Structural-Variant Discovery Algorithms

Inferring rearranged chromosomes from discordant sequencing evidence

Bioinformatics

Algorithms

Genomics

Structural Variation

How structural-variant callers combine anomalous read pairs, split reads, depth shifts, and assemblies to locate and genotype genomic rearrangements.

Author

Jiahong Sun

Published

August 24, 2026

Modified

August 24, 2026

1 Topic at a Glance

One-sentence summary: Structural-variant discovery finds genomic breakpoints by clustering sequence reads whose placement, orientation, span, or sequence disagrees with a linear reference genome.

Item	Description
Biological question	Has a segment been deleted, duplicated, inserted, inverted, or rearranged, and where are its breakpoints?
Computational problem	Distinguish coordinated breakpoint evidence from mapping artifacts in a large, repetitive reference genome
Method type	Evidence clustering, probabilistic integration, local assembly, and genotyping
Input	Reference-aligned short or long reads; optionally a matched normal, assembly, or cohort evidence
Output	SV records with type, breakpoint intervals, orientation, support, genotype/allele fraction, and filters
Prerequisites	Read mapping and seed chaining, mapping quality estimation, copy-number segmentation algorithms

2 One rearrangement, several observable signatures

The reference describes one possible chromosome sequence. A sample deletion creates a new adjacency between the reference bases on either side of the removed segment. No short read need span the entire deletion: different read patterns provide partial evidence for that new adjacency.

Signal	Observation in an aligned library	Strength	Limitation
Discordant pair	Mates map too far apart, too close, or in an unexpected orientation	Detects events larger than a read	Breakpoints are interval-sized, set by insert distribution
Split read	One read aligns in pieces to distant reference positions	Can resolve a breakpoint near base precision	Needs an alignable sequence across the junction
Read depth	Coverage falls or rises over an interval	Strong for unbalanced CNVs	Cannot locate balanced inversions/translocations reliably
Local assembly / contig	Assembled sequence crosses a novel adjacency	Resolves complex sequence and insertions	Coverage, repeats, and computation can limit assembly

Integrated callers are powerful because these signals fail differently. A depth decrease supports a deletion’s dosage; discordant pairs bracket it; split reads can refine its ends. Conversely, an inversion is usually copy-neutral, so depth may be uninformative while orientation anomalies and split reads are decisive.

3 From anomalous alignments to a call

3.1 Candidate generation and clustering

The caller first removes or downweights dubious alignments, then converts each remaining observation into a breakpoint interval: the range of genomic coordinates consistent with its alignment. It groups observations whose intervals and orientations agree. For paired-end reads, the expected insert-size distribution makes a pair’s evidence inherently broad; split reads often narrow the range.

sorted, duplicate-marked alignments + reference
       -> discordant-pair, split-read, depth, or assembly evidence
       -> cluster compatible breakpoint intervals and orientations
       -> classify the proposed adjacency / SV type
       -> genotype, score, filter, and write VCF or BCF

Some frameworks represent each signal as a probability distribution over the two breakpoints and combine compatible distributions. This is preferable to treating all supporting reads as equally precise or independent. A call’s quality should also consider the strongest alternative placement, local repeat structure, strand/orientation balance, and matched-normal evidence where relevant.

3.2 A deletion toy example

Assume a paired-end library normally has inward-facing mates about 350 bp apart. Near a true 1 kb deletion relative to the reference, several inward-facing pairs map about 1,350 bp apart, with left mates upstream and right mates downstream of the missing interval. Their different start sites form a cluster, so a caller proposes a deletion bracketed by the cluster. If split reads cross the two novel junctions, their clipped portions map across the gap and can refine the breakpoint. A coverage drop across the same interval adds support.

The evidence is not proof by itself: repeats can produce the same pattern when reads map to the wrong copy. A robust pipeline records support and filters, rather than translating every discordant pair into a high-confidence variant.

