bisulfite sequencing and cluster detection

[dawe]

Hi all, we are trying to perform BS-seq with the protocol described by Smith ZD et al - Methods 2009, which relies on preliminary digestion with MspI.
As the 5' of most of the reads will be the restriction site (CCGG), the base composition will be greatly unbalanced.
We have noticed that RTA can identify a low number of cluster which is not apparently compatible with the images (ok, this is a visual inspection... I may detail it later, sorry!). We asked Illumina and found that A and C are used in the first 5 cycles to "calibrate" the number of clusters... hence we suspect that may be the cause of our worse performance.
I wonder how many of you have noticed something similar and, more important, if the sequences I can get are biased or representative of the whole sample.
This may apply to any restriction enzyme treatment before Illumina run.

d

[fkrueger]

Hi Dawe,

we have recently performed a lot of reduced representation BS-Seq so I'll share what we experienced.

If you prepare a directional MspI-library (i.e. you will only sequence the bisulfite converted forward or reverse strand), all of your sequences should start with either CGG or TGG (depending on it's methylation state). If you have non-directional libraries (we analysed mostly paired-end, non-directional libraries) you will end up with CGG, TGG or CAA as the first three bases.

Up until RTA 1.6, this sequence composition resulted in a lower cluster identification (in a density dependent manner), typically yielding 1-3 million less clusters than you would have gotten without an intially biased sequence. RTA 1.8 uses the first 5 cycles for cluster detection instead of the first 4 for RTA 1.6, which alleviates the problem somewhat. We found that RTA 1.8 deals much better with RRBS data than RTA1.6 (yielding 10-15% more sequences overall), but you can still loose 1-2 million sequences per lane due to the biased sequence composition.

We realised however that, especially for BS-Seq, greater read depth means more methylation calls which are precious. We have devised a method to circumvent the problem of low-diversity base composition at the start of sequences, which will hopefully be published within the next few weeks. The method will require you to save the images of the run in order to work, however. I am happy to give you more information on this if you are interested.

We have not noticed any particular bias in the resulting sequences (other than the fact that RRBS is not very complex compared to shotgun sequencing anyway, but this is the whole idea behind the method). I hope this helps, and please get in touch if you require further details.

Best wishes,
Felix

[dawe]

Hi Felix, thanks for the answer

Hi Dawe,

We realised however that, especially for BS-Seq, greater read depth means more methylation calls which are precious. We have devised a method to circumvent the problem of low-diversity base composition at the start of sequences, which will hopefully be published within the next few weeks. The method will require you to save the images of the run in order to work, however. I am happy to give you more information on this if you are interested.

We have not noticed any particular bias in the resulting sequences (other than the fact that RRBS is not very complex compared to shotgun sequencing anyway, but this is the whole idea behind the method). I hope this helps, and please get in touch if you require further details.

Best wishes,
Felix

I'm looking forward to read the paper, unfortunately we would need to upgrade the workstation attached to the GA: our hardware configuration doesn't allow to save images since SCS 2.8.
I'm glad to read you did not noticed any particular bias.
Just to know, which downstream software have you used to align tags? I'm currently testing Bismark but I still have to figure out if the results are "good" or not.

thanks

d

[hylei]

ask suggestions for new users

Hi, All:

I am totally new to the NGS data analysis. I want to get more idea about how to begin with the NGS, and where to get the practice data? Thanks.

hylei

[fkrueger]

I'm looking forward to read the paper, unfortunately we would need to upgrade the workstation attached to the GA: our hardware configuration doesn't allow to save images since SCS 2.8.

If you still have some version 4 chemistry kits in your lab there is currently a nice workround: just do the cluster generation and the run with with SCS2.6 which still allows you to save the images, and then once the run is finished you use the OLB1.8 on the images. Works flawlessly, I have just done it yesterday and today with very biased libraries and 3 RRBS libraries as well. If you are interested I can even send you some comparisons between 1.6, 1.8 and 1.8 + bareback-processing (which is short for barcode-backprocessing (https://www.bioinformatics.bbsrc.ac....ects/bareback/).

