Copy Number Variant Analysis
Copy number variant (CNV) analysis identifies genomic deletions and duplications ranging from approximately 50 base pairs to several megabases. This guide is for clinical researchers, bioinformaticians, and laboratory scientists who need a practical, evidence based framework for CNV detection from next generation sequencing (NGS) data. It draws on authoritative resources such as the NCBI Bookshelf 1 and EMBL EBI Training 2 to help you avoid common pitfalls and interpret results responsibly.
At a Glance
| Aspect | Key Points |
|---|---|
| Definition | Detection of structural variants (deletions, duplications) with size from ~50 bp to multi Mb |
| Common Platforms | Whole genome sequencing (WGS), whole exome sequencing (WES), targeted panels, CNV seq |
| Key Tools | GATK gCNV, Canvas, CNVnator, PennCNV, ExomeDepth, Bioconductor packages |
| Prerequisites | Paired end reads, consistent coverage (ideally 30x WGS or 100x WES), high quality alignment |
| Critical Quality Metrics | Read depth uniformity, GC bias, mappability, read pair discordance |
| Validation Need | Orthogonal method (MLPA, ddPCR, array CGH) for all clinical findings |
Core Concepts and Decision Points
CNV analysis relies primarily on read depth, split reads, or paired end approaches. The Galaxy Training Network 3 provides modular tutorials that illustrate the three main detection signals. Your choice of method depends on read length, coverage, and fragment size.
Decision Criteria
Sequencing type: WGS is superior for detecting copy neutral variants and smaller CNVs. WES has uneven capture efficiency and requires specialized tools like ExomeDepth that model exon depth variation. The NCBI Sequence Read Archive 5 contains many datasets that can help you benchmark your pipeline against known CNVs.
Coverage depth: For germline CNV detection, aim for at least 0.1x coverage depth of coverage for WGS, for clinical grade detection of exonic CNVs, 100x WES is typical. Low coverage produces high false positive rates.
Library type: PCR free libraries reduce GC bias and improve depth uniformity. When using PCR amplified libraries, apply correction steps available in Bioconductor 4.
Tool selection: Statistical models differ. For example, GATK gCNV uses a Gaussian mixture model that requires multiple samples for accurate denoising. Single sample tools like CNVnator rely on local mean depth and can work with one sample but have lower sensitivity for small events.
Practical Workflow
A robust CNV analysis pipeline follows these sequential steps. Each step includes a source for further guidance.
1. Read Quality Control and Alignment
Start with FASTQ data and check for adapter contamination, base quality, and duplication rates. Use tools like FastQC and apply adapter trimming if needed. Align paired end reads to the reference genome (GRCh38 preferred) with BWA MEM or similar. Report alignment metrics: insert size distribution, mapping rate, and percentage of properly paired reads. The EMBL EBI Training 2 offers a detailed module on read mapping quality.
2. Depth of Coverage Calculation
For read depth based methods, compute per base or per bin depth across the genome. For exome data, sum reads per targeted region. Be aware that repetitive regions and low mappability areas produce unreliable depth. Filter out regions with abnormal GC content or extreme read depth.
3. CNV Calling
Run your chosen caller with appropriate parameters. For cohort analysis, provide at least 20 reference samples to model local depth distribution. For single sample analysis, use a tool that normalizes against a control group or uses internal statistics. The Galaxy Training Network 3 includes a step by step tutorial on using CNVnator for WGS data.
4. Filtering and Quality Control
Filter callset using:
- Confidence score threshold (e.g., quality score > 10)
- Minimum number of supporting reads
- Reciprocal overlap with known segmental duplications or centromeres
- Reproducibility across replicates if available
The NCBI Bookshelf 1 has a chapter on guidelines for CNV interpretation that includes filtering criteria for clinical use.
5. Functional Annotation and Prioritization
Annotate CNVs with affected genes, OMIM morbid genes, and known pathogenic regions. Prioritize rare events (population frequency < 0.1%), large events (> 1 Mb), and events overlapping known dosage sensitive genes. Several CNVs in neurodevelopmental disorders demonstrate the clinical relevance of this step, as reported in a study of 1,252 children with neurodevelopmental disorders 6.
6. Orthogonal Validation
Every CNV intended for clinical reporting must be validated by an independent method such as MLPA, qPCR, ddPCR, or CMA. Long read sequencing can resolve complex CNVs, as shown in the integrated analysis of STRC variants in hereditary hearing impairment 10. Validation is not optional for diagnostic use.
