How to filter bam file in structural variants detection with short reads?

Question

I have WGS resequencing data sequenced on an Illumina platform. The length of the reads is about 150bp. I have aligned them to the reference genome using bwa-mem2 and deleted the PCR duplicates using Picard. As the next step, I plan to filter the bam file, but I haven't found a comprehensive workflow for this, so I don't know the factors I should take into account.

I try to summarize the insert size distribution and found that there is a peak around insert size 0. But I think it may be normal as the existence of variation.

The GATK best practice for stuctural variant detection is not detailed enough, and I don't plan to use the GATK workflow; GATK uses manta, but I plan to use the smoove or delly workflow.

I want to know what factors I should take into acount in BAM file filtering for SV calling.

Do I need to exclude alignments with low quality or abnormal distribution of insert size? Is the filtering necessary? If so, what factors do I need to pay attention to?

I have multiple samples data instead of one. I just want to know whether I need to do filtering, as recently published work Next-generation data filtering in the genomics era has mentioned the filtering of alignment for variants calling, and had stressed the importance of filtering.

I have read this paper and it mentioned we should be concerned about the quality of alignment and read depth or coverage for SNP calling or SV calling. In the GATK best practice for SNPs, it seems there is a need to do local realignment for BAM files. I have tried to do this for SV calling and it didn't not produce a substantially different result.

Answer 1

You don't need any further special filtering to call SVs. Just follow the standard bam cleanup approach for small variants, and once you have your bam, feed it to delly or any other caller you want to use.

Local realignment is almost certainly not needed with modern sequencers. In fact, GATK don't recommend it any more as far as I know. But sure, you need to filter, but I don't think you need anything special for SVs. Of course you should check the coverage and quality etc, but the variant caller also takes these into account, so many prefer to filter on the variants called instead of on the bam file. There's no golden bullet, I'm afraid.

The important bit is to be aware that not only is short read NGS a bad (OK, a limited) tool for SV discovery, doing it with a single sample instead of a cohort is the least powerful approach. Don't get me wrong, I am not saying you shouldn't do it or that it isn't worth doing, only that you should be aware of the limitations and not expect to get a comprehensive list of all SVs your sample may have.

If you can, try to analyze multiple samples (ideally from the same sequencing run) together as that will significantly improve the accuracy of your results.