Re: [svtoolkit-help] Filter criteria for SVs

[Bob H. <han...@br...>]

First, I should warn you that I haven't done too much work with 
discovery on targeted sequencing.
Your results may vary depending on how even your coverage is across the 
target region.

Filtering good calls from the candidates is still somewhat of an art, 
and it depends on whether you want more sensitivity or more specificity.
There is a default set of filters in the queue scripts is based on the 
1000 genomes pilot.

The metrics most often used for filtering are the following:

DEPTHRATIO / DEPTHPVALUE
The depth ratio is the mean read depth for samples with observed 
aberrant read pairs divided by the mean read depth for the other samples.
It should ideally be 0.5 or below for real deletions.  If you plot this 
metric it should be bi-modal and you can select a threshold from the data.
The depth p-value indicates whether there was sufficient depth 
information for depth ratio to be reliable.
Default 1kg pilot filters: DEPTHPVALUE < 0.01 and DEPTHRATIO < 0.63 (or 
DEPTHRATIO < 0.8 if MEMBPVALUE < 0.01)
The cutoff of 0.63 was chosen as the approximate midpoint of the bimodal 
distribution for depth ratio in the 1kg pilot data set.
See below for MEMBPVALUE.

DEPTHCALLTHRESHOLD < 1
The not-very-well-named depth call threshold is the median normalized 
sequencing depth of samples with observed aberrant read pairs.
A normalized depth of 1 in this case should be approximately copy number 
2.  Ideally this number would be 0.5 or below and this filter
excludes regions of the genome with excessive coverage.

COHPVALUE > 0.01
The coherence metric (not a true p-value, despite the name) indicates 
whether the spacing of the aberrant read pairs are consistent with a
single deletion breakpoint.  Read pairs generated by mismapping, for 
example, tend to be more uniformly spaced.

MEMBPVALUE
This metric tests whether the deletion seems to be appearing more in 
some samples than others, taking into account uneven sequencing.
Lower values are better, but unless you have a lot of samples it can be 
hard to find a good absolute cutoff for this metric.
For the 1kg pilot, what we did was to use this to "boost" some samples 
with a marginal depth ratio between 0.63 and 0.8.

If you are trying to identify high confidence calls, in general longer 
calls tend to have better depth signal and thus be of higher confidence
(all other metrics being equal).  In the 1kg phase 1 data set, we also 
used a filter where we required at least one sample to have two aberrant
read pairs.  This may be important if your sequencing is low coverage 
(e.g. 4x) but for higher sequencing depth I think this is not necessary.

I would start with the 1kg pilot filters as a guide, as they proved to 
be reasonable on the 1kg phase 1 data as well
(see SVDiscoveryDefaultFilter in SVQScript.q).

In theory, it should be possible to calibrate your filters based on 
existing gold standard data sets, if you have any.
Another useful thing to do is to prospectively genotype some of the 
sites (using Genome STRiP).
Sometimes the genotyping results and metrics can be used to help 
determine whether marginal calls are good or not and this can help
influence your discovery thresholds, although I would not recommend 
trying to do large scale filtering via genotyping.

-Bob

On 9/5/11 4:56 PM, Hyun Ji Noh wrote:
> Hi,
>
> Now I have vcf files that are generated by modified discovery.sh script. There's huge amount of information in the vcf files and there are even deletion calls that are not in my target region. So I'm wondering what is your recommended criteria to filter high quality deletion calls?
>
> BW,
> Hyun Ji
>
> P.S. Thank you for you help on mismatched read pair records error. Your advice fixed the problem!
>
>
> ------------------------------------------------------------------------------
> Special Offer -- Download ArcSight Logger for FREE!
> Finally, a world-class log management solution at an even better
> price-free! And you'll get a free "Love Thy Logs" t-shirt when you
> download Logger. Secure your free ArcSight Logger TODAY!
> http://p.sf.net/sfu/arcsisghtdev2dev
> _______________________________________________
> svtoolkit-help mailing list
> svt...@li...
> https://lists.sourceforge.net/lists/listinfo/svtoolkit-help