CNV by read_depth approach

[MrMimic]

Hey guys !

I'm actually working on a CNV project for the first time and need your help.

I mapped on a ref genome 44 Ilumina pack coming from 44 different individues. I have extract the number of read mapped on each genic feature and try to normalize all of this mess.

I applied the following formula, found in a publication :

NormalizedData = (Number of reads in feature * genome size) / (Total number of read in sample * read size)

They told then :
< 0.05 = Gene isnt present
< 0.5 = deletion
> 1.5 = duplication

What do you guys think of this technique ? I read a lot of possibilities on the net, but what do you usually do with this kind of data ? I heard a fiew words about LDA and tried it on R with MASS package but I got some issues ... And i'm not so good at statistics, so ..

I want to maximize the differences and "group" my values arround thresholds. But I think weak the way I actually do.

Thanks a lot

[maxsalm]

Hi there,
a lot of approaches have already been developed for calling CNVs from Illumina data, and are tailored to different sources of data (WGS, WES, etc.). A good overview can be found here: http://www.biomedcentral.com/1471-2105/14/S11/S1
Normalisation is a key step, as you have noticed, but also often includes additional covariates such as local GC content and mappability. I recommend using some of the existing solutions; from the sound of the problem, cnvHiTSeq may be suitable.

Cheers

[MrMimic]

Your solution looks good, but I work on very large genomes and havent kept bam file from mapping, I just extracted all of the information I wanted, such number of read in each feature with samtools, bedtools, etc.

You think I'll have to remap everything piped with samtools to get lighter files ? Or is there a method to normalize my matrix ?

[maxsalm]

Personally, I always keep the BAM files till the end of the analysis: nearly all of the CNV calling methods I've tried use these as a starting point anyway, so this offers you some flexibility to try different options. Although it is of course feasible to normalise the matrix manually, you will be creating a lot more work for yourself (and potential bugs!) by not simply applying the existing tools unless you have a strong methodological reason not too a priori.

[dpryan]

As I mentioned on biostars, storage is cheap. Just remap and use one of the many excellent and peer reviewed CNV callers out there.