What a clinical genome is, and how it differs from the VCF on your laptop

“Clinical genome” refers to two different things, and search results mix them. The first is a clinical genome test: whole-genome sequencing performed in a CLIA/CAP-accredited lab, analyzed against a defined gene list, and issued as a signed report with variants classified pathogenic, likely pathogenic, VUS, likely benign, or benign. The second is ClinGen, the Clinical Genome Resource. It is an NIH-funded consortium that publishes the curated knowledge those reports depend on. That knowledge covers which genes are validly linked to which diseases, which are dosage-sensitive, and expert-panel classifications of specific variants. If you have raw sequencing data and want to understand it, you need the second to make sense of the first. This page covers both, and where the line sits between analysis you can do yourself and interpretation that requires a clinician.

What separates a clinical genome from a research genome

The FASTQs look the same. The difference is in coverage guarantees, confirmation, and classification process.

A clinical lab commits to a minimum depth across a defined set of medically relevant genes, typically ≥20x at ≥98% of target bases, and reports the regions that fail. Research pipelines usually report a genome-wide mean (30x, 40x) and stop there, which hides the gaps. Coverage is not uniform. GC-rich first exons, segmental duplications and homopolymer runs drop out regardless of mean depth. The segmental duplications include PMS2, SMN1, CYP21A2 and STRC. One WGS clinical interpretation study found that 10% to 19% of inherited disease genes, depending on platform, had coverage insufficient to call pathogenic variants reliably. It also found that concordance between platforms was high for SNVs but poor for small insertions and deletions ¹. That paper is a decade old and callers have improved, but the shape of the problem has not changed.

A clinical lab also orthogonally confirms reportable variants, usually Sanger or a targeted assay, before it puts them in a report. And it classifies variants using the ACMG/AMP framework with documented evidence codes rather than whatever ClinVar happens to say today.

WGS as a first-line clinical test has held up well in practice, with diagnostic yields that exceed targeted panels and the advantage of a single assay covering SNVs, indels, structural variants, and mitochondrial DNA ². Early clinic deployments showed it works in ordinary medical genetics workflows. Its use is not confined to research settings ³. The limiting factor is interpretation labor. It is not sequencing.

The files, and what to do with each

A complete delivery should include:

• sample.cram plus .crai, aligned to GRCh38 (ideally with ALT contigs handled properly, or aligned to a graph reference). CRAM is about 60% the size of BAM at the same information content when the reference is available.
• sample.hard-filtered.vcf.gz (SNVs/indels) plus a gVCF if you want to distinguish “reference” from “no coverage.” The gVCF matters: a position absent from a plain VCF could be homozygous reference or unsequenced, and only the gVCF tells you which.
• sample.sv.vcf.gz from Manta/DRAGEN/GRIDSS, and CNV calls, which are a separate call set with separate error modes.
• sample.repeats.vcf from ExpansionHunter if repeat expansions were genotyped.
• QC: coverage distribution and contamination estimate (VerifyBamID FREEMIX, want < 0.02). Also the ratio of heterozygous to homozygous alternate calls, and the sex check.

First thing we do with a new VCF is normalize it, because everything downstream depends on consistent representation:

bcftools norm -m -any -f GRCh38.fa -Oz -o norm.vcf.gz sample.vcf.gz
bcftools index -t norm.vcf.gz

Then annotate. Ensembl VEP with the cache is the workhorse:

vep -i norm.vcf.gz --cache --assembly GRCh38 --offline \
    --everything --pick_allele_gene --vcf --compress_output bgzip \
    --plugin dbNSFP,/data/dbNSFP4.7a_grch38.gz,REVEL_score,CADD_phred,SpliceAI_pred \
    -o annotated.vcf.gz

--pick_allele_gene is opinionated: it collapses transcripts to one consequence per gene, which makes filtering sane and occasionally throws away the transcript that mattered. For any variant you take seriously, go back and look at all transcripts, and check whether the MANE Select transcript is the one being reported.

Then intersect with ClinVar and check coverage of the genes you care about:

bcftools annotate -a clinvar_20260101.vcf.gz \
  -c INFO/CLNSIG,INFO/CLNREVSTAT,INFO/CLNDN -Oz -o clinvar.vcf.gz annotated.vcf.gz

mosdepth --by acmg_sf_v3.2.bed --thresholds 10,20,30 sample sample.cram

The CLNREVSTAT field is the one to filter on, not CLNSIG. ClinVar assertions carry a review status from zero to four stars. A single submitter with no assertion criteria (one star or fewer) is a hypothesis. Three stars means a ClinGen Variant Curation Expert Panel reviewed it. Four means a practice guideline. We filter to reviewed_by_expert_panel or practice_guideline for anything we would act on, and treat one-star pathogenic calls as leads to investigate rather than findings.

