How to reduce the runtime?

[spcspin]

The most time-consuming part of the workflow is the calling variant using the HaplotypeCaller. Therefore we focused on the HaplotypeCaller step.

Here are a few ways to try:

1. HaplotypeCallerSpark:
   HaplotypeCallerSpark is a tool designed by gatk to replace the threading functionality in gatk3. However, it is still in BETA stage, and
   many attempts to use the bee data caused problems. After a discussion with the gatk team, it is confirmed that it is a problem with
   HaplotypeCallerSpark itself.

   The discussion link

2. break down Reference into smaller chunks for HaplotypeCaller:
   scattered intervals based on N masked regions of the reference genome and collecting each intervals calls at the end using
   GatherVcfs tool.

3. Optimize JAVA setting:
   Trying to adjust the parameters related to garbage collection:

   • -XX:ParallelGCThreads
   • Heap Space -Xmx
     There is not much difference in the results.

4. CPU utilization:
   Using the --native-pair-hmm-threads option in HaplotypeCaller there is not much difference in the results.
   3 and 4 can refer to this website

5. Try different variant calling tools:

   • VarScan
   • DeepVariant

[spcspin]

To deal with the low integrity problem when breaking down the Reference genome into smaller chunks for HaplotypeCaller, the GATK team replied as follows:
Depending on how you scatter your intervals it should still hold true. Worst case scenario you may have to run HaplotypeCaller per contig/chromosome which is probably the safest way but if your reference is split by long repeats of N then you may want to split your intervals based on the positions of N repeats.