Skip to content

Step 14: Imputation Preparation

What This Does

Prepares a WGS VCF for submission to the Michigan Imputation Server (MIS) or TOPMed Imputation Server. Splits the VCF by chromosome, filters to PASS variants, and converts to the required format.

Why

Imputation servers statistically infer missing genotypes using large reference panels. For WGS data, imputation is primarily useful for phasing (determining which alleles are on the same chromosome) rather than filling in missing variants. Phased data is required for haplotype-level analyses and accurate PRS calculation.

Tool

  • bcftools (samtools/bcftools) — for VCF filtering, splitting, and indexing

Docker Image

  • BCFTOOLS_IMAGE

Pinned in versions.env; Image versions lists the current tag.

Command

export GENOME_DIR=/path/to/your/data
./scripts/14-imputation-prep.sh your_sample

The script starts one container and makes one bcftools view pass per chromosome, chr1-22 and chrX. Each file keeps the PASS records (and records with no filter), is written with its index under a temporary name, and is renamed when both are complete: the old index is removed first and the new one is moved in last, so an index never sits beside a VCF it was not built from. A chromosome the VCF has no record on gets no file (a file an earlier run left for it is removed), and the log names it. The commands it runs:

source versions.env   # from the repository root
SAMPLE=your_sample
GENOME_DIR=/path/to/your/data
mkdir -p ${GENOME_DIR}/${SAMPLE}/imputation/mis_ready

docker run --rm \
  -v ${GENOME_DIR}/${SAMPLE}:/data \
  "${BCFTOOLS_IMAGE}" \
  bash -c 'for chr in $(seq -f "chr%g" 1 22) chrX; do
    bcftools view -f PASS,. -r "$chr" -Oz --write-index=tbi \
      -o /data/imputation/mis_ready/'"${SAMPLE}"'_${chr}.vcf.gz /data/vcf/'"${SAMPLE}"'.vcf.gz
  done'

# Output: 23 per-chromosome VCFs, each with its .tbi, in ${SAMPLE}/imputation/mis_ready/

By default the input is variant-only. The step reads the VCF from step 3, which lists only the sites where the sample differs from the reference. A site that is missing from it is not a confirmed homozygous-reference genotype, and an imputation server treats it as missing.

With panel sites, hom-ref genotypes come from the gVCF. Step 3 also writes a gVCF (vcf/${SAMPLE}.g.vcf.gz), which records where the sample matches the reference. Give the step the reference panel's sites and it genotypes each of them from the gVCF: a variant call, a 0/0 call where the sample matches the reference, and nothing where the site was not covered. The per-chromosome files then hold those sites only.

The 0/0 calls matter when you merge this sample's files with other samples before the upload: there they tell a 0/0 genotype from a missing one. In a single-sample upload they add nothing, because the Michigan server's input QC excludes monomorphic sites (a 0/0 or 1/1 site in the only sample is one) and alleles other than A, C, G and T (a 0/0 call has the ALT .).

# CHROM<TAB>POS per line, or a VCF of the panel's sites, inside GENOME_DIR
IMPUTATION_SITES=${GENOME_DIR}/reference/panel_sites.tsv ./scripts/14-imputation-prep.sh your_sample

A panel of tens of millions of sites needs a few GB of memory for this. Without IMPUTATION_SITES and with a gVCF present, the step prints how to use it.

Server Options

Server Panel Samples Build URL
Michigan (MIS) HRC r1.1 32,470 GRCh37/38 imputationserver.sph.umich.edu
TOPMed TOPMed r2 132,070 GRCh38 native imputation.biodatacatalyst.nhlbi.nih.gov

Your data leaves the machine here

This is the only step whose output is meant to be sent somewhere else. The other steps keep your genome on your disk; an imputation server receives your genotypes for every chromosome you upload. Uploading is a separate decision you make, not something the step does.

Before you upload, read the server's data policy. The Michigan server's security page says it deletes the input once it is no longer needed, keeps only the number of samples and markers, and encrypts the results with a one-time password. Its getting started guide says the results are deleted 7 days after the job ends. Neither page gives an exact time for deleting inputs or backups, or says whether data is encrypted at rest. Check the TOPMed server's own terms; they are not the same service. Both need an account, so the upload is tied to your email address.

Important Notes

  • MIS requires a minimum of 20 samples per job — a single WGS sample is useful mainly for phasing, not imputation
  • TOPMed r2 panel is recommended for European ancestry (132K samples, GRCh38 native — no liftover needed)
  • Registration is required at the imputation server before submitting jobs
  • Upload per-chromosome VCF files (not the whole-genome file)
  • Servers accept .vcf.gz format — ensure files are bgzipped (bcftools output is bgzipped by default)
  • chrX is prepared too; servers usually take it as a separate job with ploidy-aware settings
  • Results include phased haplotypes and imputation quality scores (R-squared) — filter imputed variants with R2 < 0.3