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

If you want to work with your own genetic data, buy 30x (or deeper) whole-genome sequencing from a provider that hands you FASTQ or CRAM plus a gVCF, and skip the consumer genotyping arrays entirely. An array measures a few hundred thousand pre-chosen positions and infers the rest statistically. A 30x genome measures all ~3.1 billion bases directly, calls roughly 4–5 million small variants against the reference, and gives you files that stay useful for a decade as annotation databases improve. The price difference is a few hundred dollars versus a couple of thousand. The information difference is two orders of magnitude.

Arrays versus sequencing

The top-selling kits (AncestryDNA, 23andMe) run a SNP microarray: a fixed panel of probes, typically 600,000 to 1 million sites, chosen for ancestry informativeness and common-variant tagging. Ancestry estimates and relative matching work well on that data because those tasks depend on common haplotypes. Anything that depends on a rare variant does not.

Two specific failure modes matter if you plan to analyze the raw download yourself:

1. Rare-variant genotyping on arrays has poor positive predictive value. Array calls come from clustering fluorescence intensities across a batch of samples. With a variant at 0.01% frequency, almost every sample sits in the homozygous-reference cluster, and the rare-allele cluster is defined by a handful of points or none. Heterozygote calls at those sites are frequently artifacts. This is why third-party interpretation of raw array files flags “pathogenic” variants that disappear on confirmatory sequencing.
2. Coverage of clinically interpreted genes is token. A consumer BRCA1/BRCA2 report typically covers three Ashkenazi founder variants. There are thousands of pathogenic and likely pathogenic variants catalogued in those two genes in ClinVar. A negative result on three of them is close to uninformative about the other several thousand.

Imputation fills the gaps for common variation and does it well. Using a large reference panel, you can impute tens of millions of common SNPs from an array backbone at high accuracy, which is enough for polygenic scores. It gives you nothing at low allele frequency, no indels beyond the panel, no structural variants, and no coverage of exons the array never touched.

What 30x whole-genome sequencing resolves, and what it doesn’t

At 30x mean coverage with 150 bp paired-end reads, you get reliable SNV and small-indel calls across roughly 95% of the genome, including essentially all coding exons that are uniquely mappable. You also get read depth and split-read signal, which means copy-number and structural variant calling becomes possible: deletions, duplications, inversions, mobile element insertions. Mitochondrial DNA comes along free at very high depth (often 1,000x+), so heteroplasmy fractions down to a few percent are measurable.

What short-read 30x still handles badly, and what you should not assume your VCF covers:

• Highly homologous regions. SMN1/SMN2, PMS2/PMS2CL, CYP2D6/CYP2D7, HBA1/HBA2. Reads multimap and callers either drop them or emit garbage. Dedicated callers (SMNCopyNumberCaller, PharmCAT with a CYP2D6 caller) exist for a reason.
• Repeat expansions. ExpansionHunter can genotype known loci if the expansion is shorter than the read insert, but large expansions are invisible.
• HLA typing. Use a dedicated tool (HLA*LA, OptiType) rather than trusting variant calls in the MHC.
• Phasing. Short reads phase only within a few hundred base pairs unless you add statistical phasing or long reads. Compound heterozygosity across a gene often can’t be resolved from the VCF alone.

If any of those matter to you specifically, long-read sequencing (PacBio HiFi or ONT) at 15–20x solves most of them and is now the better choice for structural variation and phasing, at roughly 2–4x the cost.

Files to insist on before you pay

Ask the provider, in writing, what you receive. The minimum useful set:

• sample.cram aligned to GRCh38 (with the reference FASTA and its .fai), or raw sample_R1.fastq.gz / sample_R2.fastq.gz. A CRAM of a 30x genome is roughly 15–20 GB; gzipped FASTQs run 80–100 GB.
• sample.g.vcf.gz plus .tbi. The gVCF matters because it records reference confidence at non-variant sites. Without it you cannot tell “no variant here” from “no coverage here”, and that distinction is the difference between a real negative and a gap.
• A QC report: mean and median coverage, percent of callable bases at ≥10x and ≥20x, duplicate rate, insert size distribution, contamination estimate (VerifyBamID FREEMIX, should be < 0.02), and Ti/Tv ratio (~2.0–2.1 genome-wide, ~3.0 in exons).

If a provider offers only a PDF report and a “raw data” CSV of genotypes, that is an array, whatever the marketing says.

