DiffBind Normalization ATAC-Seq Question: When to drop a sample?

Question

[John]

The first image with less Differntially Accessible Regions (DARs) was generated by excluding a sample that had an FDR of 0.11 and the second image with more DARs is just not excluding that low FDR sample.

My question is can I get away with using that second image even though the normalization line has a hump in it or is that a sign that I should throw out that sample? I am trying to avoid this as it significantly decreases the number of reads I have to work with and decreases the number of DARs by about 65%.

-sample

VS

+sample

Answer 1

[ATpoint]

Usually, you perform some exploration like PCA and plotting metrics like sample2sample correlation, sequencing depth etc. to see whether a sample is an outlier that merits removal. From the MA-plots you see the lower one has notably larger fold changes and some sort of non-linear bias. That itself could be corrected by something like loess normalization e.g. from the csaw package. Would need to see more plots to decide on removal of samples.

[John]

The PCA plots were kind of a mess. I excluded the sample FOX4, which made the normalization linear but sacrificed many DARs.

[ATpoint]

which made the normalization linear but sacrificed many DERs.

Frankly, I dont think that the "DERs" in the lower plot a real but a normalization artifact. The upper pink points should be much lower.

[John]

Thank you for the advice. I went back and re-normalized using a different method. I used all methods, which I thought applied the loess method, but I am unsure. However, it seems to have helped flatten the curve, although I am unsure of what method was responsible. Found the method here: https://support.bioconductor.org/p/9152726/#9152964

atacV2Loes <- dba.normalize(atacV1Loes, method = DBA_ALL_METHODS, offsets = TRUE)

[John]

I did a little bit of digging and found that I used Deseq2 to normalize.