Depth higher after subsampling?

[emham]

Hi, I've been noticing that my median deduplicated depth goes up after sub-sampling (with samtools view -s) compared to the original median depth. Wondering if anyone else has observed something similar and if it's an artifact of samtools view -s? I don't think it is an issue with picking the same seed in multiple sequential sub-samplings because the original bam that is being sampled from was not already sub-sampled previously.

My workflow is:
complete non-deduped bam file -> sub-sample or no sub-sampling -> barcode deduplication -> calculate median depth

[Ben3]

emham I don't have an answer at the moment, but let me see if I can find something helpful.

[Ben3]

Hello emham

A few things I've been told. Subsampling before deduplication will bias your downstream analysis. This eliminates many low coverage areas by chance, and so you should expect the median to go up.

Are you using UMIs? They should reduce this kind of artifact and are used for deduplication.

I was specifically told, "If your analysis is sensitive to depth, consider whether non-UMI deduplication could make it worse instead of better: if your reads aren't uniformly distributed over the whole genome (e.g. you did targeted sequencing or RNA-seq or ATAC-seq or anything else like that), non-UMI deduplication is likely to incorrectly distort high-coverage regions to have lower coverage (pre-PCR duplicate molecules that occur by chance will be lost as false-positive PCR duplicates) and if the reads are really densely concentrated even short UMIs can start to run into problems"

Samtools shouldn't be wrong. If you want to understand what's happening just open up your BAM file, with duplicates marked but not removed, in IGV and take a look at interesting regions.

I have two final questions:

Why are you subsampling? And why/how are you deduplicating?