genomic-variants

Genomic Variant Analysis

Guidelines for working with called variants: quality filtering, annotation against catalogued resources, genome-wide association, and population-genetics summaries.

Where This Pack Starts

Your input is a VCF/BCF, a PLINK-format cohort, or an aligned BAM/CRAM you read evidence from. Variant calling is upstream and out of scope — no caller is installed and none can be. If you are handed reads, say what called form you need and stop.

Knowing what produced the input still matters for reading it. Germline call sets usually come from GATK HaplotypeCaller, DeepVariant or FreeBayes; somatic sets from Mutect2; structural variants from Manta, DELLY or Sniffles2; copy number from CNVkit or FACETS. Use those names to interpret the FILTER values, INFO fields and headers you are given — not as steps to run.

Method-Selection Decision Tree

1. Variant filtering

Called VCF
  → Programmatic pass → cyvcf2 (Python), full control over per-record logic
  → Expression pass over a large file → bcftools view/filter
  → Rare-variant subset → allele frequency from an annotation resource,
    or cohort AF computed from AC/AN when none resolves

• Hard-filter thresholds for a supplied germline call set, when the INFO fields are present: QD > 2.0, FS < 60.0, MQ > 40.0, MQRankSum > -12.5, ReadPosRankSum > -8.0. Check the distributions before applying them — these are starting points, not universal cutoffs, and a caller that did not emit a field cannot be filtered on it.
• Report the FILTER field as the caller populated it. Never present an unfiltered set as filtered.

2. GWAS

Genotype data (VCF or PLINK format)
  → QC: missingness (<0.02), HWE (p > 1e-6), MAF (>0.01)
    → Population structure: PCA (top 10-20 PCs as covariates)
      → Relatedness: PLINK2 --king-cutoff (remove or model related pairs)
        → Association: PLINK2 --glm (linear/logistic)

• Verify plink2 is on PATH before building a pipeline around it — it is installed on x86_64 only. There is no linux-aarch64 build, so on an arm64 host it is absent by design, not by misconfiguration, and no-egress means it cannot be installed at runtime. Probe (command -v plink2), and if it is missing, report the reason and fall back: bcftools/cyvcf2 for filtering, allele frequencies and missingness, statsmodels for per-variant association on a cyvcf2-extracted dosage matrix, scikit-learn PCA for population structure. references/plink2-cli.md maps each step to its substitute.
• ALWAYS include PCA covariates to adjust for population structure. Omitting this produces inflated results.
• Check genomic inflation factor (lambda_GC); values >1.05 suggest residual confounding.