Quality Checks and Common Mistakes
Ignoring GC bias. GC bias is the most common confounder in read depth CNV analysis. Always compute GC content per bin and apply correction. Failure to do so creates spurious calls in GC rich and GC poor regions.
Overfiltering small events. Many pipelines discard CNVs smaller than 1 kb by default. This is too aggressive for exome data where exon sized deletions (100 500 bp) are clinically relevant. Adjust your size filter to the resolution of your assay and tool.
Using default parameters without checking model assumptions. For example, GATK gCNV assumes the number of samples is sufficient to train the model (minimum 20). Using fewer samples inflates false positive rates.
Relying solely on computational scores. No single algorithm is perfect. A CNV with borderline quality but high biological relevance (e.g., deletion covering a known recessive disease gene) should be manually reviewed in IGV and validated. The case report on generalized nevus lipomatosus cutaneous superficialis 8 illustrates how comprehensive genetic analysis combined with clinical correlation is essential.
Neglecting sex chromosome analysis. Male and female samples differ in X and Y coverage. Many tools treat sex chromosomes incorrectly unless explicitly told. Use a tool that models chromosomal ploidy.
Limits of Interpretation
CNV analysis has inherent resolution, sensitivity, and specificity limits. Read depth methods cannot detect balanced translocations or inversions. Paired end analysis can detect these but has lower breakpoint precision. The lower size limit depends on coverage: at 30x whole genome coverage, events as small as 1 kb can be called with reasonable sensitivity, but smaller events require higher coverage or specialized tools.
False positive rates are higher in segmentally duplicated regions, repeat rich loci, and centromeric/telomeric areas. The NCBI Sequence Read Archive 5 hosts many control datasets that can help you estimate your platform specific false positive rate.
Uncertainty also arises from technical factors: library complexity, sequencing errors, and inconsistent coverage across samples. The study on CNV seq in a Vietnamese cohort with miscarriages 9 demonstrates both the power and limitations of CNV detection in low quality samples.
Clinical interpretation requires careful variant classification. A CNV overlapping a gene may not be pathogenic if it is common in population databases or if it does not disrupt coding sequence. The HNF1B related disease review 7 highlights how a single CNV can have variable penetrance and expressivity.
Do not interpret CNVs in isolation. Integration with other data types (SNVs, methylation, RNA expression) improves diagnostic yield. The ctDNA study 11 shows how combining mutation status with copy number can guide oncology decisions, but such approaches remain investigational for liquid biopsy.
Frequently Asked Questions
What is the minimum coverage needed for CNV detection?
For whole genome sequencing, at least 0.1x coverage depth is required for CNVs larger than 1 Mb. For exome sequencing, 80 to 100x on target is typical for detecting exon level events. Low coverage projects (e.g., 0.1x) can only detect very large events.
Can I use single end reads for CNV analysis?
Single end reads are not recommended because they lack paired end information for split read and discordant pair methods. Read depth only approaches can work but are less accurate and cannot determine breakpoint precision. Always use paired end data if possible.
Why do my CNV calls differ between tools?
Each tool uses different statistical models and assumptions. For example, GATK gCNV models read depth as a mixture of Gaussians, while CNVnator uses a mean shift algorithm. Concordance between two independent callers increases confidence. Always validate with an orthogonal method.
How do I handle noisy exome data?
Exome data is inherently noisier than whole genome data. Use tools designed for exome (e.g., ExomeDepth, XHMM). Normalize against a panel of reference samples. Filter out exons with extreme GC content or low mappability. Validate all clinical calls with MLPA or long read sequencing, as demonstrated for STRC 10.
References and Further Reading
- NCBI Bookshelf: Guidelines for the interpretation of sequence variants and copy number variants 1
- EMBL EBI Training: CNV detection from NGS data module 2
- Galaxy Training Network: Copy number variation detection tutorials 3
- Bioconductor: CNV analysis packages (exomeCopy, HMMcopy, CNAnorm) 4
- NCBI Sequence Read Archive: Public sequencing datasets for benchmarking 5
- Craniofacial features and pathogenic variants in 1,252 children with neurodevelopmental disorders 6
- HNF1B related disease: developmental origins, molecular mechanisms, and multisystem clinical challenges 7
- Generalized Nevus Lipomatosus Cutaneous Superficialis: a case report with comprehensive genetic analysis 8
- Genetic etiology of miscarriage in a Vietnamese cohort using CNV seq and exome sequencing 9
- Integrated analysis of STRC variants in hereditary hearing impairment using long read sequencing with MLPA validation 10
- ESR1 and PIK3CA circulating tumor DNA mutation status as predictive biomarkers in metastatic breast cancer 11