Bio Phasing Imputation Genotype Imputation

Version Compatibility

Reference examples tested with: Beagle 5.4 (22Jul22), Minimac4 4.1+, IMPUTE5 1.2, GLIMPSE2, bcftools 1.19+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Minimac4 (4.x) uses POSITIONAL arguments (minimac4 panel.msav target.vcf.gz); the old --refHaps/--haps/--prefix/--cpus style is Minimac3 and obsolete. Beagle 5.x emits DR2, AF, and IMP only (AR2 is a legacy 4.x field) and its default ne=100000 (not 1,000,000). The panel build (GRCh37 vs GRCh38) must match the data; record the panel name, version, and build with every result.

Genotype Imputation -- Inferring Untyped Genotypes as Dosages

"Fill in the variants I did not directly measure" -> Align the (phased or low-coverage) sample to a reference panel of phased haplotypes and infer the untyped alleles via the Li-Stephens HMM - because the output is a posterior over genotypes summarized as a dosage with a self-estimated quality, not a measured call, so the uncertainty must be carried downstream.

• CLI: java -jar beagle.jar gt=phased.vcf.gz ref=panel.bref3 map=plink.chr20.map out=imputed (or minimac4 panel.msav phased.vcf.gz, or GLIMPSE2 for low-coverage WGS)

Scope: imputing untyped genotypes from a panel (array data) or from genotype likelihoods (low-coverage WGS), the dosage/quality output, chunking, chrX, and the servers. Phasing the input -> haplotype-phasing. Panel selection/preparation/strand -> reference-panels. Quality metrics and filtering thresholds -> imputation-qc. The GWAS test on the dosages -> population-genetics/association-testing. The genotype likelihoods that low-coverage imputation consumes -> variant-calling/vcf-basics. End-to-end orchestration -> workflows/gwas-pipeline.

The Single Most Important Modern Insight -- An Imputed Genotype Is a Posterior, and the Deliverable Is a Dosage Plus a Self-Estimated Quality, Not a Hard Call

Imputation aligns a sparsely-genotyped (or low-coverage-sequenced) sample to a densely-typed reference panel of phased haplotypes and infers, via a Li-Stephens HMM, the alleles at positions the sample never observed (Browning 2018 Am J Hum Genet 103:338). The output at each untyped variant is a distribution, summarized as an expected allelic dosage in [0,2]. Three facts define the field:

1. Downstream analysis uses dosages, not hard genotypes. The dosage DS is the conditional expectation E[genotype | data, panel], the minimum-variance summary; hard-calling forces an uncertain 0.5 dosage to 0 or 1, injecting genotype error that attenuates effects and inflates standard errors. GWAS regresses the trait on DS -> population-genetics/association-testing.
2. The quality metric (Beagle DR2, Minimac R2, IMPUTE INFO) is an ESTIMATE of r2 from the posterior spread, computed without ever seeing the truth. Poorly-imputed dosages shrink toward the allele-frequency mean 2p, so low posterior variance relative to the binomial expectation 2p(1-p) flags a low-confidence site. This is NOT a validation against held-out genotypes (that is empirical r2 / EmpRsq, a masked-site quantity). Say "DR2/R2/INFO is an estimate of imputation quality," never "the imputation accuracy was 0.9" as if measured. The metric also cannot detect panel-ancestry mismatch -> imputation-qc.
3. Low-coverage WGS (0.5-4x) plus GLIMPSE2 has become a credible array replacement. Because it samples the whole genome rather than a fixed ascertained SNP set, it imputes rare variants and under-represented ancestries better than a dense array at comparable cost (Rubinacci 2023 Nat Genet 55:1088). The input is genotype likelihoods, not calls; the array-vs-low-coverage-WGS choice is an ascertainment decision (see Array vs Low-Coverage WGS below).

Tool Taxonomy

Decision Tree by Scenario

Array vs Low-Coverage WGS: the Imputation-Input Fork

The upstream decision is how to generate the genotypes that will be imputed, and it is an ascertainment question, not just an accuracy one. An array assays a fixed, designed SNP set (biased to its design population); low-coverage WGS samples whatever is in the genome.

The judgment: common-variant GWAS in a well-paneled ancestry -> array plus imputation is cheap and adequate; rare variants, under-represented ancestry, or a need for unbiased genome-wide ascertainment -> low-coverage WGS plus genotype-likelihood imputation, the direction the field is moving as sequencing costs fall. Low-coverage WGS is only as good as its panel and its likelihoods (bad mapping, contamination, or damage produce garbage GLs that impute garbage).

Output Formats and Why Dosages

The central object is the posterior genotype distribution; everything else summarizes it. Request the fields up front (Minimac4 -f GT,DS,HDS,GP; Beagle gp=true ap=true).

GP is the distribution; DS is its mean - two variants with different GP spreads can share a DS. Use DS for association (it propagates the uncertainty); use HDS/AP for phased/allele-specific analyses. Beagle computes GP from allele probabilities assuming Hardy-Weinberg and sets GT from the per-haplotype argmax, so its GT can occasionally disagree with the argmax of its own GP.

The Phasing Prerequisite

The reference panel is phased haplotypes; the target must align to that haplotype structure two ways:

• Pre-phase then impute (Minimac4, IMPUTE5): phase the target FIRST (Eagle2 or SHAPEIT) into haplotypes, then impute. The server default (Eagle2 -> Minimac4) and the fast local pattern -> haplotype-phasing.
• Phase-and-impute together (Beagle, GLIMPSE2): the tool phases internally. Low-coverage tools MUST do this, because there is no confident genotype to phase up front; GLIMPSE2 alternates haploid imputation and phasing, and gains accuracy by imputing all target samples jointly.

Low-Coverage WGS Workflow (GLIMPSE2)

The input is genotype likelihoods (PL/GL), not calls, because at 0.5-4x no genotype is certain. GLIMPSE2 can read BAM/CRAM directly (computing GLs internally) or a GL BCF made with bcftools mpileup ... -T panel_sites.vcf.gz | bcftools call -Aim -C alleles -T panel_sites.tsv.gz (the -C alleles constraint needs the panel sites supplied to call via -T; note the two -T files differ in format - a VCF for mpileup, a tab-delimited sites file for call). The pipeline:

1. GLIMPSE2_chunk defines windows with buffers.
2. GLIMPSE2_split_reference precomputes a binary panel per chunk (the speed innovation that made UK Biobank-scale imputation feasible).
3. GLIMPSE2_phase imputes and phases per chunk (--bam-list or --input-gl; --ne default 100000).
4. GLIMPSE2_ligate stitches chunks using the overlap buffers to keep phase. Output FORMAT: GT, DS, GP, HS plus a per-variant INFO score.

For chrX with GLIMPSE2, declare each sample's ploidy with --samples-file (sample and copy number) and run the PAR/nonPAR split as for the array tools (male nonPAR is haploid) -> reference-panels.