I am trying to use Sarek to align and call variants on a WGS tumor vs a WES normal (long story). I understand that I will need to use the --wes, is there a standard targets file for WES that is used with Sarek? Are there any other "precautions" that I need to take?

When you use a file with

--intervals

, everything from the BQSR on will be restricted to those intervals. Which file you use for that depends on the kit that was used for the WES. (If you had WGS, you would use by the default GATK file that only excludes centromers, telomers, etc.) Most of these kits come from Agilent (at least the ones I've worked with; if you have something else, you'll have to figure it out). You'll have to find out the exact kit (e.g. order number or similar), because especially for human they have a bewildering array of options and version with very similar names. If you do this often, you can create an account on their website and log in to view all of their prefabricated designs and download the reference data corresponding to one that matches your kit. If in doubt or you don't want an account, you can e-mail them with whatever kit information you have, and they will send you a download link. You get a ZIP file with typically BEDs for the intended targets, everything that is sufficiently homologous to the targets to be likely to be enriched, a padded version of that, etc. I typically use the padded version for the pipeline to be inclusive (since everything outside that will be discarded). If you want/need, you can later use a "narrower" file for filtering. For human these sets come in hg38 and sometimes hg19 as well, so these work well with iGenomes reference files. The only problem is that these BEDs have header lines intended for the UCSC genome browser, as well as extra columns with annotation information, both of which will trip up sarek. But you can easily clean them up with

cut -f 1-3 "$origbed" | grep -ve '^browser ' -e '^track ' > "$cleanbed"

.

Excellent, this is great. Thank you very much!

Adding on to Thomas: It should technically work but I suspect there is a chance of some weird mystery artefacts. Typically you’d want so-called ‘process-matched’ samples and a panel of normals that you could get to correct for common errors unique to the sequencing method, but you have a weird hybrid setup here so, I’m not sure whether that would work. I think you will have to try running it and then be quite cautious about your final calls, really scrutinise anything you find.