Bisulfite sequencing - filtering by min. conversion rate

[mixter]

Hi,

I am looking into filtering bisulfite reads by a minimum conversion rate. Something high like a 95% CpG and non-CpG conversion rate. I've been working with Bismark and really like it.

I know that I get conversion rates in the end results, but I would like to also filter individual reads by conversion rate, either before or after methylation calling / mapping.

Are there any available tools that could do it, or otherwise a suggested way to do this?

Many thanks!

[simonandrews]

It should be fairly simple to do this by parsing the methylation call string in each line of the bismark output, but I guess I would also ask why you wanted to do this. I'm aware that some groups have applied this filter in the past (though only ever in non-CpG context, requiring full conversion in CpG context would definitely be a mistake), but that was under the assumption that there is effectively no non-CpG methylation, which increasingly appears to not be the case. By removing highly (or even moderately) methylated reads you run the risk of biasing your results and potentially removing interesting data. If non-conversion of specific reads does happen in your library then it should only be a problem if it's targeted in some way, otherwise randomly distributing a few methylated base calls shouldn't bias your results too much.

What we do filter for in our analyses is regions which show unusually high coverage. Mismapping of repetitive regions does happen, and can produce odd results, but this type of filtering removes a region of the genome from the analysis, rather than individual reads.

[mixter]

Dear Simon,

What you wrote here made a lot of sense and I have meanwhile adopted this practice of not filtering reads to correct for conversion rate. There are several options I'm exploring, including correcting observed methylation levels and filtering for unusual coverage peaks. I'd like to argue for that in future papers and that it does not make sense to throw away significant amounts of the data due to suspected low conversion.

Are there any studies that are helpful to support the approach of not filtering for highly converted reads? Otherwise, would you perhaps agree with a reference to our correspondence?

Thanks

[simonandrews]

I'm not aware of any studies which have looked in a systematic way at the dynamics of bisulphite conversion so I'm not sure there's a fixed conclusion one way or the other. We therefore stick with the more conservative approach of not biasing our data by systematically removing parts of the data which we have no actual evidence are wrong.

Removing overrepresented sequences is easily justified since we know that we can't get that many correct sequences from a region - therefore we must be mis-measuring our data in those regions and we can therefore ignore them.