Copy Number Variation (CNV)

CNVs span a wide range of sizes, from 1 kb to several megabases, distinguishing them from smaller insertions/deletions (indels). Their structural complexity includes:

• Simple deletions or duplications: A single contiguous segment is lost or gained.
• Complex rearrangements: Events involving multiple breakpoints, inversions, or triplications nested within a single locus.
• Mobile element insertions (MEIs): The insertion of transposable elements like Alu or LINE-1 sequences, which can create variable copy numbers. Accurately simulating this size spectrum and structural diversity is a primary challenge for generative models like GANs and VAEs.

CNVs arise from distinct DNA replication and repair errors. Synthetic data generators must model the sequence signatures left by these mechanisms to be biologically plausible:

• Non-allelic homologous recombination (NAHR): Mediated by flanking low-copy repeats (LCRs), leading to recurrent CNVs with clustered breakpoints.
• Non-homologous end joining (NHEJ): An error-prone repair mechanism that produces non-recurrent CNVs with unique breakpoints and often small insertions or deletions at the junction.
• Fork stalling and template switching (FoSTeS): A replication-based mechanism that can generate complex, discontinuous copy number changes.

The primary biological effect of a CNV is altering the gene dosage—the number of functional copies of a gene. This can lead to disease through several mechanisms:

• Dosage sensitivity: Genes whose function depends on precise protein levels (e.g., transcription factors) are disrupted by a 1.5x or 0.5x change in copy number.
• Gene disruption: A breakpoint can fall within a gene, truncating the protein and causing a loss of function, independent of copy number change.
• Position effects: A duplication can place a gene near new regulatory elements, altering its expression pattern. Synthetic data must preserve the correlation between CNV location and its predicted effect on gene expression.

CNVs exist on a spectrum from common, benign polymorphisms to rare, highly penetrant pathogenic variants. Key concepts include:

• Common CNVs: Often small, gene-poor, and found in >1% of the population. They contribute to normal human diversity.
• Rare CNVs: Typically larger, gene-rich, and found in <0.1% of individuals. They are statistically enriched in cohorts with neurodevelopmental disorders like autism and schizophrenia.
• Incomplete penetrance: A pathogenic CNV may not cause disease in all carriers, a phenomenon that synthetic cohort simulations must model by incorporating genetic background and environmental modifiers.

CNVs leave distinct computational footprints in sequencing data that synthetic read generators must replicate to benchmark bioinformatics tools:

• Read Depth (RD): A deletion shows a consistent decrease in read coverage, while a duplication shows an increase. This is the primary signal for detecting large CNVs.
• Split Reads (SR): Reads that align to the genome in two disjoint parts directly span a breakpoint, providing single-nucleotide resolution.
• Discordant Read Pairs (DP): Paired-end reads that map with an abnormal distance or orientation indicate a deletion, duplication, or inversion.
• De novo assembly: Assembling reads into contigs can resolve the full structure of a complex CNV without relying on a reference genome.

For synthetic genomic data to be useful, it must realistically simulate CNVs. This involves:

• Preserving allele frequency spectra: The generator must not create an unrealistic excess of rare, large, gene-damaging CNVs.
• Modeling linkage disequilibrium (LD): CNVs must be phased with nearby SNPs and other variants to create realistic haplotypes.
• Simulating technical artifacts: Synthetic reads should include the systematic biases in read depth caused by GC content and mappability that complicate real CNV detection.
• Benchmarking via TSTR: A model trained on synthetic data with known CNV ground truth must successfully identify CNVs in real sequencing data, validating the generator's utility.

The proportion of sequencing reads supporting a genetic variant at a specific locus. In the context of CNVs, shifts in allele frequency across a region signal a duplication or deletion event. Synthetic data generators must accurately model these frequency distributions to simulate realistic cohort-level copy number polymorphisms, ensuring that downstream variant calling pipelines are tested against biologically plausible allele dosage scenarios.

The non-random association of alleles at different genomic loci. CNVs often disrupt or create novel linkage patterns. Synthetic genomic models must preserve these complex correlation structures to maintain realistic haplotype architectures. Failure to model linkage disequilibrium leads to synthetic chromosomes that lack the population-genetic constraints found in real human cohorts.

The computational process of determining which genetic variants—including large CNVs—are inherited together on the same parental chromosome. Synthetic data generators must accurately phase copy number states relative to surrounding single nucleotide polymorphisms to produce realistic diploid genomes. This is critical for simulating compound heterozygous scenarios where a deletion on one allele combines with a point mutation on the other.

A rigorous evaluation technique that trains a classifier to distinguish between real and synthetic genomic data. For CNV-aware generators, the classifier must fail to identify systematic differences in copy number distributions, segmental duplication patterns, and breakpoint characteristics. A generator is considered high-quality only when the discriminator performs no better than random chance, proving distributional parity.

An artificially generated Variant Call Format file containing simulated structural variants, including deletions, tandem duplications, and copy number gains. A high-fidelity Synthetic VCF must replicate the statistical properties of real population cohorts, including variant size distributions, breakpoint precision, and allele frequency spectra. These files serve as ground-truth benchmarks for validating CNV detection algorithms.

An evaluation paradigm where a predictive model is trained exclusively on synthetic genomic data and tested on real patient samples. For CNV research, a successful TSTR benchmark proves that synthetic copy number profiles contain sufficient biological signal to train a pathogenic variant classifier. TSTR is the gold standard for measuring the utility of synthetic data as a privacy-preserving substitute.