Simulating whole exome sequencing (WES) for suspected rare Mendelian disease — from coding-region capture to a confirmed molecular diagnosis
Whole exome sequencing (WES) is built on a simple economic and biological insight: although the coding exome makes up only about 2% of the human genome (~30–40 Mb of the ~3.1 Gb total), it harbors an estimated 85% of known disease-causing Mendelian variants. Rather than sequence all 3 billion bases at high depth, WES uses hybridization capture probes to selectively enrich exonic fragments before sequencing — cutting cost and turnaround time dramatically while still catching the large majority of clinically actionable variation.
Genomic DNA is sheared into ~200bp fragments, ligated to sequencing adapters, and then hybridized in solution against a library of biotinylated RNA or DNA probes (baits) that are complementary to exonic sequence across the genome (e.g. Agilent SureSelect, IDT xGen, Twist Exome).
Streptavidin-coated magnetic beads pull down probe-fragment hybrids; everything that did not bind a probe — introns, intergenic DNA, most regulatory sequence — is washed away and discarded. The enriched library is then PCR-amplified and sequenced.
Capture is imperfect: GC-rich exons, first exons, and small single-exon genes are systematically under-captured, creating blind spots that any yield estimate must account for.
Three sequencing strategies compete for the rare-disease diagnostic workup, each with different tradeoffs:
• Targeted gene panel (10s-100s of genes): cheapest, fastest, highest depth per gene, but only useful when the phenotype strongly implicates a known disease gene list. Misses everything outside the panel.
• Whole exome sequencing (WES): broad coverage of all coding genes at moderate-high depth (~100x), reasonable cost, but misses ~15% of pathogenic variation that lies in introns, regulatory regions, or is structural (large deletions/duplications, repeat expansions).
• Whole genome sequencing (WGS): captures coding AND non-coding variation uniformly, better structural variant detection, no capture bias — but costs more, generates far more incidental/uncertain findings, and requires much more computational and interpretive infrastructure.
In practice, WES has become the first-tier test for undiagnosed suspected-Mendelian patients in most health systems, with WGS reserved for exome-negative cases or specific structural-variant-suspicious phenotypes.
Because coding variants are so enriched for disease relevance, WES currently offers close to WGS diagnostic yield at a fraction of the cost and analytic burden — which is why it remains the dominant first-line test, even as WGS prices continue to fall.
The capture-based design of WES creates predictable blind spots that directly cap its diagnostic ceiling:
• Deep intronic variants that create cryptic splice sites — invisible because introns are not captured • Regulatory/promoter/enhancer variants affecting gene expression without altering protein sequence • Large structural variants (deletions, duplications, inversions) that span capture boundaries poorly • Repeat expansions (e.g. trinucleotide repeat disorders) — read-length and capture-efficiency limited • Mitochondrial genome variants — usually requires separate mtDNA-specific assays • Variants in genes not yet linked to any disease — even if captured and called, they cannot be prioritized
Each of these categories represents a share of the ~60-75% of exomes that end up unsolved, and each motivates a specific complementary test (WGS, RNA-seq, repeat-expansion assays, mitochondrial panels).
Once exonic DNA fragments are enriched, they are loaded onto a short-read sequencer (typically Illumina NovaSeq) which reads both ends of each ~200bp fragment as paired ~150bp reads. Millions of these reads are generated per sample and computationally aligned back to the GRCh38 reference genome, stacking up at each coding position to build a depth-of-coverage profile. Depth and uniformity of this coverage directly determine which true variants can be confidently detected.
A variant can only be called with confidence if enough independent reads span that position and agree on the alternate allele. At 10x coverage, a true heterozygous variant (expected ~50% allele fraction) might by chance be represented by only 1-2 alt reads — statistically indistinguishable from sequencing error. Clinical-grade WES pipelines therefore target a mean depth of 80-150x, aiming for at least 20x coverage across more than 95% of the targeted bases, with 8x commonly cited as an absolute analytic floor.
Coverage is never uniform: GC-rich exons, repetitive regions, and the edges of capture-probe boundaries systematically receive fewer reads, creating coverage "valleys" that can hide a causal variant even when the gene itself is nominally "covered" by the test.
