Why Did My Structural Variant Caller Miss That Deletion? Troubleshooting False Negatives in SV Detection

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-21

Key Takeaways

Structural variant (SV) callers frequently generate false negatives for deletions due to insufficient sequencing coverage at breakpoints, leading to inadequate read depth or split read evidence.
Library insert size distribution mismatches with caller assumptions can cause genuine discordant read pairs from deletions to be missed or normal pairs to be misclassified.
Overly stringent parameter settings or filtering thresholds in SV callers can systematically exclude valid deletion signals, particularly those with moderate evidence support.
Alignment ambiguity in repetitive regions or poor mapping quality at breakpoints prevents the generation of reliable split read and discordant pair evidence, a primary cause of missed deletions.
The inherent limitations of short-read sequencing, such as inability to resolve repetitive regions and challenges with large insertions, contribute to a fundamental ceiling on SV detection sensitivity.

Structural variant (SV) calling from short-read sequencing data frequently produces false negatives, where real deletions and other large genomic alterations are absent from the final call set. When your variant caller misses a deletion, the cause typically falls into one of three categories: insufficient sequencing coverage at the breakpoint, library insert size that does not match the caller's assumptions, or parameter settings that filter out legitimate signals. This article provides a systematic diagnostic approach for laboratory professionals and researchers who need to determine why specific SVs were missed and how to adjust their workflows to recover those calls.

The problem of missed SVs is also an inconvenience. False negatives in structural variant detection have direct consequences for both research and clinical applications. In rare disease diagnosis, a missed deletion can mean a failed molecular diagnosis. In cancer genomics, an undetected structural rearrangement can lead to incorrect treatment stratification. Understanding the sources of false negatives requires examining the entire detection pipeline, from sequencing library preparation through alignment to variant calling and filtering.

The Scope of False Negative Problems in SV Detection

Structural variants include deletions, insertions, duplications, inversions, and translocations that typically exceed 50 base pairs in length. Unlike single nucleotide variants, which can be detected through simple base mismatches in aligned reads, structural variants require evidence from read depth, discordant read pairs, split reads, or assembly-based approaches. Each evidence type has distinct failure modes that can produce false negatives.

The challenge of SV detection is well documented in the literature. A 2021 study in Genome Biology examined structural variant detection across multiple sequencing centers and analysis pipelines using DNA from a family quartet. The researchers found that mapping methods contributed the largest source of analytical variability, followed by sequencing center effects and replicate differences. Notably, structural variants supported by only one center or one replicate often represented true positives, with approximately 47% of single-center calls and 45% of single-replicate calls overlapping with long-read sequencing results. This finding indicates that the false negative rate for SV calling is substantially higher than the false positive rate, and that many genuine variants are being missed by individual pipelines.

The clinical implications of missed SVs are equally significant. A 2024 study in Genome Medicine evaluated whether genome sequencing could replace existing workflows for germline genetic diagnosis in rare disease. The researchers analyzed 1,000 cases with 1,271 known clinically relevant variants and found that overall detection was 95%, but detection rates varied by variant category. Small variants were detected in 96% of cases, large variants in 93%, and other variants including structural variants and aneuploidies in only 87%. The true positive rates varied between workflows from 79% to 100%, with only 7 of 10 workflows being replaceable by genome sequencing. These results demonstrate that structural variant detection remains a weak point in clinical sequencing pipelines.

At a Glance: Common Causes of Missed Deletions

Cause Category	Primary Evidence Type Affected	Typical Symptom	First Diagnostic Step
Insufficient coverage	Read depth and split reads	Deletion present in long-read data but absent in short-read calls	Check mean coverage and coverage uniformity at the deletion locus
Insert size mismatch	Discordant read pairs	Caller reports low paired-end support for known deletion	Compare library insert size distribution to caller default parameters
Overly strict filtering	All evidence types	Deletion visible in manual inspection of BAM file but filtered from final VCF	Review filter flags and quality score thresholds in the VCF
Alignment ambiguity	Split reads	Breakpoints map to repetitive or low-complexity regions	Examine mapping quality at breakpoint-flanking regions
Caller parameter limits	All evidence types	Deletion size below caller minimum threshold	Verify minimum and maximum SV size parameters

Core Principles of SV Detection and False Negative Generation

Evidence Types and Their Failure Modes

Structural variant callers rely on multiple lines of evidence to identify deletions. Understanding each evidence type helps diagnose why a specific deletion was missed.

Read depth analysis detects deletions by identifying regions where the number of aligned reads drops below expected levels. A heterozygous deletion should produce approximately half the normal read depth, while a homozygous deletion should produce near-zero depth. Read depth methods fail when coverage is too low to distinguish a genuine depth reduction from stochastic variation. They also fail in regions with high GC content or other sequencing biases that naturally produce variable coverage.

