How to detect CNV with panel sequencing?

Question

[chen]

It's relative simple to detect CNV with whole genome sequencing or whole exome sequencing, but how to do it with panel sequencing?

For example, to detect ERBB2 gene amplification (copy number gain) with a panel sequencing (i.e. a 100 genes panel), I can count the reads mapped into ERBB2 and other genes and calculate its relative gain/loss, but here comes the problems:
1, capturing has bias
2, PCR has bias
3, The gene region is too small, so this naive method is very instable.

Questions:
a), any good way to do this?
b), how to design the panel? Shall I add some CONTROLs into the panel?
c), how to do normalization/correction so that the result will be stable?

Currently I do intra-sample normalization to normalize sequencing depth, and calcate the average value within different samples (with a same panel), then compare value of current sample to the average values. This method is expected to work, but the results show big difference with other clinical methods like FISH. Why?

Thanks

Answer 1

[Irsan]

Like you noticed, there is a lot of systematic bias in read counts between regions because of many sequence-specific reasons. Most well-known factors are GC content and mappability, which drastically influence the number of reads obtained from a particular region. Fortunately most of the between-region bias is systematic and can therefore be predicted and normalized for. To do so (normalize for region specific trends) you need a cohort of control samples (meaning they have diploid genomes) and train your model to recognize the systematic trends and correct for that in your samples of interest. In the best situation you have patient-matched controls that you can use to correct for copy number polymorphisms specific for the patient from which the tumor sample was derived from.

Answer 2

[Garan]

We use a custom targeted panel ( Ellard, Sian, et al. "Improved genetic testing for monogenic diabetes using targeted next-generation sequencing." Diabetologia 56.9 (2013): 1958-1963. ) which has it's baits balanced during a design phase (calculated from coverage data from Exome library preps from the same manufacturer / technology - we're trying to get balanced read depth across the regions).

From there we use ExomeDepth with bespoke interval sizes (set according to the read length) and a set of selected reference samples (which we've confirmed not to contain CNVs over ther targeted regions).

Answer 3

[Chris Miller]

To add to what Irsan said, you'd ideally:

1) add spaced out control probes to expand the number of areas that are covered beyond just your 100 genes

2) use a pool of normal samples as a comparator to help normalize coverage

As far as tools, a collaborator of mine wrote CC2 (copycat2) to solve this problem, and I've seen some impressive results from panels with about 250 genes. https://github.com/abelhj/cc2/

I don't think it's very polished or user-friendly yet, but if you can get it to run, it may be helpful.

[lffu_0032]

can you give more details about the algorithms in CC2?

Answer 4

[l.johansson]

We have developed CoNVaDING to detect (single exon) copy number changes in targeted NGS gene panels.In addition I can recommend XHMM, When using parameters 1e-06 for exome-wide CNV rate and 2 for mean number of targets in CNV, the sensitivity is really high.

Both tools have a high sensitivity and specificity. Specificity is especially high when the both tools are in agreement.

http://molgenis.github.io/software/CoNVaDING

http://atgu.mgh.harvard.edu/xhmm/tutorial.shtml

[lffu_0032]

for panel sequencing, how to adjust the XHMM parameters properly？and why "2 for mean number of targets in CNV" is high sensitivity for XHMM?

[lffu_0032]

For germline cnv calling by using capture panel data, i recommend CoNVaDING