A meaningful fraction of exome-negative cases are not biologically unsolvable — the causal variant was simply sitting in a low-coverage or capture-dropout region and was never confidently called in the first place. Studies auditing "exome-negative" cohorts with targeted re-sequencing of poorly covered exons recover a real, if modest, share of additional diagnoses.
This is one of the main arguments for genome sequencing as a second-tier test after an unrevealing exome: WGS uses PCR-free, capture-free library preparation, giving flatter and more uniform coverage across the genome, including exons that WES chronically under-captures.
Sequencing cost has fallen roughly 10-fold per decade since the Illumina short-read platforms matured, which is what made population-scale exome sequencing economically viable for routine clinical diagnostics rather than only research. A modern clinical WES run — capture, sequencing, alignment, variant calling, and interpretation — typically completes in 2-6 weeks for a standard case, and can be compressed to 24-72 hours for rapid/ultra-rapid protocols used in critically ill neonates (rWES/rWGS), where a molecular diagnosis can directly change acute clinical management.
With reads aligned and depth established, a variant-calling algorithm (most commonly GATK HaplotypeCaller, alongside DeepVariant and others) walks the genome and emits every position where the sample differs from the GRCh38 reference. A typical individual exome yields somewhere between 20,000 and 40,000 raw variant calls — the overwhelming majority of which are common, benign polymorphisms shared by most humans, not the disease-causing needle hiding in this haystack.
Raw sequencer output moves through a standardized processing pipeline before any variant is trusted:
1. Alignment — reads mapped to GRCh38 with BWA-MEM, producing a BAM/CRAM file 2. Duplicate marking — PCR/optical duplicate reads flagged so they do not inflate apparent depth 3. Base quality score recalibration (BQSR) — systematic sequencer error patterns corrected 4. Variant calling — HaplotypeCaller (or DeepVariant) performs local reassembly around candidate sites and emits genotype likelihoods 5. Joint genotyping — when parents are also sequenced (trio), all samples are genotyped together, which sharply improves accuracy at low-coverage sites 6. Variant quality score recalibration (VQSR) / hard filtering — statistical filtering separates likely-true variants from likely artifacts
Only after this pipeline completes does the resulting VCF (Variant Call Format) file — the 20,000-40,000-row starting point — enter the interpretive filtering funnel.
The vast majority of calls are single-nucleotide variants (SNVs) — a single base substitution — with a smaller fraction of small insertions/deletions (indels), typically under 50bp. Short-read technology is well-suited to detecting both reliably in well-covered, non-repetitive sequence.
What it detects far less reliably: large indels, structural variants (deletions/duplications spanning many exons), and variants in low-complexity or highly repetitive regions, where short 150bp reads cannot be uniquely placed. These categories require complementary methods — CNV-calling algorithms applied to the same read-depth data, chromosomal microarray, or long-read sequencing.
Every human genome carries tens of thousands of variants relative to the reference simply because the reference is itself just one arbitrary genome — human populations are enormously polymorphic. The overwhelming majority of an individual's 20,000-40,000 exome variants are common benign polymorphisms carried by a large fraction of the population, several thousand are rare-but-benign private variants, and only a tiny handful — often exactly one, if the case is solvable at all — is the actual pathogenic cause of the patient's disease. This is precisely the needle-in-a-haystack problem the filtering pipeline in Stage 4 is built to solve.
Turning 20,000-40,000 raw calls into an actionable candidate list requires a cascade of filters, each removing variants that are statistically or biologically unlikely to explain the patient's disease. The three dominant filters are population allele frequency, inheritance-model segregation (dramatically sharpened by trio sequencing), and phenotype-gene matching — and the order and rigor with which they are applied is the single biggest lever on diagnostic yield.
Rare Mendelian disease is, definitionally, rare — so the first and most powerful filter removes any variant that is too common in reference population databases (gnomAD, aggregating >800,000 sequenced individuals) to plausibly cause a rare severe disorder. A typical threshold excludes variants with an allele frequency above roughly 0.1%-1% (adjusted for the disease's inheritance pattern and prevalence). This single filter eliminates well over 95% of the raw call set in one step, because the vast majority of an individual's variants are shared, common human polymorphisms.
Once frequency filtering has trimmed the list to a few hundred rare variants, the inheritance model filter tests each variant against a mode of inheritance: is it de novo (absent in both parents), recessive with two hits (compound heterozygous or homozygous), dominant, or X-linked?
