpG,CHH,CHG region) and wanna perform downstream analysis such as DMR.
here are my question:
1,how to identify DMRs between two sample without replicated?
2, any tools can do this kind of analysis?
3,Is there any wrong within my analysis workflow?(ps:i am a green hand)

[simonandrews]

In general there are two approaches you can take to call differential methylation from this kind of data.

1) You can call differences at the level of individual Cs by comparing the proportion of meth:unmeth in your two samples using a contingency test such as a Fisher's exact or Chi-square test. The problem with doing this is that you have to apply a large amount of multiple testing correction so in order for something to be significant after correction you need a high level of observation for each C to be able to achieve sufficiently low raw p-values. You can extend this test to larger regions by summing up the meth:unmeth counts over the region but this isn't ideal as the representation of each C won't be the same so your result may be unduly biased by a subset of positions within the region.

2) Alternatively you can calculate a percentage methylation for each C (maybe filtering for coverage level so you only use Cs which have a sensible level of observation). Then you can group together Cs falling within a larger region and compare the distribution of methlyation percentages between your samples using a t-test (or ANOVA if you have more than 2 samples). This test is fairer in that it represents each C equally, but may be problematic in that you get more power for your test if you have more Cs in a given context in your analysed region.

We've been incorporating some of these kinds of tests into SeqMonk, but it's only in our internal development version at the moment and hasn't made it into a public release. There are some R packages which I think do something similar though we've not used these ourselves.

[zinky]

thanks Simon! I think I got your piont ,the method to detect DMR depends on the region i intend to analysis . Anyway ,i will try both way first!

[mariebreen]

Hi all,

I was just wondering if anyone knew what coverage you would need in a sequencing project to detect a 10% difference in methylation?

I am currently optimizing the Methyl-Seq protocol and I am trying to figure out how many samples I can multiplex efficiently.

Thanks,
Marie

[zinky]

Hi all,

I was just wondering if anyone knew what coverage you would need in a sequencing project to detect a 10% difference in methylation?

I am currently optimizing the Methyl-Seq protocol and I am trying to figure out how many samples I can multiplex efficiently.

Thanks,
Marie

it's hard to say , maybe you can draw a saturation curve for certain kind of data by a coverage gradient. In my view, different samples have unequal meth-level, there are no standard or minimum coverage limited to dectect such diff-meth proportion.
waiting for further explanation。

[simonandrews]

Hi all,

I was just wondering if anyone knew what coverage you would need in a sequencing project to detect a 10% difference in methylation?

I am currently optimizing the Methyl-Seq protocol and I am trying to figure out how many samples I can multiplex efficiently.

You should be able to run a standard power analysis to work this out. It will depend on the size of the genome you're using, the size of the area you're looking at and the distribution of CpG densities (assuming CpG methylation).

In any case you're going to need pretty high coverage to detect this with any sensitivity. My gut feeling says you'd need an average coverage of >50X which if you're using a big genome isn't going to be cheap.