A copy number variant is, at root, a dosage problem. The reference human genome is diploid — two copies of every autosomal segment — and a CNV is any deviation from that count: a deletion drops a region to one copy (or zero, if homozygous), a duplication raises it to three. The clinical question is never simply is there a CNV but rather does changing the dose of the genes inside it matter. That reframing is the whole discipline of CNV interpretation.

Most genes tolerate having one copy knocked out or an extra copy added — the remaining allele, plus feedback regulation, keeps protein output close enough to normal. These genes are dosage-insensitive, and CNVs over them are usually benign. A minority of genes are dosage-sensitive: their phenotype depends on having exactly two functioning copies. Lose one and you get haploinsufficiency (50% of normal product is not enough); gain one and you get a triplosensitivity effect (150% is too much). Pathogenic CNVs are, almost by definition, the ones that sweep up a dosage-sensitive gene.

This is why size alone is a poor guide. A 2 Mb deletion across a gene desert can be benign, while a focused 300 kb deletion taking out a single haploinsufficient transcription factor is unambiguously pathogenic. CNVs span kilobases to whole chromosome arms, and the same locus can be benign as a duplication but pathogenic as a deletion (or vice versa), because haploinsufficiency and triplosensitivity are independent properties.

• Benign population CNVs are the genomic background noise of being human — common (often >1% frequency), frequently in gene-poor or highly polymorphic regions, and catalogued in the Database of Genomic Variants and gnomAD-SV.
• Pathogenic CNVs are rare, typically arose recently (often de novo), and overlap genes with established dosage sensitivity.

In neurodevelopmental practice this single variant class carries real weight: chromosomal microarray finds a diagnostic CNV in roughly 15–20% of children referred for unexplained developmental delay, intellectual disability, autism or multiple congenital anomalies Miller et al. 2010, which is why it became the first-tier test for these indications. Exome and genome sequencing now find more diagnoses overall, mostly sequence variants, and genome sequencing can call many CNVs as well.

Every CNV detection method ultimately answers the same question — how much DNA is here compared to a normal genome? — but they do it through different physics, and those differences explain both what each platform catches and what it silently misses.

Chromosomal microarray (CMA) is the workhorse. It became the recommended first-tier test for unexplained developmental delay, intellectual disability, autism, and multiple congenital anomalies because it raised the diagnostic yield several-fold over the karyotype it replaced (~15–20% vs ~3%) — the consensus that drove this shift in clinical practice Miller et al. 2010. There are two flavors. Array CGH competitively hybridizes patient DNA against a reference and reads the fluorescence ratio: more signal means a gain, less means a loss. SNP arrays instead genotype hundreds of thousands of polymorphic sites and infer copy number from allele intensity and genotype pattern. That second channel is the key clinical advantage: SNP arrays see copy-neutral loss of heterozygosity — long stretches where copy number is a normal 2 but the two copies are identical, either because both came from one parent as the same chromosome (isodisomic uniparental disomy) or because they descend from one shared ancestor (identity by descent). aCGH is blind to this because the dose is normal; only the genotype pattern reveals it. The finding is a clue rather than a diagnosis: UPD has to be confirmed by parental studies or methylation testing, and heterodisomic UPD, in which a child inherits both different homologs from one parent, shows normal heterozygosity and is invisible to the array. That matters clinically, because maternal UPD15 in Prader-Willi syndrome is mostly heterodisomic. Paternal isodisomy of chromosome 15, by contrast, appears as absence of heterozygosity along the whole of 15q with normal copy number and causes Angelman syndrome, as in a boy first referred for autism whose diagnosis methylation testing confirmed Hnoonual et al. 2021.

Understanding the mechanism also predicts the blind spots. CMA measures quantity, not arrangement, so it cannot see a balanced rearrangement — a reciprocal translocation or inversion has no net gain or loss of DNA, so the array reads normal even though a gene may be disrupted at the breakpoint. It also misses sequence-level variants, repeat expansions, and low-level mosaicism, because a small mutant fraction barely shifts the averaged signal.

Genome sequencing detects CNVs through three complementary signatures: read depth (fewer or more reads piling up over a deleted or duplicated region), discordant read pairs (mate pairs that map too far apart or in the wrong orientation), and split reads (single reads that straddle a breakpoint). Combining these gives finer resolution — down to hundreds of base pairs — and near base-pair breakpoint mapping, and in principle it can flag balanced events that CMA never could. The cost is bioinformatic: CNV calling from sequencing is genuinely hard, uneven coverage in GC-rich regions and segmental duplications degrades sensitivity exactly where many recurrent disease CNVs live, and callers disagree, so results need careful curation.