Singleton sequencing (patient only) cannot directly test any of this — it can only flag variants that are theoretically consistent with a given mode, leaving many false leads in a candidate list.
Trio sequencing (patient + both biological parents) is transformative: a de novo variant is instantly and definitively identified because it is absent in both parental exomes, and compound heterozygous variants can be phased — confirmed as sitting on two different parental chromosomes rather than co-inherited on one. This single change removes the great majority of false-positive candidates that singleton analysis cannot distinguish, which is why trio sequencing so reliably boosts diagnostic yield.
Trio sequencing does not just add more raw sequence — it adds segregation information. A single de novo dominant variant can be pinpointed among hundreds of rare candidates almost instantly once both parents are sequenced, something no amount of additional singleton analysis can replicate.
The final filter ranks remaining candidate genes by how well their known disease associations match the patient's clinical presentation. Clinicians encode the patient's features using the Human Phenotype Ontology (HPO) — a standardized vocabulary of ~18,000 terms describing everything from "microcephaly" to "atrial septal defect." Algorithms such as Exomiser and PhenIX combine variant-level pathogenicity scores with semantic similarity between the patient's HPO terms and each candidate gene's known phenotype associations (via OMIM, Orphanet, and model-organism databases) to rank candidates.
The sharper and more complete the phenotype description, the more precisely this filter can separate the true causal gene from incidental rare variants — a vague or incomplete phenotype (e.g. "developmental delay" alone) leaves many genes plausible, while a rich, specific phenotype (e.g. a constellation of five or six distinctive features) can often narrow the field to a single strong candidate.
After the filtering funnel, each remaining candidate variant is formally classified using ACMG/AMP criteria into one of five tiers: pathogenic, likely pathogenic, uncertain significance (VUS), likely benign, or benign. The case outcome follows directly: a pathogenic/likely pathogenic variant that explains the phenotype means the case is solved; a plausible but unclassifiable variant leaves the case partially solved (VUS); and no compelling candidate at all leaves the case unsolved — not necessarily forever.
Solved: a variant meeting pathogenic/likely-pathogenic ACMG criteria, consistent with the patient's inheritance pattern and phenotype, is returned as the molecular diagnosis. This can end a multi-year diagnostic odyssey, redirect clinical management, inform reproductive counseling, and connect families to disease-specific research and support communities.
Partially solved (VUS): a variant is biologically plausible — right gene, right region, some evidence of impact — but current evidence is insufficient to classify it confidently. These cases are flagged for follow-up: functional studies, additional family testing, or future reclassification as evidence accumulates.
Unsolved: no candidate survives the funnel with adequate support. This does not mean no genetic cause exists — it often means the causal variant lies outside what exome sequencing can see (Stage 1's blind spots), or in a gene not yet linked to any disease.
Across published rare-disease cohorts, moving from singleton to trio sequencing typically raises diagnostic yield by roughly 15-30 percentage points, driven almost entirely by the segregation-filtering advantage described in Stage 4: instant de novo identification and compound-heterozygous phasing. This is why current clinical guidelines increasingly recommend trio sequencing as the default starting configuration whenever both biological parents are available, rather than a "singleton-first, trio-if-negative" fallback strategy — trio-first meaningfully shortens the diagnostic odyssey.
Roughly 3,000-5,000 human genes currently have a confidently established disease association out of ~20,000 protein-coding genes — and dozens of new gene-disease relationships are published every month. A variant sitting quietly in an unsolved patient's exome, in a gene with no known disease link at the time of the original analysis, can become diagnostic the moment that gene is characterized.
Systematic reanalysis of previously unsolved exome data — without re-sequencing, just re-running current variant interpretation pipelines against updated gene-disease databases — recovers an additional diagnosis in roughly 10-15% of unsolved cases within 2-3 years, with yield continuing to climb the longer a case sits unsolved. Many clinical genetics programs now build periodic reanalysis directly into their unsolved-case workflow rather than treating a negative exome as a final answer.
For cases that remain unsolved even after reanalysis, whole genome sequencing is typically the next step — reaching into the non-coding, structural, and repeat-expansion territory that exome capture structurally cannot see. That comparison of exome versus genome sequencing yield is the natural next chapter in the rare-disease diagnostic odyssey.