Choosing genomic tests by phenotype — CMA, WES, and WGS yields from real cohort data, and the reasoning behind test selection.
Tags: Neurogenetics · Clinical Decision-Making
The single most important habit in genomic medicine is to treat a published yield figure as a property of a cohort, not of a test. The same exome is reported at 4% in isolated dystonia (Zech 2020) and at 83% in a dedicated leukodystrophy clinic (Zerem 2023). Neither number is wrong; they describe different populations. A yield is a fraction, and most of the variation lives in the denominator: who was selected, what was already ruled out, and how the result was interpreted.
Cohort selection — the dominant force. Diagnostic yield rises with the prior probability of a monogenic cause. A leukodystrophy clinic enrolls children whose MRI already implies a Mendelian white-matter disease, so most of the denominator is genetically diagnosable. A broad developmental-delay cohort sweeps in milder, multifactorial phenotypes and dilutes the Mendelian fraction. In the meta-regression of Clark et al. 2018, every additional 1,000 probands in a study lowered the odds of diagnosis by 28%: small cohorts tend to be hand-picked, large laboratory series take broad referrals. The first question about any yield figure is therefore "who was in the denominator?"
Singleton vs. trio. Trio sequencing (proband + both parents) roughly doubles the odds of a diagnosis (OR 2.04, 95% CI 1.62–2.56; Clark et al. 2018). Doubling the odds is not doubling the yield: a 30% singleton yield becomes about 47%. In the DDD study, recruitment as a trio was the strongest predictor of diagnosis (OR 4.70; Wright et al. 2023). The mechanism is interpretive power: parental data shows a variant is de novo, phases two variants in a recessive gene, and removes inherited variants from the candidate list.
Prior testing erodes the denominator. A cohort that already had microarray or a gene panel has lost the easiest diagnoses, so the next test works on a harder, pre-filtered population. Sequential testing studies must be read in this light.
What counts as a diagnosis. Most studies count pathogenic and likely pathogenic variants; some also count variants of uncertain significance. The GEMINI study's 49% genome "molecular diagnostic yield" included VUS (Maron et al. 2023).
Database maturation makes yield dynamic. In Clark's meta-regression the odds of diagnosis rose about 16% per year from 2013 to 2017. Reanalysis of unsolved exome or genome data finds new diagnoses in about 10% (95% CI 6–13%), mostly from newly published gene-disease links (Dai et al. 2022). A negative exome is time-stamped, not permanent.
CMA platform. SNP arrays detect isodisomy and long runs of homozygosity; oligonucleotide-only arrays miss both. UPD underlies a subset of imprinting disorders (Prader-Willi, Angelman), and runs of homozygosity flag parental relatedness, raising the prior for recessive disease.
Severity and specificity. Early onset, encephalopathy, developmental comorbidity, multi-system involvement and distinctive examination or imaging features each raise yield, because each moves the phenotype toward the rare, single-gene tail. Isolated autism without intellectual disability sits in the multifactorial middle and shows the lowest yields.
Key Points
The three first-line genomic tests are not a quality ladder where more sequence is simply better; they are different windows onto the genome, each blind to what the others see best. Choosing well means matching the window to the kind of variant the phenotype predicts. The pooled yields below are from Clark et al. 2018 (37 studies, 20,068 children with suspected genetic disease of any kind).
Chromosomal Microarray (CMA) — pooled yield ~10% (95% CI 8–12%); 15–20% in developmental delay/ID, autism and congenital anomalies (Miller et al. 2010)
CMA measures dosage: it counts copies of DNA across the genome and flags deletions and duplications plus aneuploidy. SNP arrays add genotype, which reveals isodisomy and runs of homozygosity. CMA cannot read sequence, so it misses single-nucleotide variants, small indels, balanced rearrangements, repeat expansions and low-level mosaicism. Its enduring value is the CNV space, which is why it stays first-tier in multiple congenital anomalies, infantile spasms and intellectual disability when an exome is sent.
Exome Sequencing — pooled yield ~36% (95% CI 33–40%); 36% in NDD overall, 31% in isolated NDD and 53% in NDD with associated conditions (Srivastava et al. 2019)
Exome reads the ~1–2% of the genome that codes for protein, where most interpretable Mendelian variants lie. It misses deep intronic and regulatory variants, calls CNVs less reliably than CMA (capture depth is uneven), resolves structural variants poorly, cannot size repeat expansions and does not read methylation.