Before 2019, CNV interpretation was inconsistent between laboratories, and size thresholds largely drove which CNVs were reported at all; the 2011 ACMG guideline offered five categories but no quantitative scoring Kearney et al. 2011. The joint ACMG/ClinGen technical standards replaced that intuition with a quantitative, point-based scoring system, deletions and duplications scored on separate scoresheets Riggs et al. 2020. The logic mirrors the ACMG/AMP sequence-variant framework — accumulate evidence, sum it, land in one of five tiers (Pathogenic, Likely Pathogenic, VUS, Likely Benign, Benign) — but the evidence categories are tailored to dosage.

The core conceptual move is that points add up. Each line of evidence is worth a positive or negative value; a CNV reaching ≥0.99 points is Pathogenic, 0.90–0.98 is Likely Pathogenic, −0.89 to 0.89 is VUS, −0.90 to −0.98 is Likely Benign, and ≤−0.99 is Benign. Some evidence is decisive by itself: complete overlap of an established haploinsufficient gene or region scores 1.00, which reaches Pathogenic alone. Most CNVs, though, have no such anchor, and their classification emerges from converging evidence. De novo status illustrates the limit: it is worth at most 0.45 points, so on its own it can never carry a CNV to Pathogenic.

The scoresheet walks through five evidence sections:

• Section 1 — genomic content. Does the CNV even contain protein-coding genes or known functional elements? A region with no protein-coding genes or known functional elements scores −0.60; one that contains them scores 0 and evaluation continues. This is where the framework deliberately abandoned size cutoffs in favor of what's inside.
• Section 2 — overlap with established dosage-sensitive regions. Complete overlap of a ClinGen-curated haploinsufficient (for a deletion) or triplosensitive (for a duplication) gene or established region scores 1.00 on its own, enough for a Pathogenic call; partial overlap scores differently depending on whether the critical gene's coding sequence is disrupted, and a CNV contained within an established benign region scores −1.00.
• Section 3 — number of protein-coding genes. A crude proxy: for a deletion, 0–24 genes score 0, 25–34 score 0.45 and 35 or more score 0.90 (the duplication bins are 0–34, 35–49 and ≥50).
• Section 4 — literature/database evidence for this specific region and phenotype, including case-level data and segregation.
• *Section 5 — inheritance and de novo status in the patient at hand: a confirmed de novo* CNV scores 0.15–0.45 depending on how specific and consistent the phenotype is; a CNV inherited from an apparently unaffected parent scores −0.30 by default when the child's phenotype is specific and −0.15 when it is nonspecific.

The engine underneath Section 2 is ClinGen's curated dosage map. Each gene and region gets a haploinsufficiency (HI) score and a triplosensitivity (TS) score on a 0–3 scale — 0 no evidence, 1 little evidence, 2 emerging evidence, 3 'sufficient evidence that losing (or gaining) a single copy causes disease' — and only a score of 3 counts as established in Section 2. Two further codes mark a gene associated with an autosomal recessive phenotype (30) and one where dosage sensitivity is unlikely (40); a score of 40 carries no points of its own, since benign evidence enters the scoresheet through overlap with established benign CNVs (Section 2), common population variation (Section 4) and population data. Crucially, the standards insist on uncoupling the variant's intrinsic classification from its meaning for the individual: a CNV can be objectively Pathogenic yet, at a reduced-penetrance or imprinted locus, still not explain a given patient's presentation — two separate judgments that should not be blurred on the report.

Some deletions and duplications appear over and over, in unrelated patients, with nearly identical breakpoints and the same size every time. This recurrence is not coincidence — it is written into the architecture of the genome itself. These loci are flanked by low-copy repeats (segmental duplications): long blocks of DNA, often >95% identical, sitting on either side of the region.

The mechanism is non-allelic homologous recombination (NAHR). During meiosis, chromosomes pair up and the recombination machinery looks for matching sequence to align. The flanking repeats are so similar that the machinery can mistake the repeat on one side for the repeat on the other — aligning paralogous (non-allelic) copies instead of true alleles. When the repeats are directly oriented, a crossover in that misaligned configuration between chromatids excises the intervening segment from one product (deletion) and adds it to the reciprocal product (duplication); inverted repeats instead give inversions. NAHR within a single chromatid, which loops the interval out, produces a deletion with no reciprocal duplication, one reason deletions are found more often than duplications. Because the repeats define the alignment, the breakpoints — and therefore the CNV size — are stereotyped in most carriers, although nested deletions between internal repeats also occur (at 22q11.2, LCR22A–B and A–C). This is what makes these conditions genomic disorders: the disease is a property of where the repeats sit, not of any single mutation.

