The Best Genetic Test for Health Is 30x Whole-Genome Sequencing With Raw Data Access

Oak
genomics wgs variant-interpretation bioinformatics pharmacogenomics
Studio-lit feathered dinosaur specimen on a steel plinth, densely plumaged except for a band of sparse, widely spaced quills on its flank.

If you want genetic data you can work with for health purposes, buy 30x (or deeper) short-read whole-genome sequencing from a lab that returns raw FASTQ or CRAM plus a gVCF, and treat every clinically consequential finding as a hypothesis until a CLIA/CAP lab confirms it on a second sample. Genotyping arrays (23andMe, AncestryDNA, most “$99 health DNA kits”) are the wrong instrument for this: they interrogate roughly 600,000 to 700,000 pre-chosen positions out of 3.1 billion, they call rare variants poorly, and the “raw data” you export is a tab-separated file of probe calls with no read evidence behind it. You cannot audit a probe. You can audit a BAM.

The four tiers, and what each one measures

Genotyping array. Hybridization probes at fixed sites. Good for common-variant work: ancestry, polygenic scores, common pharmacogenomic star alleles. Rare pathogenic variants are the failure mode. Clustering algorithms are trained on common allele frequencies, so a heterozygous call at a site where the rare allele appears in 1 in 10,000 people is frequently a cluster-boundary artifact. Confirmatory sequencing of array-derived “pathogenic” calls in clinical labs has repeatedly reclassified large fractions of them as reference. If an array report tells you something alarming, the correct next step is a clinical lab, not a forum.

Exome (WES). ~1-2% of the genome, coding only, typically 100x mean depth. Cheaper per coding base, but capture is uneven, deep intronic and regulatory variants are invisible, and copy-number calling from capture data is noisy. Exome made sense when WGS cost $5,000. It mostly doesn’t now.

Short-read WGS at 30x. The default. Mean depth 30x gives ~95% of the genome at ≥10x callable depth, roughly 4-5 million SNVs and indels versus GRCh38, plus structural variant and copy-number calling that exomes cannot do. Deliverables run ~60-100 GB as gzipped FASTQ, ~15-25 GB as CRAM with a reference.

Long-read WGS (PacBio HiFi, ONT). Solves the regions short reads cannot: repeat expansions, segmental duplications, phasing across tens of kilobases, and direct methylation calling on ONT. More expensive, lower per-base accuracy on ONT, and the interpretation ecosystem is thinner. We would buy short-read 30x first and add long reads later if a specific question demands it.

What to demand from the provider before you pay

  1. Raw FASTQ (both reads, with intact read groups) or CRAM aligned to GRCh38 with the reference build and its accession named explicitly. “PDF report only” is disqualifying.
  2. A gVCF, not just a filtered VCF. The gVCF distinguishes “reference” from “no coverage here,” which matters when you are checking whether a gene was callable at all.
  3. Mean depth, percent of genome ≥10x and ≥20x, duplicate rate, and estimated contamination. Ask for the numbers, not a pass/fail badge.
  4. PCR-free library prep. PCR amplification inflates duplicates and degrades indel calling in homopolymers.
  5. CLIA/CAP accreditation if you want any finding to be usable by a clinician, or a stated path to orthogonal confirmation. Clinical WGS reporting has published best practices around confirmation, classification, and what belongs in a report, and a research-grade FASTQ does not meet them 1.

The QC you run yourself, in order

Before any interpretation, validate the data:

mosdepth --by 1000 --fast-mode sample sample.cram      # depth distribution
samtools flagstat sample.cram                          # mapping + duplicate rate
verifybamid2 --SVDPrefix 1000g.phase3 --Reference GRCh38.fa --BamFile sample.cram
somalier extract -d extracted/ --sites sites.GRCh38.vcf.gz -f GRCh38.fa sample.cram
somalier relate extracted/*.somalier                   # sex check, sample identity

Thresholds we use: contamination (FREEMIX) under 0.02, duplicate rate under 10% for PCR-free, mean depth ≥28x, and ≥90% of the genome at ≥20x. Sex check mismatching your records means a sample swap, and a sample swap is the single most consequential error in this whole pipeline. Check it first.

