negative CNN_1D scores

[MarionFresquet]

Hello,

I used the Sarek pipeline to detect germline variants ( working with 30X WGS). After variant calling with haplotypecaller, the pipeline runs CNNScoreVariants and FilterVariantTranches. However, the number of variants in the filtered and unfiltered VCF files is the same.

When I checked the CNN_1D scores in the INFO section, most were negative, like -4.993, -5.226, 0.447; -7.3, -3.864, -8.959...

is this normal as an output of CNNSCOREVARIANTS ?

Thank you for your help !

[gokalpcelik]

Hi @MarionFresquet
Did you use the docker version of GATK? If not, how was gatk python environment created in the compute environment?
CNNScoreVariants tool, although deprecated, relies on gatk python environment with certain versioned dependencies of tensorflow. If those versions do not match what gatk uses then it is possible that scores are not as expected. Can you also try NVScoreVariants from version 4.6.1.0 and see if the issue persists? Keep in mind that conda environment must be created based on the documentation we provide.

Can you check on that?
Regards.

[MarionFresquet]

Thank you for your response.
The Sarek pipeline use GATK with Singularity as a Docker. It seems to me that the pipeline shouldn't have any dependencies problems and the tools should match.
Best regards