What Whole Genome Sequencing Measures

“Whole genome” means all ~3.1 billion base pairs of your haploid genome sequence, coding and non-coding, read redundantly enough that you can call variants at most positions with confidence. In practice you get a pile of short reads (typically 2×150 bp), aligned to a reference, and a variant file listing the roughly 4–5 million places where you differ from that reference: about 3.5 million single-nucleotide variants, 500–700k small insertions and deletions, and several thousand to tens of thousands of structural variants depending on the caller and how generous you are about the definition. That is the product. Everything else is interpretation layered on top.

The important distinction against a genotyping array (23andMe, AncestryDNA) is that an array measures a fixed panel of ~600k–1M pre-chosen positions and infers the rest statistically. WGS measures the positions directly, including ones no one has catalogued yet. Exome sequencing measures the ~1–2% that codes for protein, at higher depth per dollar. Whole genome measures everything at uniform depth, which is why it also gives you structural variants, mitochondrial DNA, non-coding regulatory regions, and enough off-target coverage to estimate copy number.

Depth, and why 30× is the floor

Coverage depth is the mean number of independent reads overlapping a given base. At 30× mean depth, a heterozygous site has enough reads on both alleles that you rarely lose one to sampling noise. Below about 20×, heterozygote sensitivity degrades noticeably and you start miscalling hets as homozygous reference.

Be careful about what “30×” is measured against. Some providers quote raw depth (total bases sequenced ÷ genome size) rather than mean mapped, deduplicated depth. The gap between them can be 15–25% once you remove optical duplicates, PCR duplicates, and unmapped reads. Ask for the post-deduplication figure and the fraction of the genome covered at ≥10× and ≥20×. A good 30× PCR-free run gives you ≥95% of GRCh38 at 10× or better.

Two other quality numbers worth demanding: the fraction of bases at Q30 or above (should be >85–90% for a 2×150 run) and whether the library was PCR-free. PCR-free matters because amplification introduces GC bias, which thins coverage in GC-rich promoters and first exons, exactly the regions you care about. The underlying chemistry, reversible-terminator sequencing-by-synthesis, has been the workhorse since the first human genome was sequenced with it in 2008 ¹.

The files you should receive

Insist on these, in this order of importance:

• FASTQ (gzipped, R1/R2). Raw reads plus per-base quality. ~50–60 GB for 30×. This is the only truly reprocessable artifact. If the reference genome changes, or a better aligner ships, you can redo everything from here.
• CRAM (with the exact reference FASTA and its .fai). Aligned reads, reference-compressed, ~15–25 GB for 30×. BAM is the same content at 3–4× the size. CRAM without the matching reference is unreadable, so get the reference URL and MD5.
• gVCF. Per-position records including reference-confident blocks. This is what tells you the difference between “you match the reference here” and “we had no coverage here.” A plain VCF cannot distinguish those, and that distinction is the difference between a real negative and a blind spot.
• VCF with variants normalized and left-aligned.
• Run metrics, which are the following.
  • mosdepth output
  • samtools flagstat
  • duplicate rate
  • insert size distribution
  • contamination estimate (VerifyBamID or equivalent, FREEMIX below 0.02)

If a vendor will only give you a PDF report and a “raw data” download that turns out to be a 23andMe-style genotype text file, that is not whole genome sequencing.

The pipeline we would run

Align with bwa-mem2 mem -t 32 -K 100000000 -Y against GRCh38 with alt contigs and decoys (the GCA_000001405.15_GRCh38_no_alt_analysis_set build, or the full-alt build if you intend to do HLA typing). Mark duplicates with samblaster or gatk MarkDuplicatesSpark. Sort and compress to CRAM with samtools sort --write-index --output-fmt CRAM --reference.

Call small variants with DeepVariant (--model_type=WGS) rather than GATK HaplotypeCaller. On a single sample with no cohort to joint-call against, DeepVariant’s indel precision is better out of the box and it requires no base-quality recalibration or hard-filter tuning. Then normalize: bcftools norm -m -any -f GRCh38.fa -Ou | bcftools norm -d exact. Skipping normalization is the single most common reason a variant fails to match a ClinVar record that is sitting right there.

Structural variants need a separate pass. Run Manta and smoove (LUMPY plus duphold) and take the intersection as your high-confidence set, the union as your candidate set. Expect short-read SV calling to be the weakest part of the whole exercise: sensitivity for deletions is decent, for insertions poor, for balanced events near zero.

Annotate with VEP or bcftools csq against a pinned Ensembl release, then join against ClinVar (dated VCF, not a web query), gnomAD v4 for allele frequencies, and SpliceAI for non-coding candidates. Pin every database version in the output filename. An annotation from March is a different annotation from one in September.

