Whole Exome Sequencing: What It Covers, What It Misses, and How to Analyze It

Whole exome sequencing reads the protein-coding portion of your genome. That is roughly 20,000 genes and about 180,000 exons. It comes to 30–60 Mb of sequence depending on the capture kit, or 1–2% of the 3.1 Gb genome. Because Mendelian disease variants are concentrated in coding sequence, an exome captures a large fraction of high-penetrance pathogenic variation for a fraction of the sequencing cost of a genome. The tradeoff is that everything outside the capture targets is invisible, coverage across the targets is uneven, and structural variation is hard to call from capture data. If you want one sentence: an exome is a good instrument for coding single-nucleotide variants and small indels, a mediocre one for copy number, and blind to deep intronic and regulatory variation.

What “exome” means in practice, and why kits differ

Exome sequencing is hybrid capture. Biotinylated RNA or DNA baits, 60–120 bp, are hybridized to a fragmented sequencing library, pulled down with streptavidin beads, washed, and amplified. The bait set defines the assay. Common kits and their nominal target sizes:

• Twist Bioscience Exome 2.0: ~36.8 Mb core target, with optional RefSeq and mitochondrial add-ons.
• Agilent SureSelect Human All Exon V8: ~35 Mb.
• IDT xGen Exome Hyb Panel v2: ~34 Mb.
• Illumina Exome (formerly Nextera/TruSeq): ~45 Mb with padding.

“Exome” is not a standard. A variant in a UTR or a poorly conserved last exon may be covered by one kit and absent from another. When you get exome data, the first file you should ask for is the target BED. The VCF should not come before the BED. Without it you cannot distinguish “no variant” from “no reads.”

The original demonstration of the method sequenced 12 exomes and found the causal variant in a patient with congenital chloride diarrhea by looking for homozygous coding changes in a consanguineous background, which is still the cleanest illustration of why exomes work: the search space after filtering is small enough to read by eye.¹ The approach generalized quickly across Mendelian disease.²³

The pipeline we would run

Assume you have paired-end FASTQs, 2×150 bp, and the target BED. This is the sequence of commands, with the flags that matter.

Alignment. BWA-MEM2 against GRCh38 with the full ALT-aware analysis set:

bwa-mem2 mem -t 32 -K 100000000 -Y \
  -R '@RG\tID:S1\tSM:S1\tPL:ILLUMINA\tLB:lib1' \
  GRCh38_full_analysis_set_plus_decoy_hla.fa R1.fq.gz R2.fq.gz \
| samtools sort -@ 8 -o S1.bam -

-K 100000000 fixes the batch size so output is deterministic across thread counts. -Y uses soft clipping for supplementary alignments, which structural variant callers expect. Use the decoy-containing reference. Without decoys, reads from unplaced contigs and HLA pile onto the wrong loci and manufacture false variants.

Duplicate marking. gatk MarkDuplicates or samtools markdup. For exomes with PCR in the library prep, duplicate rates of 10–25% are normal. Above 40%, the library was over-amplified from too little input DNA and effective coverage is much lower than the raw number suggests. If the prep used UMIs (Twist and IDT both offer them), collapse by UMI with fgbio GroupReadsByUmi instead of positional dedup.

Base quality recalibration. gatk BaseRecalibrator then ApplyBQSR with dbSNP and the known indel sets. Effect is modest on modern NovaSeq data, but it costs an hour.

Variant calling. gatk HaplotypeCaller -ERC GVCF -L targets.bed -ip 100 with 100 bp interval padding to catch splice-region variants just outside the baits. Then GenotypeGVCFs. For a single sample, hard filtering beats VQSR, which needs many samples to train:

gatk VariantFiltration -V raw.vcf.gz \
  --filter-expression "QD < 2.0 || FS > 60.0 || MQ < 40.0 || MQRankSum < -12.5 || ReadPosRankSum < -8.0" \
  --filter-name "snp_hard"

DeepVariant with --model_type=WES is the alternative and, in our experience, gives fewer false-positive indels in homopolymer runs. It is slower and needs the target BED passed as --regions.

Annotation. Ensembl VEP with --everything --pick_allele_gene --assembly GRCh38, plus plugins for CADD, SpliceAI, and gnomAD v4 allele frequencies. Annotate against MANE Select transcripts so your HGVS strings match what a clinical lab would report.

Coverage is the metric that decides your result

Run mosdepth --by targets.bed --thresholds 1,10,20,30 S1.bam. The number to look at is the fraction of target bases at ≥20×. A good exome hits 95–98%. That still leaves 700,000 to 1.8 Mb of coding bases under-covered. The gaps are not random, and they cluster in the following places.

• GC-rich first exons and CpG islands, where capture and PCR both underperform.
• Segmental duplications and paralogs, where multi-mapping reads get MAPQ 0 and are filtered out. Examples are PMS2, SMN1/SMN2, CYP21A2 and GBA.
• Genes with recent pseudogenes, where reads align to the wrong copy.

Before you conclude a gene is clean, check its per-exon coverage directly:

mosdepth --by gene_of_interest.bed --no-per-base gene S1.bam

Uneven coverage across the target is the structural limitation of capture, and it is why exome negatives are weak evidence.⁴

Copy number from exome data

Exome read depth carries CNV signal, but it is noisy: capture efficiency varies per bait by an order of magnitude, so you need a panel of 20+ technically matched samples to normalize against. Tools worth running: GATK gCNV (DetermineGermlineContigPloidy then GermlineCNVCaller in cohort mode), CNVkit (cnvkit.py batch --targets targets.bed --reference ref.cnn), and XHMM. Single-exon deletions are detectable but require careful filtering, and methods designed for small exonic events have been applied to autism cohorts specifically for this reason.⁵ Expect false-positive rates high enough that anything clinically relevant needs orthogonal confirmation by MLPA, qPCR, or a genome.