And yes, we are using Bismark for mapping and methylation calling (if you get stuck or want some advice on Bismark I am happy to help). Most of the further downstream analysis is then carried out in SeqMonk or custom scripts.

[dawe]

If you still have some version 4 chemistry kits in your lab there is currently a nice workround: just do the cluster generation and the run with with SCS2.6 which still allows you to save the images, and then once the run is finished you use the OLB1.8 on the images. Works flawlessly, I have just done it yesterday and today with very biased libraries and 3 RRBS libraries as well. If you are interested I can even send you some comparisons between 1.6, 1.8 and 1.8 + bareback-processing (which is short for barcode-backprocessing (https://www.bioinformatics.bbsrc.ac....ects/bareback/).

And yes, we are using Bismark for mapping and methylation calling (if you get stuck or want some advice on Bismark I am happy to help). Most of the further downstream analysis is then carried out in SeqMonk or custom scripts.

I didn't realized it was you :-) I've seen bareback page when I've started this BS-seq analysis and realized I couldn't use it with my instrumentation :-( I probably push some bosses to upgrade the GA workstation if we will have more BS-seq experiments...
I probably will write you directly asking for details of SeqMonk usage with Bismark results.
About the chemistry... I have to ask, I'm not dealing with the instrument directly (yes, I'm a bioinfo).
Thanks!

d

[kalyankpy]

RRBS sequencing!

Dear Felix,

I have prepared RRBS libraries recently (My first NGS library) and am waiting for data! Meanwhile I downloaded K562-rep2 (Myers lab RRBS data) to practise data analysis and understand how the biasness look like. I was going through one of your reply where you mentioned that one will notice only CGG, TGG and CAA in case of non-directional RRBS sequencing. However, I am not convinced (dont understand) with this statement. I can understand how CGG and TGG dominate in case of directional sequencing. But as per my understanding there will be 32 possibilites (2[C/T]*4[A/T/G/C]*4[A/T/G/C]) in case of non-directional sequencing. I am attaching a slide which will show two (of 32) possibilites. Could you explain me if I understand the library preparation and data properly!

Hi Dawe,

we have recently performed a lot of reduced representation BS-Seq so I'll share what we experienced.

If you prepare a directional MspI-library (i.e. you will only sequence the bisulfite converted forward or reverse strand), all of your sequences should start with either CGG or TGG (depending on it's methylation state). If you have non-directional libraries (we analysed mostly paired-end, non-directional libraries) you will end up with CGG, TGG or CAA as the first three bases.

Up until RTA 1.6, this sequence composition resulted in a lower cluster identification (in a density dependent manner), typically yielding 1-3 million less clusters than you would have gotten without an intially biased sequence. RTA 1.8 uses the first 5 cycles for cluster detection instead of the first 4 for RTA 1.6, which alleviates the problem somewhat. We found that RTA 1.8 deals much better with RRBS data than RTA1.6 (yielding 10-15% more sequences overall), but you can still loose 1-2 million sequences per lane due to the biased sequence composition.

We realised however that, especially for BS-Seq, greater read depth means more methylation calls which are precious. We have devised a method to circumvent the problem of low-diversity base composition at the start of sequences, which will hopefully be published within the next few weeks. The method will require you to save the images of the run in order to work, however. I am happy to give you more information on this if you are interested.

We have not noticed any particular bias in the resulting sequences (other than the fact that RRBS is not very complex compared to shotgun sequencing anyway, but this is the whole idea behind the method). I hope this helps, and please get in touch if you require further details.

Best wishes,
Felix

[fkrueger]

Dear kalyankpy,

it looks to me like you left out the end-repair and A-tailing step in the schematic drawing you attached. I have prepared a couple of slides some time ago which try to describe potential problems with (non-directional) libraries a bit better. I will attach it here, please don't hesitate to contact me directly via email ([email protected]) if you got further questions.

Cheers,
Felix