# Pooling samples for SV calling

Pavel Dimens

●

Published 2024-08-01

One of the cool benefits of linked-read data is the fact that you can call structural variants with it. Depending on the depth of your data, you may want (or need) to pool samples together. This page walks you through why you may want to do that and the logic of doing so.

# Sample depth

# Depth, explained

In bioinformatics, the terms "coverage" and "depth" and often used interchangeably, which is incorrect and leads to confusion. Coverage refers to the proportion of a genome that is sequenced, and depth refers to the number of sequences present at a given genomic position. Think of coverage as horizontal and depth as vertical. If depth is given as "15X", that means every sequenced genomic position has about 15 sequences per position (a.k.a. 15 replicates per locus). For clarity, when we say "depth", we are referring to this "vertical" description of replicates per locus.

The difference between depth and coverage. The locus on the left would be considered 5X.

# Depth, in context

Historically, one would have wanted to sequence fewer individuals at higher depth to get confident genotype calls, rather than sequence more individuals at lower depth. Recent advances in bioinformatics have enabled low-coverage whole genome sequencing a.k.a. lcWGS to be a viable option for studies of outbred non-model systems. More and more studies are emerging that have adopted lcWGS, which is pretty cool. However, calling structural variants at low depth is challenging, especially with short reads.

# The problem

It's recommended to have at least 10X-12X depth to get decent structural variant calls (definitely read that in a paper that I would like to link here, but I can't seem to find it). If your data already has a minimum of 10X for each individual, great! Feel free to use variant callers like naibr and leviathan to identify structural variants. However, if you opted to sequence more individuals at lower depth (lcWGS is often between 0.5-5X), then calling structural variants in individuals may be a challenge.

# The solution

One way to get your low-coverage (low depth) data and still call structural variants is to pool samples together, which would effectively boost the depth. By doing this, you will no longer be able to make per-individual assessments, which can be fine depending on the nature of your study.

# Pooling considerations

If pooling samples, you must pool them sensibly and with a biological context to do so. In other words, you don't just pool random samples together to inflate depth. Since haplotag data is just whole genome sequence data plus a little extra information, you should use the SNPs of your data to first identify genetic clusters as per standard population genetic practices. Once population structure/stratification has been identified, you can use that as a basis to pool together samples from these groups. As an example, you can use pcangsd to perform a PCA on low-depth SNP data and get results like these:

PCA of Alosa sapidissima (SNPs from low-depth haplotag dataset)

Given these results, a sensible pooling strategy may be:

Pool	Sample Locations
1	Miramichi
2	Annapolis
3	St_Johns
4	Santee_Cooper
5	Roanoke, Chowan-Blackwater
6	Potomac
7	Delaware
8	Hudson
9	Connecticut
10	Merrimack, Kennebec

Do not concatenate manually!

It would seem natural to use samtools cat to combine alignment files quickly and easily, but do not do this. The harpy sv workflows will intelligently concatenate files and will make sure every individual will have unique MI values that are not shared with any other individual in the pool. If you need to concatenate linked-read alignment files outside of a workflow, use concatenate_bam.py shipped with Harpy instead of samtools cat or other similar tools.

The reason is, samples aligned using Harpy have their linked-read barcodes deconvolved and supplemented with a unique molecule ID, given as an MI:i SAM tag in the alignment records. The molecule ID's start at 1 for each sample and increment for every identified unique molecule. What this means is that doing a typical concatenation via samtools cat will not resolve conflicting molecule IDs.

Think of it this way, every sample has MI's from 1 to N, and as soon as you do a basic concatenation, the MI 1..N from sample2 will immediately conflict with 1..N from sample1. Why? Well, it's impossible that MI:i:1 from sample1 and MI:i:1 from sample2 came from the same molecule. A basic concatenation will flood the resulting file with clashing MI tags, which will make it impossible for a linked-read aware SV caller to make sense of the data and do its job well.