This explains several clinical observations at once. It explains why deletion and duplication of the same region often coexist as paired syndromes (they are reciprocal products of one event). It explains the stereotyped sizes at each locus, from about 600 kb at 16p11.2 (BP4–BP5) to about 3 Mb at 22q11.2. And it explains why these are among the most common identifiable causes of syndromic neurodevelopmental disease — the genomic architecture keeps regenerating the same CNVs in the population, generation after generation. Contrast this with non-recurrent CNVs, which have scattered, patient-specific breakpoints and arise through different routes (non-homologous end joining, or replication-based mechanisms such as FoSTeS/MMBIR), where the deleted region varies from case to case.

The classic neurogenetic genomic disorders — 22q11.2 deletion (DiGeorge/velocardiofacial; about 85% are the ~3 Mb LCR22A–D deletion), Williams-Beuren (7q11.23), the imprinted 15q11–q13 region underlying Angelman and Prader-Willi (BP1 or BP2 to BP3), and proximal 16p11.2 (BP4–BP5, about 600 kb; the deletion is associated with developmental delay, autism and obesity, the reciprocal duplication with microcephaly and lower weight) — all share this NAHR signature, which is why a microarray can recognize them at a glance from their reproducible coordinates. Their reciprocal duplications are real disorders too, usually milder and more variable: the 22q11.2 duplication is often inherited from a mildly affected or unaffected parent, and the 7q11.23 duplication is associated with speech delay, anxiety and autism. At 15q11–q13, imprinting decides the phenotype: in neurons UBE3A is expressed only from the maternal chromosome, while SNRPN and the SNORD116 cluster are expressed only from the paternal one, so a maternal deletion causes Angelman syndrome and a paternal deletion Prader-Willi syndrome. Because a parent carrying a typical 5–6 Mb deletion would be affected, these deletions are almost always de novo; an inherited Angelman deletion is usually a small one within UBE3A, carried silently by a mother on the chromosome she received from her father.

1p36 deletion syndrome, the most common terminal deletion, is the instructive exception: its breakpoints vary from patient to patient and it is not NAHR-mediated.

Some recurrent CNVs are susceptibility loci rather than syndromes. The 15q11.2 BP1–BP2 deletion, for example, had an estimated penetrance of about 10% for a combined phenotype of congenital anomalies, developmental delay or intellectual disability, against a background of about 5% Rosenfeld et al. 2013. A CNV of this kind can be correctly classified Pathogenic and still be present in a healthy parent, and still not be the whole explanation for the child.

A CNV report is a clinical instrument, and its job is to let a clinician act without re-deriving the lab's reasoning. Three things make that possible: unambiguous coordinates, an honest classification, and a clear statement of what the finding means for this patient.

Coordinates must be reproducible. Because CNV size and breakpoints are the whole story, the report has to anchor them to a specified genome build — the same deletion has different numerical coordinates on GRCh37 versus GRCh38, and mixing builds is a recurring source of error. Good reports give ISCN nomenclature, the linear coordinates, the build, and a summary of the gene content so the reader can judge dosage at a glance.

Inheritance is often the deciding evidence, and this is where the report's value is highest. Testing both parents (a 'trio') converts an uninterpretable VUS into something actionable surprisingly often: a confirmed de novo CNV earns points toward pathogenicity, because a benign CNV would usually be found sitting harmlessly in a parent, while a CNV inherited from an unaffected parent loses points. The asymmetry must be read with care, though — inheritance from an unaffected parent argues against pathogenicity only when the locus shows full penetrance and no imprinting. At reduced-penetrance loci, an unaffected carrier parent is fully compatible with a pathogenic call; at imprinted loci like 15q11–q13, the parent of origin, not mere inheritance, determines the phenotype.

The VUS must be communicated as genuinely uncertain — neither reassurance nor diagnosis. A VUS cannot drive medical decisions, and the report should say so plainly while offering the concrete next steps that can resolve it: parental testing, deeper phenotyping, and periodic reclassification as databases mature. Honest uncertainty prevents both false reassurance and overtreatment.

Finally, genome-wide assays produce incidental and secondary findings — a CNV unrelated to the referral question, sometimes with implications for an adult-onset condition or for relatives. Because microarray scans the whole genome at once, this possibility is foreseeable, so it belongs in pre-test counseling rather than arriving as a surprise on the report.