somalier: extract informative sites, evaluate relatedness, and perform quality-control on BAM/CRAM/BCF/VCF/GVCF

Quick Start

somalier makes checking any number of samples for identity easy directly from the alignments or from jointly-called VCFs:

The first step is to extract sites. For VCF just use:

somalier extract -d extracted/ --sites sites.vcf.gz -f /data/human/g1k_v37_decoy.fa $cohort.vcf.gz

with a sites file from releases

For BAM or CRAM, use: This is parallelizable by sample via cluster or cloud, but here, using a for loop:

for f in *.cram; do
    somalier extract -d extracted/ --sites sites.vcf.gz -f /data/human/g1k_v37_decoy.fa $f
done

--sites is a VCF of known polymorphic sites in VCF format. A good set is provided in the releases but any set of common variants will work.

:warning: somalier can work on GVCF and individual VCFs, but it is recommended to extract from bam/cram when possible. It is also good to extract from a jointly-called VCF/BCF when only looking within that cohort. While extracting from a single-sample VCF is possible (with --unknown) and GVCF is also supported, these options are less accurate and more prone to problems.

The next step is to calculate relatedness on the extracted data:

somalier relate --ped $pedigree extracted/*.somalier

This will create text and interactive HTML output that makes it fast and easy to detect mismatched samples and sample-swaps. The files created are:

• `{output_prefix}.samples.tsv # creates a .ped like file with extra QC columns
• `{output_prefix}.pairs.tsv # shows IBS for all possible sample pairs
• `{output_prefix}.groups.tsv # shows pairs of samples above a certain relatedness
• `{output_prefix}.html # interactive html

Example html output is here

Note that the somalier relate command runs extremely quickly (< 2 seconds for 600 samples and ~1 minute for 4,500 samples) so it's possible to add/remove samples or adjust a pedigree file and re-run iteratively.

For example to add the n + 1th samples, just run somalier extract on the new sample and then re-use the already extracted data from the n original samples.

For huge sample-sets, if you run into a bash error for argument list too long, you can pass the somalier files as quoted glob strings like: "/path/to/set-a/*.somalier" "/path/to/set-b/*.somalier".

Example Output

• Interactive output from somalier relate is here
• Interactive output from somalier ancestry is here

Infer

somalier can also infer first-degree relationships (parent-child) when both-parents are present and can often build entire pedigrees on high-qualty data. To do this, use somalier relate --infer ... and the samples.tsv output will be a pedigree file indicating the inferred relationships.

See wiki for more detail.

Usage

The usage is also described above. Briefly, after downloading the somalier binary and a sites vcf from the releases run:

somalier extract -d cohort/ --sites sites.hg38.vcf.gz -f $reference $sample.bam

for each sample to create a small binary file of the ref and alt counts for the variants listed in sites.hg38.vcf.gz.

for a vcf, run:

somalier extract -d cohort/ --sites sites.hg38.vcf.gz -f $reference $cohort.bcf

somalier can extract from a multi or single-sample VCF or a GVCF. This will be much faster, in cases where it's available, this would look like:

somalier extract -d extracted/ --sites sites.vcf.gz -f /data/human/g1k_v37_decoy.fa joint.vcf.gz

following this, there will be a $sample.somalier file for each sample in the joint.vcf.gz

Note that somalier uses the AD field to extract depth information. If that FORMAT field is not present in the header, then it will use the genotypes only and use a total depth of 20 (10,10 for heterozygote), etc.

Then run:

somalier relate --ped $pedigree_file cohort/*.somalier

This will create an html file for QC in a few seconds.

Note that if a new sample is added to the cohort, it's only necessary to perform the extract step on that sample and then run the (fast) relate step again with all of the extracted files.

Extended Usage

For each command of somalier, extended parameters are listed in --help of each subcommand.

$./somalier --help
Commands:
  extract      :   extract genotype-like information for a single sample from VCF/BAM/CRAM.
  relate       :   aggregate `extract`ed information and calculate relatedness among samples.
  ancestry     :   perform ancestry prediction on a set of samples, given a set of labeled samples
  find-sites   :   create a new sites.vcf.gz file from a population VCF (this is rarely needed).

somalier extract

extract genotype-like information for a single-sample at selected sites

Usage:
  somalier extract [options] sample_file

Arguments:
  sample_file      single-sample CRAM/BAM/GVCF file or multi/single-sample VCF from which to extract

Options:
  -s, --sites=SITES          sites vcf file of variants to extract
  -f, --fasta=FASTA          path to reference fasta file
  -d, --out-dir=OUT_DIR      path to output directory (default: .)
  --sample-prefix=SAMPLE_PREFIX
                             prefix for the sample name stored inside the digest

somalier relate

calculate relatedness among samples from extracted, genotype-like information

Usage:
  somalier relate [options] [extracted ...]