Where short-read WGS quietly fails

Roughly 5–8% of GRCh38 is not reliably callable with 150 bp reads. The failures cluster:

• Segmental duplications and pseudogenes. CYP2D6 (pharmacogenetics), SMN1/SMN2, PMS2, GBA and HBA1/HBA2. Reads map ambiguously or to the wrong paralog. A standard pipeline will emit confident-looking calls here that are wrong. Use paralog-aware callers (Cyrius for CYP2D6, SMNCopyNumberCaller for SMN) or accept that you have no data.
• Short tandem repeat expansions. A 150 bp read cannot span a 500-repeat expansion. ExpansionHunter genotypes known loci from anchored read pairs, with real but bounded accuracy.
• HLA. Needs dedicated typing (HLA-LA, Optitype) against IMGT/HLA. The generic caller will not do here.
• Highly GC-rich first exons, if the library was amplified.

None of this is a reason to skip WGS. It is a reason to know which of your “no variant found” results are actual negatives and which are coverage holes. That is what the gVCF is for.

What the data supports, and what it does not

Comprehensive pharmacogenomic genotyping from WGS is one of the genuinely well-supported uses: a single genome contains the star-allele-defining variants for most of the major drug-metabolizing enzymes, and researchers have demonstrated extracting full pharmacogenomic profiles from whole-genome data ². That said, interpretation, and any decision that follows from it, belongs with a clinician and a pharmacist. We describe measurement here, nothing more.

For disease risk in an unselected healthy adult, calibrate your expectations downward. Sequencing healthy people surfaces a large number of variants of uncertain significance, and the positive predictive value of a pathogenic-looking variant found without a matching family history or phenotype is considerably lower than the same variant found in an affected patient ³. Around 80 genes are on the ACMG secondary findings list precisely because those are the ones with actionable, well-established consequences. The other ~19,000 protein-coding genes mostly yield ambiguity. The field’s central bottleneck is interpretation. Sequence generation is not the limiting step ⁴.

There are also consequences you should decide about before you sequence, not after: findings about relatives who did not consent, information you cannot un-know, and downstream disclosure questions ⁵. Professional guidance in Europe has argued for deliberately targeted analysis of genome data rather than unfiltered reporting of everything, for exactly these reasons ⁶. You can hold the raw CRAM and choose to query only specific regions. The file does not force you to look.

Questions people also ask

What does “whole genome” mean? Every base of your nuclear and mitochondrial DNA, read with enough redundancy to call variants across nearly all of it. In practice, that means ~3.1 Gb of reference sequence and ~4–5 million differences from the reference per person. About 5–8% of the genome is where short reads cannot give you a reliable answer.

What is the difference between WGS and a DNA test from a consumer array? An array measures a fixed set of ~1M known positions. WGS measures positions directly, including rare and novel variants that no array probe exists for, plus indels, copy number, and structural variation. An array also cannot distinguish “no variant” from “not measured,” because it only ever measured its own panel.

Is 30× enough, or do I need more? 30× post-deduplication is the standard for germline calling and is sufficient for nearly all single-sample work. Higher depth (50×+) helps mainly for mosaicism, somatic variants, and hard regions. If your goal is a germline baseline you will re-analyze for years, spend the marginal dollar on PCR-free library prep and keeping the FASTQs rather than on depth beyond 30×.

Can I analyze this myself? Yes, if you are comfortable with a 32-core machine, ~1 TB of scratch disk, and a day of compute. Alignment plus DeepVariant on 30× runs in roughly 6–20 hours depending on hardware. The harder part is annotation discipline: pin your database versions, normalize your VCF, and never compare an un-normalized variant to a ClinVar record.

Footnotes

1. David Bentley, Shankar Balasubramanian, Harold Swerdlow, et al. Accurate whole human genome sequencing using reversible terminator chemistry. Nature, 2008. https://doi.org/10.1038/nature07517 ↩

2. Clint Mizzi, Brock A. Peters, Christina Mitropoulou, et al. Personalized Pharmacogenomics Profiling Using Whole-Genome Sequencing. Pharmacogenomics, 2014. https://doi.org/10.2217/pgs.14.102 ↩

3. Noralane M. Lindor, Stephen N. Thibodeau, Wylie Burke. Whole-Genome Sequencing in Healthy People. Mayo Clinic Proceedings, 2017. https://doi.org/10.1016/j.mayocp.2016.10.019 ↩

4. 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 ↩

5. 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 ↩

6. 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 ↩