4 Calling, genotyping, and representation

Discovery proposes candidate breakpoints; genotyping asks whether the sample supports reference and/or alternate adjacency. In a diploid germline analysis, reads may be compared under 0/0, 0/1, and 1/1-like hypotheses, much like variant-calling likelihood models, although independence and coverage assumptions are even more fragile for SVs. Somatic analysis additionally needs normal contamination, tumor purity, subclonality, and matched-normal artifact control.

VCF can describe simple events with SVTYPE, END, SVLEN, confidence intervals such as CIPOS/CIEND, and supporting-read fields. Translocations and other junctions may use paired breakend records (BND). These coordinates and symbols are a representation of a hypothesis, not an assembled chromosome sequence; preserve the caller’s filters and reference build.

5 Failure modes and validation

Failure mode	Why it creates error	Useful response
Segmental duplications and repeats	Reads have multiple plausible loci	Use mapping quality, repeat annotations, and long reads or assembly for resolution
Library artifacts / duplicates	A few molecules yield apparently repeated anomalous pairs	Mark duplicates; inspect unique molecule support and orientation
Reference mismatch	Sample sequence absent from or divergent from reference	Use appropriate reference/decoys; interpret population-specific regions carefully
Low allele fraction	Few supporting reads compete with noise	Integrate signals, use matched normal, and report sensitivity limits
Complex rearrangement	One simple SV label cannot represent the event	Use local or whole-genome assembly and examine breakpoint graphs

Benchmark by SV class, size, genomic context, and technology. A deletion caller can have excellent performance on large unique-region deletions yet be weak for insertions or tandem repeats. Orthogonal evidence–long reads, assemblies, PCR, optical mapping, or a curated truth set–is especially valuable for clinically or biologically consequential calls.

6 Connections and method choice

Short-read mapping creates the anchors from which SV evidence is inferred, so read mapping and seed chaining and mapping confidence directly affect discovery. Copy-number segmentation algorithms supplies dosage evidence for deletions and duplications but cannot by itself establish an adjacency. Haplotype-aware evaluation can use haplotype assembly and phasing to distinguish alleles and avoid double-counting a complex event.

Data / goal	Prefer	Reason
Short-read germline WGS	Integrated pair + split-read + depth caller	Complementary signals cover common simple SVs
Somatic tumor/normal WGS	Matched-pair caller with allele-fraction modeling	Helps separate germline variation and artifacts
Novel insertions or repetitive loci	Long reads and local/whole-genome assembly	Direct sequence can resolve junctions absent from the reference
Copy-number burden	Depth plus allele-specific segmentation	Measures dosage even when exact junction is unresolved

7 Take-home Messages

Structural variants are new genomic adjacencies or dosage changes inferred indirectly from alignments and sequence.
Discordant pairs, split reads, depth, and assembly offer complementary resolution and sensitivity.
Repeats, ambiguous mapping, library artifacts, and sample mixture make support counts insufficient without context-aware scoring and validation.
A reported SV type and coordinate are model-based hypotheses; retain confidence intervals, filters, reference build, and supporting evidence.

8 References

Rausch, T. et al. 2012. DELLY: structural variant discovery by integrated paired-end and split-read analysis. Bioinformatics 28:i333–i339.
Layer, R. M., Chiang, C., Quinlan, A. R., & Hall, I. M. 2014. LUMPY: a probabilistic framework for structural variant discovery. Genome Biology 15:R84.
Sedlazeck, F. J. et al. 2018. Accurate detection of complex structural variations using single-molecule sequencing. Nature Methods 15:461–468.
DELLY documentation: official usage guidance for integrated SV and CNV discovery.

