Sliding window methylation analysis suggestions

shawpa · 09-07-2018, 11:34 AM

I want to do sliding window analysis for differential methylation calculations for WGBS and target bisulfite sequencing libraries. Obviously I am not comparing the 2 library types against each other. I have found the software to make the sliding windows with specific step sizes but my question is how do you know how big your windows should be and then what step to use. I've read papers that use 3Kb windows all the way down to 500bp windows. Is choosing the window and step just arbitrary? I would like to hear others suggestions on what they have done to determine how big their windows and steps are. I am thinking for low coverage WGBS the larger window is probably best to get enough coverage but I am not sure what "step" to use. For my targeted libraries, I have very high coverage so I should probably use a smaller window. I would just like to hear what others would do. Thanks.

simonandrews · 09-10-2018, 02:25 AM

If you have a look at the Differential Methylation Leture document on our methylation training course material and have a look at slides 11 and 12 you can see some information about this.

The short answer is that the best window size is a balance between statistical power and biological effect size. You need the window to be large enough that you get enough data in it to be able to achieve significance, but you need it to be small enough that it doesn't end up larger than the region whose methylation is changing so you don't get averaging against the surrounding regions.

We also don't generally recommend using fixed size windows (eg 500bp), since the uneven distribution of CpGs within the genome means that different windows will have very different numbers of CpGs within them, and you therefore get a statistical power bias which can mean getting biases in the regions of the genome where you find hits. We tend to prefer defining windows based on the number of CpGs then contain, which then means you get a variation in resolution, but a more even statistical power.

shawpa · 09-10-2018, 08:51 AM

If I use fixed windows versus number of CpG windows as you are suggesting, is the concern that I will get no significance or that I will be flooded with regions of significance that are really false positives?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Sliding window SNP density analysis?	balaena	Bioinformatics	0	09-27-2016 07:38 AM
Trimmomatic Sliding Window vs. Removing Adapters	arcolombo698	Bioinformatics	7	12-13-2013 02:57 AM
Trimming with sliding window before alignment	blancha	Bioinformatics	2	05-14-2013 02:21 PM
Multiple testing in a sliding window for Tajima's D	gwilymh	Introductions	0	03-17-2013 04:17 PM
Software for sliding window	mathew	Bioinformatics	6	12-12-2012 10:16 AM

09-07-2018, 11:34 AM	#1
shawpa Member Location: Pittsburgh Join Date: Aug 2011 Posts: 72	Sliding window methylation analysis suggestions I want to do sliding window analysis for differential methylation calculations for WGBS and target bisulfite sequencing libraries. Obviously I am not comparing the 2 library types against each other. I have found the software to make the sliding windows with specific step sizes but my question is how do you know how big your windows should be and then what step to use. I've read papers that use 3Kb windows all the way down to 500bp windows. Is choosing the window and step just arbitrary? I would like to hear others suggestions on what they have done to determine how big their windows and steps are. I am thinking for low coverage WGBS the larger window is probably best to get enough coverage but I am not sure what "step" to use. For my targeted libraries, I have very high coverage so I should probably use a smaller window. I would just like to hear what others would do. Thanks.

09-10-2018, 02:25 AM	#2
simonandrews Simon Andrews Location: Babraham Inst, Cambridge, UK Join Date: May 2009 Posts: 871	If you have a look at the Differential Methylation Leture document on our methylation training course material and have a look at slides 11 and 12 you can see some information about this. The short answer is that the best window size is a balance between statistical power and biological effect size. You need the window to be large enough that you get enough data in it to be able to achieve significance, but you need it to be small enough that it doesn't end up larger than the region whose methylation is changing so you don't get averaging against the surrounding regions. We also don't generally recommend using fixed size windows (eg 500bp), since the uneven distribution of CpGs within the genome means that different windows will have very different numbers of CpGs within them, and you therefore get a statistical power bias which can mean getting biases in the regions of the genome where you find hits. We tend to prefer defining windows based on the number of CpGs then contain, which then means you get a variation in resolution, but a more even statistical power.

09-10-2018, 08:51 AM	#3
shawpa Member Location: Pittsburgh Join Date: Aug 2011 Posts: 72	If I use fixed windows versus number of CpG windows as you are suggesting, is the concern that I will get no significance or that I will be flooded with regions of significance that are really false positives?