Genome Sequencing — pooled yield ~41% (95% CI 34–48%)
Genome removes the capture step and sequences at even depth, adding noncoding variants, structural variants with breakpoints, CNVs and mitochondrial DNA. How much it adds over exome is debated: Clark 2018 found no significant difference; Nurchis et al. 2023 found genome ahead in a network meta-analysis (OR 1.54, 1.11–2.12) although pooled yields were nearly identical (38.6% vs 37.8%); Pandey et al. 2025 found 30.6% vs 23.2% in within-cohort comparisons (OR 1.7, 0.94–2.92, not significant). The gain concentrates where exome is structurally blind: after a negative exome, structural variants, deep intronic splice variants and noncoding genes such as RNU4-2, whose de novo variants explain about 0.4% of NDD (Chen et al. 2024).
The repeat-expansion blind spot. Short-read sequencing assembles the genome from ~150-bp fragments, which cannot span a long tandem repeat, so the expansion cannot be sized. This is a property of the chemistry, not a coverage problem. Standard exome misses the commonest repeat disorders: Friedreich ataxia (FXN GAA), the CAG SCAs, CANVAS (RFC1 AAGGG), SCA27B (FGF14 GAA), fragile X/FXTAS (FMR1 CGG), DM1, Huntington disease and C9orf72. Short-read genome pipelines (e.g., ExpansionHunter) can flag some known loci, but a flag needs confirmation by repeat-primed PCR, Southern blot or long-read sequencing. Before declaring a genomic work-up complete, ask whether the phenotype points to a repeat disorder the sequencing could never have seen.
Key Points
Within epilepsy, yield tracks a single organising principle: the more the seizures index brain that was built wrong rather than brain that fires wrong, the higher the monogenic fraction. Encephalopathic, early-onset epilepsies with developmental impairment sit at the high-yield end; seizures in a normally developing child sit at the low end.
By test (all epilepsy). Sheidley et al. 2022 pooled 154 studies: overall yield 17%; genome 48%, exome 24%, multigene panel 19%, CMA 9%. Each estimate pools different cohorts, so these are not head-to-head comparisons. The only phenotypic factors associated with higher yield were developmental and epileptic encephalopathy and neurodevelopmental comorbidity.
By phenotype. Stefanski et al. 2021 (103 studies, 32,331 individuals): epilepsy 24% overall, 27.9% with intellectual disability, 36.8% with early-onset seizures. In a population-based Scottish cohort of every child presenting with epilepsy before age 3, a genetic cause was found in 31% (Symonds et al. 2021). In a laboratory series of 9,769 individuals tested with epilepsy gene panels, yield was 14.9–24.4% depending on the panel, and about a third of definitive diagnoses had possible precision-medicine implications (Truty et al. 2019).
Infantile Epileptic Spasms Syndrome (IESS). Meta-analysis of 30 studies: exome 26% (21–31%), multigene panel 20% (15–27%), CMA 14% (11–16%); exome and panel were comparable (P = 0.34), and 61.6% of individuals with genetic IESS may benefit from the diagnosis in clinical management (Feng et al. 2025). IESS is the epilepsy phenotype where CMA earns its place in the sequencing era: a 14% CNV yield justifies sending it alongside sequencing, not after.
Neonates and infants in intensive care. See section 6: in GEMINI, genome found a molecular diagnostic variant (pathogenic or VUS) in 49% vs 27% with a targeted neonatal panel (Maron et al. 2023).
Normally developing child with focal or generalized epilepsy. The monogenic fraction is smaller and skewed toward ion-channel and synaptic genes; a well-designed panel is a reasonable first step, and CMA contributes little.
Key Points
Across neurodevelopmental phenotypes, two levers move yield: how severe and syndromic the presentation is, and how much family structure the lab is given to interpret the data.
CP is the phenotype whose yield rises as you subtract patients: the stricter the exclusion of acquired causes, the more the remaining denominator is enriched for genetic disease. A known risk factor lowers but does not remove the chance of a genetic cause.
Key Points
Movement and white-matter disorders make the central lesson of this module concrete: the molecular architecture of a phenotype decides which test can find the answer, and front-loaded phenotyping (the examination, the MRI) decides the denominator.