--- title: "Structural-Variant Discovery Algorithms" subtitle: "Inferring rearranged chromosomes from discordant sequencing evidence" description: "How structural-variant callers combine anomalous read pairs, split reads, depth shifts, and assemblies to locate and genotype genomic rearrangements." author: "Jiahong Sun" date: "2026-08-24" date-modified: "2026-08-24" categories: [Bioinformatics, Algorithms, Genomics, Structural Variation] toc: true number-sections: true code-fold: show draft: false --- # Topic at a Glance > **One-sentence summary:** Structural-variant discovery finds genomic breakpoints by clustering sequence reads whose placement, orientation, span, or sequence disagrees with a linear reference genome. | Item | Description | |---|---| | **Biological question** | Has a segment been deleted, duplicated, inserted, inverted, or rearranged, and where are its breakpoints? | | **Computational problem** | Distinguish coordinated breakpoint evidence from mapping artifacts in a large, repetitive reference genome | | **Method type** | Evidence clustering, probabilistic integration, local assembly, and genotyping | | **Input** | Reference-aligned short or long reads; optionally a matched normal, assembly, or cohort evidence | | **Output** | SV records with type, breakpoint intervals, orientation, support, genotype/allele fraction, and filters | | **Prerequisites** | [Read mapping and seed chaining](read-mapping-and-seed-chaining.qmd), [mapping quality estimation](mapping-quality-estimation.qmd), [copy-number segmentation algorithms](copy-number-segmentation-algorithms.qmd) | # One rearrangement, several observable signatures The reference describes one possible chromosome sequence. A sample deletion creates a new adjacency between the reference bases on either side of the removed segment. No short read need span the entire deletion: different read patterns provide partial evidence for that new adjacency. | Signal | Observation in an aligned library | Strength | Limitation | |---|---|---|---| | Discordant pair | Mates map too far apart, too close, or in an unexpected orientation | Detects events larger than a read | Breakpoints are interval-sized, set by insert distribution | | Split read | One read aligns in pieces to distant reference positions | Can resolve a breakpoint near base precision | Needs an alignable sequence across the junction | | Read depth | Coverage falls or rises over an interval | Strong for unbalanced CNVs | Cannot locate balanced inversions/translocations reliably | | Local assembly / contig | Assembled sequence crosses a novel adjacency | Resolves complex sequence and insertions | Coverage, repeats, and computation can limit assembly | Integrated callers are powerful because these signals fail differently. A depth decrease supports a deletion's dosage; discordant pairs bracket it; split reads can refine its ends. Conversely, an inversion is usually copy-neutral, so depth may be uninformative while orientation anomalies and split reads are decisive. # From anomalous alignments to a call ## Candidate generation and clustering The caller first removes or downweights dubious alignments, then converts each remaining observation into a *breakpoint interval*: the range of genomic coordinates consistent with its alignment. It groups observations whose intervals and orientations agree. For paired-end reads, the expected insert-size distribution makes a pair's evidence inherently broad; split reads often narrow the range. ```text sorted, duplicate-marked alignments + reference -> discordant-pair, split-read, depth, or assembly evidence -> cluster compatible breakpoint intervals and orientations -> classify the proposed adjacency / SV type -> genotype, score, filter, and write VCF or BCF ``` Some frameworks represent each signal as a probability distribution over the two breakpoints and combine compatible distributions. This is preferable to treating all supporting reads as equally precise or independent. A call's quality should also consider the strongest alternative placement, local repeat structure, strand/orientation balance, and matched-normal evidence where relevant. ## A deletion toy example Assume a paired-end library normally has inward-facing mates about 350 bp apart. Near a true 1 kb deletion relative to the reference, several inward-facing pairs map about 1,350 bp apart, with left mates upstream and right mates downstream of the missing interval. Their different start sites form a cluster, so a caller proposes a deletion bracketed by the cluster. If split reads cross the two novel junctions, their clipped portions map across the gap and can refine the breakpoint. A coverage drop across the same interval adds support. The evidence is not proof by itself: repeats