Run Time Performance of Variant (SNP) Calling Tools?

[CindyF]

Hi all,

I read lots of papers about SNP calling and the different implementations (GATK, SAMtools, SOAPsnp, etc) in the last time. Something I noticed was that in no paper run time performances where presented, neither in the papers introducing new SNP callers nor in surveys.

The only thing I came across where comments like "xy is comparably slow" (e.g., about SAMtools or GATK), but I cannot find exact numbers for run time performances anywhere.

Does anyone know why this is the case? Despite the comments I found, are SNP calling tools maybe that performant that run time performance is nothing to speak of?

Or even better, can anyone point me to any resource (papers, websites, forums, etc) where SNP calling tools have been compared to each other regarding run time performance or at least provide ANY data about run time performance?

Thank you,
CindyF

[lh3]

Brad Chapman has a blog post having numbers you may be interested in:

https://bcbio.wordpress.com/2013/05/22/scaling-variant-detection-pipelines-for-whole-genome-sequencing-analysis/

For samtools, calling a single human chr1 at 60X coverage took overnight on a single 2.3GHz Opteron CPU. I do not have the exact timing. Disabling BAQ will make samtools several times faster at minor cost of accuracy.

[CindyF]

Hi lh3,

thanks for the link, I have not come across it yet. It is quite interesting to see that run time performance indeed seems to be a challenge for variant calling.

Regarding your overnight samtools run - did you call SNPs and Indels? Did really only the variant calling without preprocessing (base quality score recalibration etc.) took that long?

Regards,

CindyF

[lh3]

Variant calling usually include both SNP calling and INDEL calling. In the old days, samtools was usually faster than GATK. It still seems faster than GATK used in Brad's post.

I don't think variant calling is slow. Brad's post explains well: variant calling alone takes ~25% of total CPU time starting from alignment. Alignment and post processing are both slower than variant calling. That is why samtools/gatk/etc add steps known to greatly impact the speed but with only marginal gains in accuracy. As long as variant calling is not the bottleneck, we can afford that.