Dystonia. Exome sequencing of 764 people with dystonia at 33 specialty centres found a diagnosis in 19% of 728 families: 4% (16/388) with isolated dystonia, 19% (19/98) with combined dystonia, and 45% (100/222) when other neurological signs accompanied the dystonia; 65% of diagnostic variants were in genes also linked to neurodevelopmental disorders (Zech et al. 2020).
Progressive (hereditary) cerebellar ataxia. In 260 mostly adult, sporadic patients, exome gave a clinically relevant detection rate of 52%, with 25% classified as pathogenic; reanalysis of the first 76 exomes raised the pathogenic rate by 7% (Ngo et al. 2020). The authors described ~50% as a relative ceiling for exome-based methods. Much of the remainder lies in variant classes exome cannot see:
After a negative exome in ataxia, ask whether dedicated repeat testing has been sent; in late-onset sporadic ataxia, send the repeat panel first or alongside sequencing. CANVAS (ataxia + sensory neuropathy + chronic cough) is the trap case: a textbook phenotype that exome will report as negative every time.
White-matter disease. The yield depends on how the cohort was assembled:
MRI pattern recognition (hypomyelination vs. demyelination vs. cystic or vacuolating patterns) is the essential pre-test step: it collapses the differential before any DNA is drawn, generating the specialty-cohort effect from section one by imaging rather than referral. See the Hereditary Ataxias module for clinical features.
Key Points
The pooled figures below are anchors, but each is a cohort-specific number; the reasoning behind it is what transfers to the patient in front of you.
Pooled yields in children with suspected genetic disease (Clark 2018, 37 studies, 20,068 children):
| Test | Diagnostic yield | Clinical utility (management changed, % of those tested) |
|---|---|---|
| CMA | 10% (8–12%) | 6% |
| Exome | 36% (33–40%) | 17% |
| Genome | 41% (34–48%) | 27% |
| Setting | Figure | Source |
|---|---|---|
| All epilepsy | CMA 9%, panel 19%, exome 24%, genome 48% | Sheidley 2022, 154 studies |
| Infants in intensive care, rapid genome | 40% diagnosed; care changed in 32% | Dimmock 2021 (Project Baby Bear, n=184) |
| Infants, genome vs neonatal panel | 49% vs 27% (pathogenic or VUS); care changed in 19% | Maron 2023 (GEMINI, n=400) |
| Reanalysis of unsolved exome/genome | ~10% (6–13%) | Dai 2022, 29 studies |
Clinical utility ≠ diagnostic yield. A negative test can still inform recurrence counseling, and a positive one earns its value when it changes something. The strongest evidence sits in infant intensive care: in Project Baby Bear, rapid genome sequencing changed care in 32% of infants and reduced net costs. In epilepsy, about a third of definitive diagnoses carry possible precision-medicine implications (Truty 2019), and 61.6% of genetically explained IESS may benefit in management (Feng 2025). Examples: SCN1A → avoid sodium-channel blockers; KCNQ2 → sodium-channel blockers such as carbamazepine; SLC2A1 (GLUT1) → ketogenic diet; ALDH7A1 → pyridoxine; SLC6A1 → valproate often effective. The decision to test, and which test, should be argued from what a result would change, not from the yield percentage alone.
Key Points
1. A 6-month-old with severe developmental delay, seizures, and no family history presents for genetic evaluation. Both parents are available for testing. Which testing strategy is best supported by current evidence?
For severe early-onset neurodevelopmental phenotypes (DEE, GDD with seizures), trio sequencing is strongly supported by evidence. The key advantage is detecting de novo variants — spontaneous mutations present in the child but absent from both parents — which are the primary cause of many severe neurogenetic conditions. Meta-analysis shows trio sequencing roughly doubles the odds of a diagnosis compared with singleton testing (OR ~2.04), a substantial but smaller rise in yield. Trio analysis also enables phasing of compound heterozygous variants and parent-of-origin determination. While CMA remains valuable as a complementary test (detecting large CNVs, aneuploidy, UPD), it alone would miss the majority of diagnoses in this clinical context. A panel restricted to epilepsy genes would miss the many non-channel causes of GDD with seizures, and exome coverage of those genes is now comparable to panel coverage.
2. Two published studies report WES diagnostic yields for intellectual disability: Study A reports 45% yield, while Study B reports 20% yield. Both used trio WES with comparable bioinformatics. What is the most likely explanation for the difference?
