Detection Of Cnvs From Targeted Resequencing Experiments

Question

[Jorge Amigo]

My research group is slowly focusing on targeted resequencing, and although our main interests are SNP and INDEL discovery we would also like to detect CNVs if possible. as far as I know the CNV detectors available use whole genome information to calculate some kind of background noise and average read coverage, and then they go into inferring repeats. but for targeted resequencing I haven't found anything. For that reason, I have 2 questions which I'm placing here joined in this single BioStar question:

1. (bio) Is it possible to appropriately detect CNVs on targeted resequencing? is there really a need to have all the genome covered with reads in order to infer repeats? wouldn't it be enough to look for a coverage average along the targeted regions, and then infer from coverage alterations those possible repeats?
2. (informatics) Is there any program available for CNV detection for targeted resequencing? we currently use BioScope, and its CNV detection module only works on the Whole Genome Resequencing pipeline, refusing to accept data from the Targeted Resequencing one.

Answer 1

[Chris Miller]

CNV detection from targeted resequencing is a difficult problem. Imagine for a second that you're using a capture chip to pull down specific regions of the genome for resequencing. You end up with 100-fold coverage of one region and 50-fold coverage of another. Is that difference caused between a CNV, or by a difference in probe binding affinity? The same basic principle applies for other types of capture.

Really, you have two options:

1) Pair your sample with a matched normal, and divide the former by the later. This should cancel out any site-specific effects.

2) try to model the specific bias for each site so that you can correct for it.

I've seen presentations from people working on both approaches, and they both seem viable. I haven't, however, seen any software released that addresses this problem yet. (Disclaimer - I recently went through a move and started a new job, so I'm woefully behind on reading journals from the last few months).

[Sean Davis]

I have a rough implementation of #1 above. See here: https://github.com/seandavi/ngCGH . It is a little rough, but it does give useable results for paired tumor/normal samples. I typically load the results into Nexus Copy Number or into R for segmentation and downstream processing.

[Jorge Amigo]

thanks Chris for your answer. I was having in mind something like option 1, because when option 2 came to my mind it looked too complex (model particular regions, if you are not restricted to the same ones always, would take a lot of effort). but I haven't found any program nor publication that indeed implements such idea. we are in contact with a dutch research group that are aiming towards a publication on this matter, so we'll see if they finally get to something soon.