The pipeline we would run

Start from FASTQ if you have it, because you control the reference build and the caller version.

bwa-mem2 mem -t 32 -R '@RG\tID:1\tSM:me\tPL:ILLUMINA' \
    GRCh38_full_analysis_set_plus_decoy_hla.fa R1.fq.gz R2.fq.gz \
  | samtools sort -@ 8 -o me.bam -
samtools index me.bam

Use the full analysis set with decoys and alt contigs. Aligning to a naked primary assembly pushes reads that belong on decoy sequence into real genes and creates false heterozygous calls.

For small variants we would run DeepVariant rather than GATK HaplotypeCaller on a single sample. It has fewer knobs, better indel precision on Illumina WGS, and does not need a cohort for VQSR:

run_deepvariant --model_type=WGS --ref=GRCh38.fa --reads=me.bam \
  --output_vcf=me.vcf.gz --output_gvcf=me.g.vcf.gz --num_shards=32

Then structural variants with Manta plus a depth caller, repeats with ExpansionHunter, and pharmacogenes with PharmCAT. Annotate with VEP:

vep -i me.vcf.gz --cache --assembly GRCh38 --everything \
  --plugin CADD,whole_genome_SNVs.tsv.gz --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG \
  --vcf -o me.annot.vcf.gz

Then filter. Almost everything is noise: of ~4.5 million variants, a few hundred will be rare (gnomAD AF < 0.001) and predicted loss-of-function, and a handful will be in genes with established disease associations and ClinVar pathogenic assertions with review status of two stars or better. That filtering logic mirrors what clinical labs do, and published best practices for clinical WGS interpretation are worth reading before you trust your own output. ¹

For polygenic scores, download weights from the PGS Catalog and run plink2 --score <weights> cols=+scoresums --score-col-nums 3. Report your score as a percentile against an ancestry-matched reference (1000 Genomes superpopulations work), because raw sums are meaningless and scores derived in European cohorts transfer poorly to other ancestries.

Where you need a clinician

Anything you find that looks like a pathogenic variant in a clinically actionable gene needs orthogonal confirmation in a CLIA/CAP laboratory before it means anything about you, and interpretation with a genetic counselor or medical geneticist. Research-grade pipelines produce false positives in low-complexity regions, and variant classification is a structured process involving population frequency, functional data, segregation, and curated evidence, not a ClinVar lookup. ¹ The counseling piece is not a formality. The ethics literature on genome sequencing is largely about the burden of uncertain and unsolicited findings, insurance and family implications, and the fact that a result can’t be un-known once delivered. ²

Questions people also ask

Is genetic testing really worth it? For a technically fluent person who will analyze the data: yes, if you buy sequencing rather than an array, because the files remain interpretable as databases grow. Roughly 1–3% of unselected adults carry a pathogenic variant in the ACMG secondary-findings gene set (about 80 genes covering hereditary cancer, cardiomyopathies, arrhythmias, and a few metabolic conditions). Carrier status and pharmacogenetics apply to nearly everyone.

How much does a full genetic test cost? 30x WGS with raw data delivery runs roughly $400–$1,500 depending on provider and whether interpretation is included. Clinical-grade WGS ordered through a physician, with a curated report, is typically several thousand.

How long do results take? Sequencing itself is 2–4 days of instrument time. Consumer WGS providers typically quote 4–8 weeks door to door. Your own analysis on a 32-core machine takes about 12–24 hours from FASTQ to annotated VCF.

Is there a downside? Yes. Variants of uncertain significance vastly outnumber actionable ones, incidental findings implicate relatives who never consented, and results can affect life, disability, and long-term-care insurance in many jurisdictions. These are the central arguments in the debate over sequencing asymptomatic people, including newborns. ^{3 2}

What is the most commonly used genetic test? By volume, the consumer SNP array. In clinical practice, targeted gene panels and chromosomal microarray remain the most-ordered tests, with exome and genome sequencing growing as first-tier options for suspected genetic disease. ¹

What can go wrong? Sample swaps and contamination (check FREEMIX), misalignment in segmental duplications, uncalled regions silently read as negative when you skip the gVCF, and misclassified variants in ClinVar that a naive script treats as truth.

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 ↩³

2. 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 ↩ ↩²

3. Csaba Szalai. Arguments for and against the whole-genome sequencing of newborns. PubMed, 2023. https://pubmed.ncbi.nlm.nih.gov/37969196 ↩