What a WGS Sample Is, and What You Get Back From One

A WGS sample is the physical material you send to a lab: in practice 4–10 mL of whole blood in a lavender-top EDTA (K2/K3) tube, or 2 mL of saliva in an Oragene OG-500-style stabilizing kit. From it the lab extracts genomic DNA, builds a sequencing library, and runs it on a short-read instrument until every base of your genome is covered an average of ~30 times. What comes back to you, if you ask for it, starts with 80–120 gigabase-pairs of paired-end reads as FASTQ or unaligned BAM. You also get an aligned CRAM against a named reference build. The last file is a VCF or gVCF containing roughly 4–5 million single-nucleotide variants and 500,000–900,000 short insertions and deletions relative to that reference. Everything else in a WGS report is downstream interpretation of that VCF.

Blood or saliva, and why it matters more than people think

Take blood if you have the option. Two reasons.

First, yield and integrity. A 4 mL EDTA draw typically yields 50–150 µg of high-molecular-weight genomic DNA. PCR-free library preps, the ones you want, need roughly 500 ng to 1 µg of intact input. PCR-free matters because amplification introduces GC-bias that thins coverage in promoters and the first exons of many genes, and it creates duplicate reads that inflate your raw yield without adding information. Ask the lab explicitly whether the prep is PCR-free.

Second, purity. Saliva contains oral bacteria, and the human fraction of a saliva-derived library varies a lot between people and between collections. A bad saliva sample can come back 20–30% non-human reads, which means you paid for 30x and received 22x. You can check this yourself after the fact: align, then look at the unmapped fraction in samtools flagstat, and classify a sample of unmapped reads with Kraken2 if you want the species breakdown. If the lab’s QC does not report human read fraction, assume it did not check.

Practical handling: EDTA blood is stable for DNA extraction at 4 °C for several days, frozen indefinitely. Do not use heparin tubes. Heparin carries over into the extract and inhibits polymerase, and the failure shows up as a library that will not amplify rather than as an obvious sample defect.

The files, and which ones to insist on

Ask for these, in this order of importance:

1. CRAM plus its .crai index, and the exact reference FASTA used. CRAM is reference-compressed, so a 30x human genome lands around 15–20 GB instead of 60–90 GB as BAM. It is useless without the byte-identical reference, so get the MD5 of the FASTA as well as the build name. The build name alone is not enough. “GRCh38” is not a specification, because the analysis set with decoys and no ALT contigs behaves differently from the full ALT-aware build.
2. gVCF as well as VCF. With a gVCF you can distinguish “this position matched the reference” from “we had no coverage here” because it records reference-confidence blocks. With a plain VCF, absence of a variant call is ambiguous. That ambiguity will bite you when you check a specific position and find nothing.
3. FASTQ or unaligned BAM, if you want the option to realign later against a different build. This is the expensive one to store (60–100 GB gzipped) and the one labs are most likely to delete after 30–90 days. Decide early.
4. The QC bundle: per-base coverage histogram, insert size distribution, duplicate rate and contamination estimate.

QC you should run before believing any variant

Run these yourself. They take under an hour on a laptop-class machine plus a few hours of I/O.

mosdepth --by 1000 --fast-mode sample sample.cram        # coverage distribution
samtools stats sample.cram > sample.stats                # insert size, error rate, dup rate
verifyBamID2 --SVDPrefix 1000g.phase3.100k.b38.vcf.gz.dat \
             --Reference GRCh38.fa --BamFile sample.cram # contamination
somalier extract -d extracted/ --sites sites.hg38.vcf.gz -f GRCh38.fa sample.cram

What you are looking for:

• Mean coverage ≥ 30x and, more importantly, ≥ 95% of the callable genome at ≥ 20x. Mean hides a multimodal distribution. A sample with 32x mean and a long low-coverage tail is worse than one with 29x mean and tight variance.
• Duplicate rate under ~10%. Higher means the library was over-amplified or under-loaded, and your effective depth is lower than the headline number.
• Freemix (contamination) under 0.02. Above ~3%, heterozygous calls start accumulating false positives, and low-allele-fraction calls become untrustworthy.
• Insert size mode around 350–450 bp for a standard PCR-free prep. A collapsed distribution means degraded DNA.
• Sex chromosome coverage and somalier genotype fingerprint to confirm the sample is yours. Sample swaps are rare but not hypothetical, and a fingerprint is cheap insurance if you ever sequence twice.

