using haplotypecaller for 5.5mb of targeted capture across ~2400 bams
our group is calling variants using 5.5mb of targeted capture across ~2400 bams (for 830 subjects). i have successfully executed using unifiedgenotyper, but we would also like to run using haplotypecaller to compare. my basic approach has been to scatter the capture region into 25 equal sized chunks, run parallel and then CatVariants to gather afterward. this worked nicely for UG, but for HC estimated run-times for even these small intervals are >7 days and the jobs inevitably crash due to memory problems.
so, my question is how should i approach this large cohort problem in HC? i have read HC works best in cohorts <100, so should i divide up my cohort into chunks of 100 bams and combine afterward? seems this would undermine the benefits of joint calling, no? i have also read that GATK 3.0 is going to address this issue with HC -- have these improvements been made with the current version (which I am not currently using)? if so, is there a best practice or example available anywhere with cmd specs for situations similar to mine? finally, is it ok to split my capture region and run in parallel as i have done? this doesn't undermine the benefits of de novo local assembly by HC? thanks in advance for any input!