Bio Phasing Imputation Reference Panels

Version Compatibility

Reference examples tested with: bcftools 1.19+, PLINK 1.9+.

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.

A reference panel is DATA, not a tool: it has a dated release and a fixed genome build (GRCh37 or GRCh38). Record the exact panel name, version, and build with every result; "1000 Genomes" without a version is unreproducible because Phase 3 (GRCh37, low-coverage) and the high-coverage NYGC 3202 release (GRCh38, 30x) are different call sets. HRC and TOPMed are server-only (not downloadable); panel-build tools (Minimac4 --compress-reference, bref3.jar, imp5Converter) and the Will Rayner harmonization check are separate downloads.

Reference Panels -- Choosing and Preparing the Prior Imputation Copies From

"Which reference panel should I use, and how do I prepare it?" -> Match the panel's ancestry to the target population, reconcile build and strand, then convert to the engine's format - because imputation can only copy haplotypes the panel contains, so the panel IS the prior, and a mismatched ancestry or a flipped strand corrupts the result without any error.

• CLI: bcftools norm -m -any -f ref.fa then the strand/allele harmonization check against the panel sites, then minimac4 --compress-reference / bref3.jar / imp5Converter to build the engine format

Scope: panel selection (ancestry-match, build, access), strand/allele harmonization, build/liftover, and format conversion. The phasing engine that consumes the panel -> haplotype-phasing. Imputation -> genotype-imputation. PCA to establish the target ancestry -> population-genetics/population-structure. Classical HLA-allele imputation needs a dedicated HLA panel -> clinical-databases/hla-typing. VCF normalization mechanics -> variant-calling/variant-normalization.

The Single Most Important Modern Insight -- Ancestry Match Beats Panel Size, Because Imputation Copies Haplotypes and a Mismatched Panel Has None Worth Copying

The reflex "TOPMed has 97k samples, HRC has 32k, so use TOPMed" is right for a European cohort and wrong for an ancestry-mismatched one, and the reason is mechanical, not statistical: imputation copies haplotype segments from panel samples that resemble the target, so if no panel sample carries the target population's haplotypes there is nothing to copy, and adding ten thousand more European haplotypes does nothing for an East African sample (Marchini & Howie 2010 Nat Rev Genet 11:499). The binding resource is not panel size but how many panel samples share the target's ancestry. Three facts follow:

1. Ancestry composition often dominates size. HRC (32k, European-heavy) imputes African-ancestry rare variants poorly while TOPMed lifts them dramatically - not because TOPMed is 3x bigger but because it contains African-American and Hispanic/Latino haplotypes. Admixed samples need a panel with both ancestral components and admixed individuals, because admixed haplotypes are mosaics a single-ancestry panel cannot reconstruct at the switch points. The honest answer is "large AND matched"; where both are not available, which one wins depends on whether the target is common or rare variants (size matters more for rare).
2. The self-graded INFO/R2 cannot see ancestry mismatch. When the panel lacks the target's haplotypes the model still finds some template and reports a confident-looking quality about a copy that is systematically wrong (full metric theory -> imputation-qc). High R2 in an under-represented ancestry is not reassurance; validate against masked truth stratified by frequency.
3. A panel encodes whose haplotypes are trusted to fill the gaps. HRC is European-heavy because its component cohorts were; TOPMed is diverse because it was designed to be. No panel represents everyone, and the imputation inherits exactly the panel population's representation, including its gaps. This is the central equity problem of imputation, and low-coverage WGS (unbiased ascertainment -> genotype-imputation) is the main route around it.

The Major Panels

Numbers are routinely misquoted; these are the verified figures. State the exact version and build in any method.

The "1000G" trap: Phase 3 (Auton 2015, GRCh37, low-coverage) and the high-coverage NYGC 3202 release (Byrska-Bishop 2022, GRCh38, 30x, with 602 trios) are different call sets; the NYGC release is strictly better for rare variants. State which.

Decision Tree by Scenario

The governing principle: ancestry match beats panel size, and governance (can the data be uploaded to a US server?) often narrows the field before accuracy does.

The Strand / Allele Harmonization Gate

This is the step that silently corrupts results when skipped. The job: align every study variant's alleles to the panel REF/ALT, fix strand, and drop the variants that cannot be safely resolved. Normalize first (bcftools norm -m -any -f ref.fa), because the same indel represented two ways will not match.

The field-standard gate is Will Rayner's check (HRC-1000G-check-bim.pl and its bgen/VCF variants): it compares a QC'd PLINK .bim plus an allele-frequency file against the panel's sites list and EMITS a Run-plink.sh that updates positions, ref/alt, and strand, removes unresolvable SNPs, and splits by chromosome. The check diagnoses; the script fixes - both are required, and a surprising number of pipelines run the check and never execute the script.

The allele-frequency concordance plot (study AF vs panel AF) is the visual gate, not decoration: a tight diagonal is good; points on the y = 1 - x anti-diagonal are strand flips; a general smear is sample mislabeling or the wrong panel ancestry. Compare against the ancestry-MATCHED sub-panel's frequencies - an African cohort vs a European panel's AF smears even with perfect strand. Read this plot before uploading, every time.

Panel Formats and the Genetic-Map Pairing

A panel is half the input; phasing/imputation also needs a genetic (recombination) map matched to the panel build. A panel VCF comes site-only (the legend, used for the harmonization check) or full (the haplotypes, used to impute) - obtain both. The map is build-specific: an hg19 map with a GRCh38 panel silently mis-places recombination rates.

chrX and MHC

• chrX: split into PAR1 / nonPAR / PAR2 using build-correct coordinates. PAR and female nonPAR are diploid; male nonPAR is HAPLOID and must be coded haploid (a het call there is an error). Mixed ploidy in one file crashes most tools; the Michigan/TOPMed servers split, impute, and re-merge automatically.
• MHC (chr6 ~28-34 Mb): extreme LD and polymorphism. Standard panels impute SNPs there but unreliably and cannot impute classical HLA alleles at all - that needs a dedicated HLA panel and tool -> clinical-databases/hla-typing.

Per-Method Failure Modes