Then call variants. DeepVariant (--model_type=WGS) outperforms GATK HaplotypeCaller on indels in most benchmarks and needs no hand-tuned filtering. Normalize before annotating, because half of all ClinVar match failures are representation problems:

bcftools norm -m -any -f GRCh38.fa -Oz -o norm.vcf.gz calls.vcf.gz
bcftools index -t norm.vcf.gz
vep -i norm.vcf.gz --cache --offline --assembly GRCh38 --everything \
    --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT \
    --custom gnomad.v4.genomes.sites.vcf.gz,gnomAD,vcf,exact,0,AF \
    --vcf -o annotated.vcf

Short reads miss specific, known things. Add targeted callers rather than pretending the gaps do not exist: ExpansionHunter for repeat expansions, Cyrius for CYP2D6 (which short-read pipelines routinely miscall because of the CYP2D7 pseudogene), HLA-LA or T1K for HLA typing, Manta plus smoove for structural variants, and PharmCAT for star-allele pharmacogenomics from your VCF. PMS2, SMN1/SMN2, and the segmental duplications around them need either long reads or a dedicated assay.

What a genome can and cannot tell you

It can tell you, with high confidence, whether you carry a known pathogenic or likely pathogenic variant in a well-characterized gene. The ACMG secondary-findings list (v3.2, 81 genes) is the highest-yield place to look because those genes have actionable clinical pathways behind them. It can give recessive carrier status, pharmacogenomic star alleles, and polygenic scores whose calibration depends heavily on the ancestry of the training cohort.

It cannot tell you what most of your variants do. A typical genome carries thousands of rare coding variants, the majority classified “of uncertain significance,” and interpreting non-coding variation remains largely an open problem despite functional genomics atlases making progress on regulatory effects 2. Penetrance is the other trap: a variant curated from affected families often looks much less dangerous when measured in an unselected population, so “pathogenic” is a statement about a variant class, not a prediction about you. Reporting guidelines exist precisely because turning a VCF row into a clinical statement requires evidence review, orthogonal confirmation, and explicit classification criteria 1.

Sequencing also produces findings you did not go looking for, including carrier status affecting relatives and results with implications for insurance and family relationships. These are well-described consequences, not edge cases 3. Professional bodies have argued for targeted analysis of the sequence rather than open-ended interrogation of everything, exactly to keep incidental-finding burden manageable 4. You can decide to look at everything. Decide deliberately.

Finally, your genome is fixed. It says nothing about current state: it will not tell you what your transcriptome is doing this month, what your inflammatory proteins look like, or how you handle a given meal. Those need measurement over time, and they are where most near-term actionable signal lives.

Questions people also ask

Which company is best for genetic testing? The one that gives you FASTQ or CRAM, a gVCF, honest QC metrics, PCR-free 30x, and CLIA/CAP accreditation. Rank providers on deliverables and accreditation rather than on report design. Report interfaces get replaced. Your raw data does not.

What can throw off a DNA test? Sample swaps, cross-contamination between samples (check FREEMIX), degraded saliva DNA yielding low-complexity libraries, PCR duplicates, wrong reference build (GRCh37 vs GRCh38 coordinates silently mismatch), unnormalized indels failing to join to ClinVar, and pseudogene interference in CYP2D6, PMS2, and SMN1.

Can I get genetic testing for all diseases? No. Sequencing reads the bases. Whether a base matters depends on curation that exists for perhaps a few thousand genes and is thin or absent elsewhere, and most common disease risk is polygenic and environmental.

Are genetic health tests worth it? A 30x genome with raw data is worth it once, as a permanent asset you re-analyze as databases improve. A $99 array health report is worth it only for curiosity, and its rare-variant claims should not be acted on.

What percentage of genetic tests are wrong? Depends on the test and the variant class. Well-covered SNVs from 30x WGS are highly accurate. Rare-variant calls from consumer arrays fail confirmation at rates high enough that clinical labs require re-testing on a new sample before reporting. Anything consequential goes to a clinician and a clinical lab.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Christina Austin‐Tse, Vaidehi Jobanputra, Denise Perry, et al. Best practices for the interpretation and reporting of clinical whole genome sequencing. npj Genomic Medicine, 2022. https://doi.org/10.1038/s41525-022-00295-z ↩ ↩2

  2. Tuuli Lappalainen, Alexandra J. Scott, Margot Brandt, et al. Genomic Analysis in the Age of Human Genome Sequencing. Cell, 2019. https://doi.org/10.1016/j.cell.2019.02.032 ↩

  3. Wim Pinxten, Heidi Howard. Ethical issues raised by whole genome sequencing. Best Practice & Research Clinical Gastroenterology, 2014. https://doi.org/10.1016/j.bpg.2014.02.004 ↩

  4. Carla van El, Martina C. Cornel, Pascal Borry, et al. Whole-genome sequencing in health care. European Journal of Human Genetics, 2013. https://doi.org/10.1038/ejhg.2013.46 ↩