Structural Variant

A structural variant (SV) is a genomic alteration of about 50 base pairs or more that changes the arrangement of DNA: a segment deleted, duplicated, inserted, inverted, or moved to a new location. The 50 bp threshold is a convention drawn where short-read alignment stops representing a change as a simple edit to a single read and starts requiring evidence from read pairs, split reads, or depth.

The four classes people usually mean by “types of genetic variants” are SNVs (single base substitutions), small indels (under ~50 bp), structural variants, and aneuploidies or whole-chromosome changes. Copy number variants are a subset of SVs: the unbalanced ones, deletions and duplications, where the amount of DNA changes. Inversions and balanced translocations are structural variants that are not CNVs, because copy number stays the same while the arrangement does not. This distinction matters practically, since array-based and depth-based methods detect CNVs and are blind to balanced events.

How it works

Short-read callers infer SVs from three signals: discordant read pairs (insert size or orientation inconsistent with the reference), split reads (a single read aligning to two distant positions), and read depth (a drop or rise in coverage across an interval)¹. Each signal has a blind spot. Depth sees large CNVs but cannot localize breakpoints or detect inversions. Split reads give base-pair breakpoints but fail inside repeats where the split maps ambiguously.

Long reads change the arithmetic. A 15 kb HiFi read spanning both breakpoints resolves the event directly, including insertions, which short reads recover poorly because the inserted sequence is absent from the reference. Long-read assemblies of human genomes have catalogued tens of thousands of SVs per individual, the large majority of them insertions and deletions under a few kilobases, concentrated in tandem repeats and segmental duplications². A typical human genome carries thousands of SVs that collectively affect far more base pairs than all its SNVs combined³.

Consequence is not proportional to size. A 2 kb deletion inside an intron may do nothing. A deletion or inversion that removes a TAD boundary can put an enhancer in contact with a gene it never normally regulates, changing expression without touching a coding sequence⁴. This is why annotating SVs only against exons misses a whole category of effect.

In your own data

Your SV calls arrive as a VCF distinct from your small-variant VCF. Records look like:

chr7  117548000  DEL00231  N  <DEL>  .  PASS  SVTYPE=DEL;SVLEN=-4312;END=117552312;

The ALT is a symbolic allele (<DEL>, <DUP>, <INV>, <INS>, <BND>), and the coordinates come from INFO, not from the ALT string. Breakend records (BND) come in mate pairs linked by MATEID; a translocation is two BND lines, and a script that counts VCF rows will double-count it.

What we check first, in order:

1. Caller and read type. bcftools view -h your.sv.vcf.gz | grep '^##source'. Manta or Delly on short reads, Sniffles2 or pbsv on long reads. If it is short-read only, treat insertion calls as substantially incomplete.
2. Filter status. Drop everything that is not PASS before you look at anything. bcftools view -f PASS.
3. Population frequency. Intersect with gnomAD-SV and the Database of Genomic Variants, a curated catalogue of SVs observed in controls⁵. bedtools intersect -a calls.bed -b dgv.bed -f 0.5 -r with reciprocal overlap. Most of your calls will be common polymorphism.
4. What the remaining events touch. Overlap against a GENCODE exon BED, then against a TAD boundary track from a relevant cell type.

Common mistakes: comparing SV calls between two callers by exact coordinates (breakpoint estimates differ by tens to hundreds of bases; use reciprocal overlap), mixing GRCh38 and hs37d5 coordinates, and trusting calls that sit inside segmental duplications without visual inspection of the alignment in IGV.

Limitations

Most SVs you find will be unclassifiable. Variants of uncertain significance are common even in well-studied hereditary disease gene panels, and reclassification takes years⁶. That rate reflects missing population and functional data rather than hidden risk, and the reasonable default is to treat an uncertain variant as uncertain rather than as evidence of anything⁷. A VUS should not drive any medical decision, and interpreting one in a clinical context requires a genetics professional. Scaling functional assays and better population reference data are the two paths out, and neither is finished⁸.

Short-read SV calling also carries a real false-discovery rate in repetitive regions, and sensitivity for inversions and balanced events is poor across all short-read methods¹. If SVs matter to your question, sequence long.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.