Reference Panels - Usage Guide

Overview

Select and prepare the reference panel that statistical phasing and imputation copy haplotypes from. The decision that dominates results is ancestry match: imputation can only impute variants the panel contains and copies haplotypes from samples that resemble the target, so a panel whose ancestry does not match the study population imputes poorly no matter how large it is. This skill covers panel choice (1000 Genomes, HRC, TOPMed, HGDP+1kGP, CAAPA), genome-build and chromosome-naming reconciliation, the strand/allele harmonization gate that silently corrupts results when skipped, and conversion to the engine-specific panel format. It does not run the phasing or imputation engines (those are sibling skills) and routes classical HLA-allele panels out to clinical-databases.

Prerequisites

• bcftools (conda install -c bioconda bcftools) with the fixref and fill-tags plugins; PLINK for the allele-frequency file the harmonization check consumes.
• Will Rayner's HRC-1000G-check-bim.pl (or the bgen/VCF variant) for the field-standard strand/allele check, plus the panel's site/legend file.
• A reference FASTA matching the PANEL's genome build, and a QC'd, normalized study VCF.
• Conceptual prerequisites and big notes:
• A panel is data with a dated release and a fixed build (GRCh37 or GRCh38). Record the exact version and build; "1000 Genomes" alone is unreproducible.
• HRC is SNP-only (no indels) with a MAF floor near 5e-4; TOPMed is the most diverse but is server-only and never downloadable.
• Match the panel build to the study build and avoid liftover; if forced to lift, do it once with a strand-aware method and re-run the harmonization check.
• Normalize (split multiallelics, left-align) before aligning alleles to the panel.

Quick Start

Tell your AI agent what you want to do: - "Which reference panel should I use for my admixed Latino cohort on GRCh38?" - "Align my GRCh37 array data to the HRC panel strand and alleles before imputation" - "Convert my reference VCF to Beagle bref3 and Minimac msav formats" - "My study data is GRCh37 but I want a GRCh38 panel - how should I handle the build?" - "Check whether my data can use TOPMed given our data-residency rules"

Example Prompts

Panel selection

"My cohort is an admixed US population sequenced on GRCh38. Which reference panel should I impute against and why, and note any access or governance requirements."

Strand and build harmonization

"I have an Illumina array VCF on GRCh37. Walk me through aligning strand and alleles to the panel, handling palindromic SNPs, and reading the allele-frequency concordance plot before I upload to an imputation server."

Format conversion

"I have a custom reference panel as a phased VCF. Convert it to bref3 for Beagle and msav for Minimac4, per chromosome, and tell me what each engine requires."

Governance constraint

"Our IRB forbids participant genotypes leaving the country. Which diverse panels can I use locally, and what accuracy trade-off does that imply versus TOPMed?"

What the Agent Will Do

1. Establish the study genome build and the target ancestry (routing to population-genetics/population-structure for the PCA if needed).
2. Recommend a panel by ancestry match, build, variant class (SNP vs indel), and data-governance constraints, naming the exact version.
3. Normalize the study VCF and reconcile chromosome naming to the panel convention.
4. Run the strand/allele harmonization check, execute the generated fix script, drop unresolvable palindromic SNPs, and inspect the allele-frequency concordance plot against the ancestry-matched sub-panel.
5. Convert the panel to the engine-specific format (msav, bref3, or imp5), pairing it with a build-matched genetic map.
6. Hand the harmonized study data and prepared panel to haplotype-phasing and genotype-imputation.

Tips

• When someone says "I used the biggest panel," the real question is how many panel samples share the target ancestry, because that subset did the work.
• Treat "all palindromic SNPs kept" as evidence the strand was never checked; the check drops A/T and C/G SNPs above MAF 0.4 deliberately.
• Run the harmonization check and then actually execute its generated fix script; the check diagnoses, the script fixes.
• A high INFO/R2 from an ancestry-mismatched panel can be confidently wrong; the metric is panel-relative and cannot detect that the panel lacked the target's haplotypes.
• If the data cannot be uploaded to a US server, the practical choice narrows to downloadable panels (1000G, HGDP+1kGP); HGDP+1kGP is the strong diverse-and-local default.
• Obtain both the site-only legend (for the check) and the full haplotype panel (for imputation); they are different files.

Related Skills

• haplotype-phasing - The phasing engine that consumes the panel
• genotype-imputation - Impute untyped variants once the panel is prepared
• imputation-qc - INFO/R2 quality metrics, which cannot detect ancestry mismatch
• variant-calling/variant-normalization - Split multiallelics and left-align before harmonization
• population-genetics/population-structure - PCA to establish target ancestry
• clinical-databases/hla-typing - Classical HLA-allele imputation with a dedicated panel
• workflows/gwas-pipeline - End-to-end QC -> phase -> impute -> associate