Where exomes fail, and when a genome is the better buy

An exome cannot see:

• Deep intronic variants that create cryptic splice sites.
• Promoter, enhancer, and other regulatory variants.
• Balanced translocations and inversions, since breakpoints are almost always intronic.
• Repeat expansions (FMR1, C9orf72, HTT), which need long reads or targeted PCR.
• Mitochondrial genome, unless the kit includes chrM baits and the pipeline calls heteroplasmy.

Whole genome sequencing at 30× gives more even coverage of coding regions than an exome at nominally higher depth, because there is no capture bias, and it adds structural variant resolution. Genome-based studies of de novo mutation in autism families depend on exactly that even coverage to characterize mutation rates and clustering.⁶ The comparison between the two assays has been laid out clearly in the pediatric setting.⁷ Our view: if cost per sample is within a factor of two, sequence the genome. Choose the exome when you need depth on coding regions (mosaicism, somatic work, deep tumor sequencing) and want to spend reads there rather than spread them thin.

Questions people also ask

Can whole exome sequencing detect autism? No. Autism is a clinical diagnosis made by a clinician from behavior and development. A sequencer cannot make that diagnosis. Exome sequencing can identify a monogenic cause in a minority of people with an autism diagnosis, and trio designs (proband plus both parents) raise yield by isolating de novo variants.⁸ Candidate and novel genes continue to come out of exome studies, including in female-only cohorts where the genetic architecture differs.⁹¹⁰ Most autism is polygenic and no exome will explain it.

What diseases can be diagnosed with whole exome sequencing? Mendelian conditions where the causal variant is coding. They include metabolic disorders, skeletal dysplasias, ciliopathies and hereditary cancer syndromes. They also include many forms of intellectual disability, hearing loss, and neuromuscular disease. Exomes have been productive across inherited disease broadly.³² Diagnosis is a clinical act. A VCF is data, and interpretation that will affect medical decisions belongs with a genetics clinician and a CLIA-certified laboratory.

Is exome sequencing worth it? If you want analyzable data about coding variation and you will do something with the BAM and VCF, yes. If you want a complete picture, the exome will frustrate you within a week of your first coverage report. Diagnostic yield in selected clinical cohorts typically runs 25–50%, which means a negative result is common and not reassuring.⁴

How much does an exome test cost? Research-grade exome sequencing and a standard pipeline run a few hundred dollars per sample at volume. Clinical exomes with interpretation and reporting are typically a few thousand dollars, and trio testing multiplies the sequencing cost.

Does insurance cover exome sequencing? Sometimes, for a specific clinical indication with documented prior negative testing, and usually with prior authorization. Consumer or curiosity-driven sequencing is generally not covered. Your clinician’s genetic counselor will know the payer’s criteria.

Footnotes

1. Murim Choi, Ute I. Scholl, Weizhen Ji, et al. Genetic diagnosis by whole exome capture and massively parallel DNA sequencing. Proceedings of the National Academy of Sciences, 2009. https://doi.org/10.1073/pnas.0910672106 ↩

2. Bahareh Rabbani, Mustafa Tekin, Nejat Mahdieh. The promise of whole-exome sequencing in medical genetics. Journal of Human Genetics, 2013. https://doi.org/10.1038/jhg.2013.114 ↩ ↩²

3. Gerald Goh, Murim Choi. Application of Whole Exome Sequencing to Identify Disease-Causing Variants in Inherited Human Diseases. Genomics & Informatics, 2012. https://doi.org/10.5808/gi.2012.10.4.214 ↩ ↩²

4. Martine Tetreault, Eric Bareke, Javad Nadaf, et al. Whole-exome sequencing as a diagnostic tool: current challenges and future opportunities. Expert Review of Molecular Diagnostics, 2015. https://doi.org/10.1586/14737159.2015.1039516 ↩ ↩²

5. Christopher S. Poultney, Arthur P. Goldberg, Elodie Drapeau, et al. Identification of Small Exonic CNV from Whole-Exome Sequence Data and Application to Autism Spectrum Disorder. The American Journal of Human Genetics, 2013. https://doi.org/10.1016/j.ajhg.2013.09.001 ↩

6. Jacob J. Michaelson, Yujian Shi, Madhusudan Gujral, et al. Whole-Genome Sequencing in Autism Identifies Hot Spots for De Novo Germline Mutation. Cell, 2012. https://doi.org/10.1016/j.cell.2012.11.019 ↩

7. David Bick, David Dimmock. Whole exome and whole genome sequencing. Current Opinion in Pediatrics, 2011. https://doi.org/10.1097/mop.0b013e32834b20ec ↩

8. Xiujuan Du, Xueren Gao, Xin Liu, et al. Genetic Diagnostic Evaluation of Trio-Based Whole Exome Sequencing Among Children With Diagnosed or Suspected Autism Spectrum Disorder. Frontiers in Genetics, 2018. https://doi.org/10.3389/fgene.2018.00594 ↩

9. Merlin Butler, Syed Rafi, Waheeda Hossain, et al. Whole Exome Sequencing in Females with Autism Implicates Novel and Candidate Genes. International Journal of Molecular Sciences, 2015. https://doi.org/10.3390/ijms16011312 ↩

10. Elif Funda Sener, Halit Canatan, Yusuf Ozkul. Recent Advances in Autism Spectrum Disorders: Applications of Whole Exome Sequencing Technology. Psychiatry Investigation, 2016. https://doi.org/10.4306/pi.2016.13.3.255 ↩