Cohort selection is the single largest driver of yield variation across published genomic studies. Referral-centre or specialty-clinic cohorts are enriched for complex, multi-system or syndromic disease, so more of them have identifiable Mendelian conditions; broader cohorts that include milder, less specific phenotypes yield less with the same test. In the Clark 2018 meta-regression, larger studies (typically broad laboratory referral series) reported lower yields. Before citing any yield number, always ask: 'Who was in the denominator?' Severity, syndromic features, younger age at onset and comorbid epilepsy all raise yield; platform and bioinformatics differences contribute less.
3. A teenager with progressive ataxia and sensory neuropathy has whole exome sequencing that returns negative. Which critical next step should the clinician consider?
This is one of the most important blind spots in clinical neurogenetics. Common hereditary ataxias — Friedreich ataxia (FXN GAA repeat), the CAG spinocerebellar ataxias, SCA27B (FGF14 GAA) and CANVAS (RFC1 AAGGG) — are caused by repeat expansions. Standard exome sequencing does not size these expansions, and short-read genome sequencing detects them only variably. Dedicated testing (repeat-primed PCR, Southern blot or long-read sequencing) is required. In a teenager, progressive ataxia with sensory neuropathy points first to Friedreich ataxia; CANVAS (RFC1) causes a similar pattern but usually begins in later adulthood (Cortese 2019). A negative exome in ataxia should always prompt the question: 'Has dedicated repeat expansion testing been ordered?'
4. A colleague argues that chromosomal microarray (CMA) is obsolete now that whole exome sequencing is widely available. Which statement best counters this argument?
CMA is not obsolete: it provides information that exome does not. CMA detects copy number variants with higher sensitivity than exome-based CNV calling, and SNP-array CMA detects uniparental isodisomy and runs of homozygosity, relevant to imprinting disorders (Angelman, Prader-Willi) and to recessive disease in families with parental relatedness. CMA yields ~10% across children with suspected genetic disease (Clark 2018), 15–20% in DD/ID, autism and congenital anomalies (Miller 2010) and 14% in infantile spasms (Feng 2025). Many centres send CMA and exome together rather than sequentially.
5. Among the following pediatric neurogenetics phenotypes, which consistently shows the HIGHEST diagnostic yield from exome/genome sequencing?
White-matter disease in MRI-selected cohorts achieves among the highest diagnostic yields in clinical neurogenetics: in a dedicated leukodystrophy clinic, exome was diagnostic in 83% (34/41) and pathogenic variants were found in 89.6% overall (Zerem 2023); in patients undiagnosed after standard work-up, exome solved 59%, rising to 72% with reanalysis and genome (Schlüter 2022). MRI pattern recognition (hypomyelination vs. demyelination vs. cystic or vacuolating) narrows the differential before sequencing begins. By contrast, autism has the lowest NDD yield (pooled across all autism cohorts 17.1%, Stefanski 2021; lowest of all for autism without additional features, Tammimies 2015), and mild GDD yields less than severe, syndromic presentations. The more phenotypically specific and severe the presentation, the higher the yield.
6. Rapid whole genome sequencing in critically ill infants with a suspected genetic disorder diagnoses roughly 40–50%. Beyond identifying a diagnosis, what makes this one of the most impactful applications of genomic testing in pediatrics?
Infant intensive care provides the strongest evidence for the clinical utility of genomic testing. In Project Baby Bear, rapid genome sequencing diagnosed 40% of 184 acutely ill infants in a median of 3 days, changed medical care in 32%, and saved more than it cost ([Dimmock 2021](https://pubmed.ncbi.nlm.nih.gov/34089648/)). In GEMINI (400 infants), genome found a molecular diagnostic variant (pathogenic or VUS) in 49% vs 27% with a targeted neonatal panel, and care changed in 19% ([Maron 2023](https://pubmed.ncbi.nlm.nih.gov/37432431/)). Management changes include starting targeted therapy (e.g., ketogenic diet for GLUT1 deficiency), avoiding harmful medications, redirecting care when a devastating prognosis is confirmed, and surgical decisions. Speed matters because these decisions are made within days. Across pediatric genomic testing, Clark 2018 estimated clinical utility at 27% for genome, 17% for exome and 6% for CMA among all children tested.