RNAseq variant filtering criteria

Question

[prasundutta87]

Hello,

Can anyone guide me or share what parameters should be considered for filtering RNAseq variants and why? I haven't got any published material on it as nothing is standardised for this. I am working on a non-model organism (water buffalo) for which there is no truth set data (such as dbsnp, although dbsnp data is itself questionable). If anyone has come across on any document mentioning about RNAseq variant filtering criteria, I would be grateful if it can be shared with me.

I have made some distribution graphs but I am unable to decide a threshold based on that. Parameters chosen- QUAL, DP, GQ (Genotype quality), SP (phred scaled strand bias P value)

PS- Variants called using bcftools mpileup and bcftools call. Kindly let me know if any more information is needed for this question

Answer 1

[grant.hovhannisyan]

Here is the GATK pipeline with some recommendations regarding variant filtering https://gatkforums.broadinstitute.org/gatk/discussion/3892/the-gatk-best-practices-for-variant-calling-on-rnaseq-in-full-detail. But overall the answer on your question is quite tricky, and for different organisms (at least if they are taxonomicaly distant) parameters should be different. In my case (yeasts) GATK hard filtering parameters eliminate 70% of SNP. And I guess that GATK filters are designed for human data (but I am not sure).

[prasundutta87]

Thanks for sharing this. I had already gone through this. You are right that gatk is developed, tested and validated keeping human data in mind. Anything other than human may not perform as expected with gatk pipelines. Furthermore, their RNAseq pipeline is not validated or tested as their dnaseq variant calling pipeline (mentioned in their document).

Answer 2

[andrew.j.skelton73]

As @grant.hovhannisyan shared the GATK RNA Seq best practises, there's not much more you can do than their suggested hard filters. Hard filtering is, by definition, hard. Without truth sets (of which there are none for RNASeq, regardless of species), you can't use VQSR, and therefore have to rely on the hard filters.

[prasundutta87]

I agree. Thats one of the main reasons I am stuck.