How to use ClinGen specifically

ClinGen publishes machine-readable curations you can pull into a pipeline:

• Gene-Disease Validity: each gene-disease pair classified Definitive, Strong or Moderate, or else Limited, Disputed, or Refuted. Downloadable as CSV. If a gene’s link to the phenotype is Limited, a “pathogenic” variant in it carries much less weight than the label suggests.
• Dosage Sensitivity: haploinsufficiency and triplosensitivity scores per gene and region. This is what you use to interpret a CNV call, alongside the ACMG/ClinGen CNV scoring rubric.
• Variant Curation Expert Panels: gene-specific rules that modify the generic ACMG/AMP criteria. The BRCA1/2 ENIGMA specifications, the CDH1 panel, the hearing loss panel, and others specify exactly what frequency threshold triggers BA1 or BS1 for that gene, which is usually not the generic 5%.
• The ClinGen Allele Registry gives canonical allele IDs (CA identifiers) so you can join across databases without fighting over HGVS formatting.

Add gnomAD v4 allele frequencies and filter hard. A variant claimed pathogenic for a dominant disease with population frequency above a few times the disease prevalence is almost always misclassified. This single filter removes most of the false positives in a naive ClinVar intersection.

For phenotype-driven ranking, Exomiser with HPO terms works well and is scriptable:

java -jar exomiser-cli.jar --analysis analysis.yml \
     --vcf clinvar.vcf.gz --assembly hg38

What this pipeline will and will not find

It will find well-characterized coding SNVs and indels in genes with good coverage. That includes the ACMG secondary findings list, and pharmacogenomic star alleles if you run PharmCAT or Aldy on top. The list is about 80 genes. It covers hereditary cancer syndromes, cardiomyopathies and arrhythmias, familial hypercholesterolemia and malignant hyperthermia.

It will underperform on the cases below.

• PMS2 and other genes with pseudogenes
• repeat expansions unless you specifically genotyped them
• deep intronic splice variants (SpliceAI helps, prediction is not proof)
• somatic mosaicism below the variant-caller’s allele-fraction threshold
• CNVs smaller than a few kilobases from short-read data

And it will produce variants of uncertain significance, many of them. A VUS is a statement about the state of evidence. It is not a statement about you. The correct action for a VUS is almost always to record it, set a reminder to recheck ClinVar in a year, and do nothing else.

Where you need a clinician

Anything you find in a research-grade call set should be treated as a hypothesis until a clinical lab confirms it on a fresh sample. Sample swaps happen. Aligner artifacts happen. A pathogenic variant in a cardiomyopathy gene or a cancer predisposition gene changes screening schedules and has implications for your relatives, and that conversation belongs with a board-certified genetic counselor or clinical geneticist who can order the confirmatory test, take a family history, and put penetrance in context. Penetrance estimates from disease cohorts are systematically higher than what is observed in unselected populations, which is exactly the situation you are in if you sequenced yourself while healthy. We do not interpret findings for medical action here, and neither should a pipeline you wrote on a weekend.

Questions people also ask

Is a geneticist an MD or PhD? Both exist and they do different jobs. A clinical geneticist is an MD who sees patients and signs reports. A clinical molecular geneticist or clinical cytogeneticist is typically a PhD certified by the American Board of Medical Genetics and Genomics who directs the laboratory and signs out variant classifications. Genetic counselors hold a master’s degree and are usually the person who explains results and coordinates family testing.

What does genomic testing tell you? Which variants you carry, and where the current evidence places them: pathogenic, benign, or uncertain. It covers monogenic disease risk, carrier status for recessive conditions, pharmacogenomic metabolizer status, and polygenic scores of limited and population-dependent accuracy. It does not tell you what will happen to you.

When is genomic testing usually recommended? For an undiagnosed condition with a suspected genetic basis, a family history clustering in a suggestive pattern, reproductive carrier screening, or characterizing a tumor. WGS is increasingly the first-line test in these settings rather than a sequential panel-then-exome workup ², and it has been shown to fit into routine genetics clinic operations ³.

What are the benefits? For a diagnostic case, ending the search and identifying management pathways. For a healthy person, the realistic benefits are actionable secondary findings in a small number of genes, carrier status relevant to reproductive planning, pharmacogenomics, and a permanent data asset you can reanalyze as ClinGen and ClinVar improve. Reanalysis is underrated: the same file interpreted three years later yields new answers without new sequencing.

Footnotes

1. Frederick E. Dewey, Megan E. Grove, Cuiping Pan, et al. Clinical Interpretation and Implications of Whole-Genome Sequencing. JAMA, 2014. https://doi.org/10.1001/jama.2014.1717 ↩

2. Petar Brlek, Luka Bulić, Matea Bračić, et al. Implementing Whole Genome Sequencing (WGS) in Clinical Practice: Advantages, Challenges, and Future Perspectives. Cells, 2024. https://doi.org/10.3390/cells13060504 ↩ ↩²

3. Pamela Fraser, Michael F. Gutzeit, Jeremy M. Harris, et al. Successful Application of Whole Genome Sequencing in a Medical Genetics Clinic. Journal of Pediatric Genetics, 2016. https://doi.org/10.1055/s-0036-1593968 ↩ ↩²