Depth, and what 30x buys

Thirty-fold is the germline convention because it is roughly where heterozygous SNV sensitivity plateaus for short reads: you need enough reads that a true het is unlikely to be seen as a hom-ref by chance. Deeper sequencing mostly buys you rare-variant confidence and indel accuracy. Deep WGS of 1,070 individuals at ~32x identified millions of novel variants absent from existing reference panels, most of them rare and population-specific, which is the practical argument for sequencing rather than genotyping an array.¹ If your ancestry is underrepresented in the panels your annotation tools use, a larger fraction of your variants will come back with no population frequency at all.

Somatic work is a different regime entirely. Calling mutations present in a subset of cells requires far more depth and a matched normal, and a benchmarking study across many pipelines found wide variation in sensitivity and precision between groups analyzing the same tumor-normal data.² Do not read germline-depth WGS as a cancer screen.

What WGS misses

Short-read WGS at 30x is blind or unreliable in several places that matter:

• Segmental duplications and homologous pseudogenes. Genes like SMN1/SMN2, PMS2, CYP21A2, and CYP2D6 have near-identical paralogs. Standard alignment mismaps reads between them. These need targeted callers or long reads.
• Repeat expansions. Trinucleotide expansions longer than the read length cannot be sized from 150 bp reads. Tools like ExpansionHunter estimate them from anchored read pairs, with real uncertainty.
• HLA typing. Requires a dedicated caller against an allele database. The caller does not work against the reference genome.
• Structural variants. Callable in principle from split reads and discordant pairs with Manta, GRIDSS or smoove. Precision drops sharply for events under ~1 kb and in repetitive regions. Expect a high false-positive rate and treat unconfirmed SV calls as hypotheses.
• Mitochondrial heteroplasmy at low fraction, which needs a dedicated mtDNA pipeline.

The other category of downside is interpretive rather than technical. A genome contains findings you did not go looking for, including variants with implications for relatives who did not consent to the sequencing, and the ethics literature on this is worth reading before you sequence rather than after.³ Any variant you intend to act on medically needs orthogonal confirmation in a clinical lab and interpretation by a genetic counselor or physician. Research-grade calls are not a diagnosis, and we do not treat them as one.

Cost per sample

Reagent cost for a 30x human genome on a high-output flow cell has fallen into the low hundreds of dollars, which is why research and consumer-facing 30x WGS is commonly priced between about $300 and $1,000 per sample. Clinical WGS, ordered through a physician with CLIA/CAP validation, confirmatory Sanger, and a written interpretation, runs several thousand dollars, because you are paying for validation and interpretation rather than sequencing. The price difference is almost entirely non-sequencing labor.

Budget for storage too. If you keep FASTQ plus CRAM plus VCFs, plan on 100–150 GB per genome and pick cold object storage with a retrieval plan you have tested once.

Questions people also ask

What is a WGS test? Sequencing of essentially the entire genome, both coding and non-coding, typically to ~30x average depth with short reads. It differs from exome sequencing, which targets ~1–2% of the genome. It also differs from consumer genotyping arrays, which assay a fixed set of several hundred thousand to a million pre-selected positions.

What does WGS tell you? The set of positions where your genome differs from a reference build: about 4–5 million SNVs, several hundred thousand short indels, and a noisier set of structural variants. What those differences mean depends entirely on annotation databases and population frequency panels, which are incomplete and improve over time. Reanalyzing an old VCF against current databases is often more valuable than resequencing.

What are the downsides of whole-genome sequencing? Technical blind spots in repeats, segmental duplications, and long expansions. A large number of variants of uncertain significance. Incidental findings, including ones that carry information about family members.³ And a permanent, re-identifiable data asset you now have to store and secure.

How much does whole-genome sequencing cost per sample? Roughly $300–$1,000 for research-grade 30x with data delivery, and several thousand for clinical-grade with interpretation. Ask what is included: PCR-free library, delivered CRAM and gVCF, and how long the lab retains raw data.

Footnotes

1. Masao Nagasaki, Jun Yasuda, Fumiki Katsuoka, et al. Rare variant discovery by deep whole-genome sequencing of 1,070 Japanese individuals. Nature Communications, 2015. https://doi.org/10.1038/ncomms9018 ↩

2. Tyler Alioto, Ivo Buchhalter, Sophia Derdak, et al. A comprehensive assessment of somatic mutation detection in cancer using whole-genome sequencing. Nature Communications, 2015. https://doi.org/10.1038/ncomms10001 ↩

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 ↩ ↩²