Filtering in Picard HSMetrics?

[sgroenewold]

I've been using HSMetrics for a while to get 2X/10X/20X/30X coverage, as well as average probe and target coverage.

I've noticed lately that the 2X/10X/20X/30X numbers start to skew lower for alignments with high duplicate levels. In fact, if I run MarkDuplicates and then re-run HSMetrics, the 2X/10X/20X/30X numbers actually increase, which seems counterintuitive. On top of that, coverage counts that I generate using BEDTools coverageBed don't match the 2X/10X/20X/30X numbers or the average target coverage depth counts I get from HSMetrics.

I've searched high and low and not seen a good explanation of the inner workings of HSMetrics. Is it applying some of its own filtering? Is it perhaps sampling only a portion of the data?

My next steps are to try both HSMetrics and BEDTools with a small target set to see if I can replicate and explain the behavior. After that, I guess I'll try to walk through the Picard code.

But if someone out there knows what might be going on, I'm all ears.

[sgroenewold]

One bump for another shot at a response.

Can anyone shed a little light?

Thanks.

[fashar1]

This is very interesting. What is the Percent_Duplication on your samples? I just ran it with a couple of my samples and don't see a significant difference.