1
$\begingroup$

We are trying to use deeptools for analysis of ATAC Seq datasets. We have datasets with different sequencing depths and are wondering if bamCompare's SES based normalization is appropriate for comparing two ATACSeq datasets as opposed to ChipSeq datasets. If such normalization is okay to perform, does it make sense to compare difference in values across comparisons? Let's say I do bamCompare on A and B using SES, and then I do bamCompare (using SES) on A and C. Will it then make sense to compare the two comparisons (A-B & A-C)?

If such normalization isn't okay, what are the best tools to normalize ATAC-Seq datasets with different sequencing depths so that one can perform comparisons like above? Thanks for any suggestions or comments.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

I would be very hesitant to use SES normalization in a case like this. We recommend that people run plotFingerprint first and see if there is good separation between the samples and only then using SES. I would not expect there to be drastic separation between the samples for ATAC-seq data, so I don't think it will work well. Our standard ATAC-seq normalization is RPGC (aka, "1x normalization"). That should work well for your use-case as well.

As an aside, if you really want to make statistically-backed comparisons between groups I can recommend the CSAW package.

Improve this answer
$\endgroup$
7
  • $\begingroup$ Thanks Devon! multiBamSummary tool produces a scalingFactor table for respective BAMs. Does it make sense to first scale with these scaling factors, create bigwig using bamConverge and then use these bigwigs for comparison using bigwigcompare? I am wondering if scaled files go to bigwigcompare, will it make for a better comparison across comparisons. $\endgroup$
    – zerotimer
    Commented Oct 7, 2019 at 15:54
  • $\begingroup$ Yes, that's another good way of approaching things. bigwigCompare does assume that the input bigWigs are scaled appropriately and the procedure you suggest will create a nicely robust comparison. $\endgroup$
    – Devon Ryan
    Commented Oct 7, 2019 at 15:56
  • $\begingroup$ Thanks! You mentioned plotFingerprint for seeing good separation. I did the plotCorrelation heatmap. Is there some approximate threshold for "good separation"? Let's say I have a correlation between 0.5 and 0.7 across different samples. Will this make them ok for SES? $\endgroup$
    – zerotimer
    Commented Oct 7, 2019 at 17:53
  • $\begingroup$ There's no substitute for plotFingerprint here. $\endgroup$
    – Devon Ryan
    Commented Oct 7, 2019 at 20:25
  • $\begingroup$ Devon, Can we use scalingFactor table (from multiBamSummary) to create new scaled BAMs instead of bigwig/bedgraph? I want to perform peak calling only on normalized/scaled BAMs so I can properly compare peaks/footprints across different samples. I know I can downsample using samtools but some scaling factors are greater than 1 so not sure how to "upscale" it. Thanks again! $\endgroup$
    – zerotimer
    Commented Oct 15, 2019 at 19:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.