can produce the same pattern when reads map to the wrong copy. A robust pipeline records support and filters, rather than translating every discordant pair into a high-confidence variant. # Calling, genotyping, and representation Discovery proposes candidate breakpoints; genotyping asks whether the sample supports reference and/or alternate adjacency. In a diploid germline analysis, reads may be compared under 0/0, 0/1, and 1/1-like hypotheses, much like [variant-calling likelihood models](variant-calling-likelihood-models.qmd), although independence and coverage assumptions are even more fragile for SVs. Somatic analysis additionally needs normal contamination, tumor purity, subclonality, and matched-normal artifact control. VCF can describe simple events with `SVTYPE`, `END`, `SVLEN`, confidence intervals such as `CIPOS`/`CIEND`, and supporting-read fields. Translocations and other junctions may use paired breakend records (`BND`). These coordinates and symbols are a representation of a hypothesis, not an assembled chromosome sequence; preserve the caller's filters and reference build. # Failure modes and validation | Failure mode | Why it creates error | Useful response | |---|---|---| | Segmental duplications and repeats | Reads have multiple plausible loci | Use [mapping quality](mapping-quality-estimation.qmd), repeat annotations, and long reads or assembly for resolution | | Library artifacts / duplicates | A few molecules yield apparently repeated anomalous pairs | Mark duplicates; inspect unique molecule support and orientation | | Reference mismatch | Sample sequence absent from or divergent from reference | Use appropriate reference/decoys; interpret population-specific regions carefully | | Low allele fraction | Few supporting reads compete with noise | Integrate signals, use matched normal, and report sensitivity limits | | Complex rearrangement | One simple SV label cannot represent the event | Use local or whole-genome assembly and examine breakpoint graphs | Benchmark by SV class, size, genomic context, and technology. A deletion caller can have excellent performance on large unique-region deletions yet be weak for insertions or tandem repeats. Orthogonal evidence--long reads, assemblies, PCR, optical mapping, or a curated truth set--is especially valuable for clinically or biologically consequential calls. # Connections and method choice Short-read mapping creates the anchors from which SV evidence is inferred, so [read mapping and seed chaining](read-mapping-and-seed-chaining.qmd) and mapping confidence directly affect discovery. [Copy-number segmentation algorithms](copy-number-segmentation-algorithms.qmd) supplies dosage evidence for deletions and duplications but cannot by itself establish an adjacency. Haplotype-aware evaluation can use [haplotype assembly and phasing](haplotype-assembly-and-phasing.qmd) to distinguish alleles and avoid double-counting a complex event. | Data / goal | Prefer | Reason | |---|---|---| | Short-read germline WGS | Integrated pair + split-read + depth caller | Complementary signals cover common simple SVs | | Somatic tumor/normal WGS | Matched-pair caller with allele-fraction modeling | Helps separate germline variation and artifacts | | Novel insertions or repetitive loci | Long reads and local/whole-genome assembly | Direct sequence can resolve junctions absent from the reference | | Copy-number burden | Depth plus allele-specific segmentation | Measures dosage even when exact junction is unresolved | # Take-home Messages 1. Structural variants are new genomic adjacencies or dosage changes inferred indirectly from alignments and sequence. 2. Discordant pairs, split reads, depth, and assembly offer complementary resolution and sensitivity. 3. Repeats, ambiguous mapping, library artifacts, and sample mixture make support counts insufficient without context-aware scoring and validation. 4. A reported SV type and coordinate are model-based hypotheses; retain confidence intervals, filters, reference build, and supporting evidence. # References - Rausch, T. et al. 2012. [DELLY: structural variant discovery by integrated paired-end and split-read analysis](https://doi.org/10.1093/bioinformatics/bts378). *Bioinformatics* 28:i333--i339. - Layer, R. M., Chiang, C., Quinlan, A. R., & Hall, I. M. 2014. [LUMPY: a probabilistic framework for structural variant discovery](https://doi.org/10.1186/gb-2014-15-6-r84). *Genome Biology* 15:R84. - Sedlazeck, F. J. et al. 2018. [Accurate detection of complex structural variations using single-molecule sequencing](https://doi.org/10.1038/s41592-018-0001-7). *Nature Methods* 15:461--468. - [DELLY documentation](https://github.com/dellytools/delly): official usage guidance for integrated SV and CNV discovery.