Arguments:
  [extracted ...]  $sample.somalier files for each sample. the first 10 are tested as a glob patterns

Options:
  -g, --groups=GROUPS        optional path  to expected groups of samples (e.g. tumor normal pairs).
specified as comma-separated groups per line e.g.:
    normal1,tumor1a,tumor1b
    normal2,tumor2a
  --sample-prefix=SAMPLE_PREFIX
                             optional sample prefixes that can be removed to find identical samples. e.g. batch1-sampleA batch2-sampleA
  -p, --ped=PED              optional path to a ped/fam file indicating the expected relationships among samples.
  -d, --min-depth=MIN_DEPTH  only genotype sites with at least this depth. (default: 7)
  --min-ab=MIN_AB            hets sites must be between min-ab and 1 - min_ab. set this to 0.2 for RNA-Seq data (default: 0.3)
  -u, --unknown              set unknown genotypes to hom-ref. it is often preferable to use this with VCF samples that were not jointly called
  -i, --infer                infer relationships (https://github.com/brentp/somalier/wiki/pedigree-inference)
  -o, --output-prefix=OUTPUT_PREFIX
                             output prefix for results. (default: somalier)

Note that for large cohorts, by default, somalier relate will subset to interesting sample-pairs so as not to balloon the size of the output. By default, pairs that are expected to be unrelated and have a relatedness <= 0.05. If you wish to force all samples to be reported, then set the environment variable SOMALIER_REPORT_ALL_PAIRS to any non-empty value, e.g. export SOMALIER_REPORT_ALL_PAIRS=1

find-sites

To create a set of new sites, use somalier find-sites on a population VCF. More info on this tool is available in the wiki

Install

Somalier is available via bioconda, see here.

Alternatively, you can get a static binary from here.

Users can also get a docker image here which contains htslib and a somalier binary ready-for-use.

How it works

somalier takes a list of known polymorphic sites. Even a few hundred (or dozen) sites can be a very good indicator of relatedness. The best sites are those with a population allele frequency close to 0.5 as that maximizes the probability that any 2 samples will differ. A list of such sites is provided in the releases for GRCh37 and hg38.

In order to quickly calculate genotypes at these sites, somalier assays the exact base. The extraction step is done directly from the bam/cram files 1 sample at a time.

The relate step is run on the output of the extract commands. It runs extremely quickly so that new samples can be added and compared. It uses 3 bit-vectors per sample for hom-ref, het, hom-alt. Each bitvector is a sequence of 64 bit integers where each bit is set if the variant at that index in the sample is for example, heterozygous. With this setup, we can use fast bitwise operations and popcount hardware instructions to calculate relatedness extremely quickly.

For each sample-pair, it reports:

1. IBS0 -- the number of sites where one sample is hom-ref and another is hom-alt
2. IBS2 -- the number of sites where the samples have the same genotype
3. shared-hets -- the number of sites where both samples are heterozygotes
4. shared-hom-alts -- the number of sites where both samples are homozygous alternate

These are used to calculate relatedness and a measure of relatedness that is unaffected by loss-of-heterozygosity that is often seen in some cancers. The interactive output allows toggling between any of these measures.

It also reports depth information and the count of HET, HOM_REF, HOM_ALT, and unknown genotypes for each sample along with a number of metrics that are useful for general QC.

Example

Here, each point is a pair of samples. We can see that the expected identical sample-pairs (e.g. tumor-normal pairs) specified by the user and drawn in red mostly cluster together on the right. Unrelateds cluster on the lower left. The sample-swaps are the blue points that cluster with the red. In the somalier output, the user can hover to see which sample-pairs are involved each point

Ancestry Estimate

somalier can predict ancestry on a set of query samples given a set of labelled samples. You can read more about this in the wiki

Usage

Usage is intentionally very simple and running somalier extract or somalier relate will give sufficient help for nearly all cases.

By default somalier will only consider variants that have a "PASS" or "RefCall" FILTER. To extend this list, set the environment variable SOMALIER_ALLOWED_FILTERS to a comma-delimited list of additional filters to allow.

by default sites with an allele balance < 0.01 will be considered homozygous reference. To adjust this, use e.g. : SOMALIER_AB_HOM_CUTOFF=0.04 somalier relate ...

Other Work

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5499645/

https://academic.oup.com/bioinformatics/article/33/4/596/2624551

Acknowledgement

This work was motivated by interaction and discussions with Preeti Aahir and several early users who provided valuable feedback.