From FASTQ to a clinical answer

FASTQ → BAM → VCF, without the silent failures

A modern germline pipeline is four moving parts. Each one fails in characteristic ways that look like real biology if you don't know to check.

1. Align
   BWA-MEM2, minimap2 (long read), DRAGEN-Map

   Reads are mapped to a reference (GRCh38, T2T-CHM13). Reference choice matters: GRCh38 mis-maps ~5M reads per genome that T2T resolves correctly.

   Silent failure · Decoy contigs disabled → false SVs in HLA, KIR, centromeres.

2. Pre-process
   MarkDuplicates, BQSR, soft-clipping

   Optical/PCR duplicates removed; base qualities recalibrated against known sites. Skip BQSR for PCR-free WGS — it's a no-op there.

   Silent failure · Failing to mark duplicates inflates allele fractions in tumor-only calling.

3. Call
   DeepVariant, GATK HaplotypeCaller, Strelka2; Clair3 / DeepVariant for ONT/HiFi; Mutect2 for somatic

   Joint-call families/cohorts when possible. DeepVariant currently leads on F1 for short-read SNVs; PEPPER-Margin-DeepVariant for nanopore.

   Silent failure · Single-sample calling on low-coverage WGS produces 30–50% false het calls in low-complexity regions.

4. Annotate & filter
   VEP, SnpEff, Ensembl MANE select, gnomAD v4, ClinVar, SpliceAI

   Pick one canonical transcript set (MANE) and stick to it. Filter on gnomAD popmax AF, not global AF.

   Silent failure · Reporting a variant against a non-MANE transcript is the #1 cause of clinician confusion at sign-out.

ACMG/AMP, and why most VUS aren't really VUS

Richards et al. 2015 (with the 2018–2024 ClinGen Sequence Variant Interpretation refinements) is the law. Five tiers — Pathogenic, Likely Pathogenic, VUS, Likely Benign, Benign — built from weighted evidence codes.

The reference genome is mostly European

~78% of GWAS participants and ~55% of gnomAD v4 are of European ancestry. Pretending this doesn't matter is how you misclassify variants and miscalibrate polygenic risk in non-European patients.

• gnomAD popmax, not global AF

  A variant at 0.1% globally can be 4% in African or East Asian subpopulations — well above the BA1 5% line in some VCEPs. Always filter on the highest population AF, with the appropriate confidence interval.

• PRS portability collapses across ancestries

  Polygenic scores trained on UK Biobank lose 60–80% of their predictive variance when applied to African-ancestry individuals. Report PRS only when the training cohort matches the patient, or use ancestry-aware methods (PRS-CSx, BridgePRS).

• ACMG PM2 is ancestry-blind by default

  'Absent from controls' should be 'absent from a control set that includes the patient's ancestry'. Otherwise rare African variants get over-classified as pathogenic.

• GRCh38 itself is ancestry-skewed

  The pangenome reference (HPRC v1.1, 47 haplotypes) reduces small-variant errors by ~34% and SV errors by ~104% in non-European samples. Move to graph-based alignment (vg, Giraffe, DRAGEN graph) for diverse cohorts.

• Pharmacogenomic star alleles vary widely

  CYP2D6 *17, *29 are common in African ancestry and absent from many commercial panels. CYP2C19 *2/*3 frequencies differ 3–5× across populations. Always check PharmGKB tier of evidence and population frequency before dosing.

Before you sign the report

A pre-flight you can run in under five minutes per case. Each item links into the Evidence Grader so you can see how strong the underlying claim actually is.

• Coverage ≥20× across all reportable exons (≥30× for somatic)
  Grade A
• Variant called by ≥2 independent callers OR confirmed by orthogonal method (Sanger / ddPCR)
  Grade A
• Annotated against MANE Select transcript with HGVS c. and p. nomenclature
  Grade A
• ClinVar classification reviewed AND star rating ≥2 OR independent ACMG re-curation
  Grade B
• gnomAD v4 popmax AF checked against patient's reported ancestry
  Grade B
• If PRS reported: training cohort ancestry matches patient, or ancestry-adjusted method used
  Grade C
• Pharmacogenomic calls cross-checked against PharmGKB / CPIC level A guidelines
  Grade B
• Incidental findings handled per ACMG SF v3.2 secondary-finding list with documented patient consent
  Grade A
• If long-read used: SVs validated against ≥1 orthogonal method or replicated across two flow cells
  Grade B
• Direct-to-consumer raw data: never use for clinical decisions without a CLIA-validated re-test
  Grade F

The boxes that produce the reads

Reference catalog of clinical and translational sequencers. Throughput, accuracy, real per-genome cost, and what each instrument is actually good for.