Discordant read pair analysis identifies deletions by finding read pairs where the insert size is larger than expected or where the orientation is abnormal. For a deletion, reads that flank the deleted region will map with an insert size that exceeds the library fragment length. Discordant pair methods fail when the deletion is smaller than the variation in insert size, when the library has a broad insert size distribution, or when reads map to repetitive regions that create ambiguous placements.

Split read analysis detects deletions by identifying reads that span the breakpoint, with one portion of the read mapping to one side of the deletion and the other portion mapping to the other side. Split read methods provide breakpoint-level resolution but require sufficient coverage at the exact breakpoint position. They fail when coverage is low, when the breakpoint falls in a repetitive region, or when the read length is too short to uniquely map both sides of the breakpoint.

Assembly-based methods construct local assemblies from reads and compare the assembled contigs to the reference genome. These methods can detect larger and more complex SVs but require substantial computational resources and high coverage. They fail when coverage is insufficient for assembly or when repetitive regions prevent contiguous assembly.

The Role of Sequencing Coverage

Coverage is the most fundamental determinant of SV detection sensitivity. Each evidence type has minimum coverage requirements, and these requirements are higher for SV detection than for small variant detection.

For read depth methods, the ability to distinguish a heterozygous deletion from normal coverage variation depends on the number of reads in the region. With 30x coverage, a heterozygous deletion should produce approximately 15x coverage, but the variance in coverage across a genome means that some regions will naturally have 15x coverage without any deletion. Higher coverage reduces this variance and improves the signal-to-noise ratio.

For split read methods, the probability of observing a read that spans a breakpoint depends on the read length and the coverage. If the average coverage is 30x and reads are 150 base pairs, the expected number of reads spanning a particular breakpoint is related to the read length and the distance from the breakpoint to the nearest read start. Low coverage regions may have no reads spanning the breakpoint, making split read detection impossible.

The 2024 Genome Medicine study used 37x mean coverage for their genome sequencing workflow and still detected only 87% of structural variants and aneuploidies. This finding suggests that even moderate-to-high coverage does not guarantee complete SV detection, and that other factors such as library preparation and alignment methods contribute to false negatives.

Library Insert Size and Fragment Length Distribution

The insert size of the sequencing library is a critical parameter for discordant read pair detection. Most SV callers assume a particular insert size distribution, typically modeled as a normal distribution with a mean and standard deviation. Reads pairs with insert sizes exceeding the expected distribution by several standard deviations are flagged as discordant and used as evidence for deletions or other SVs.

If the actual library insert size differs from the caller's assumption, the caller may misclassify normal read pairs as discordant or fail to identify genuinely discordant pairs. For example, if the library has a mean insert size of 500 base pairs but the caller assumes 300 base pairs, many normal pairs will appear discordant, creating false positive signals. Conversely, if the library has a mean insert size of 300 base pairs but the caller assumes 500 base pairs, genuinely discordant pairs from a deletion may fall within the expected insert size range and be missed.

The insert size distribution also affects the minimum detectable deletion size. Discordant read pair methods can only detect deletions larger than the variation in insert size. If the insert size distribution has a standard deviation of 50 base pairs, deletions smaller than approximately 100 to 150 base pairs may be indistinguishable from normal insert size variation.

Alignment Methods and Their Impact on SV Detection

The choice of alignment algorithm has a substantial impact on SV detection sensitivity. The 2021 Genome Biology study found that mapping methods provided the major contribution to variability in SV detection across different analysis pipelines. This finding highlights that the aligner, not the variant caller, is often the primary determinant of whether a structural variant is detected.

Different aligners use different strategies for handling reads that span structural variant breakpoints. Some aligners clip reads at the breakpoint, producing split read alignments that SV callers can use. Other aligners may map the entire read to one side of the breakpoint with poor quality, or may fail to map the read entirely. The choice of aligner also affects mapping quality scores, which SV callers use to filter candidate variants.

Alignment to repetitive regions is particularly problematic. Reads that map to repetitive elements may have multiple equally good alignment positions, resulting in low mapping quality or random placement. SV callers typically filter out variants supported by low mapping quality reads, which can lead to false negatives for deletions in or near repetitive regions.

Practical Workflow for Diagnosing Missed Deletions

Step 1: Confirm the Deletion Is Real

Before troubleshooting why a variant caller missed a deletion, you must confirm that the deletion actually exists in the sample. This confirmation is essential because not all apparent false negatives are genuine. The deletion may be an artifact of the comparison method, a misassembly in the reference genome, or a variant that is present in the reference but absent in the sample.

Compare your short-read SV calls to an orthogonal data source. Long-read sequencing data from platforms such as Pacific Biosciences or Oxford Nanopore can provide independent confirmation of structural variants. The Genome in a Bottle Consortium has developed benchmark sets for germline large deletions and insertions that integrate multiple sequencing technologies and can be used to evaluate the sensitivity of your calling pipeline. The 2020 Nature Biotechnology benchmark paper describes a sequence-resolved benchmark set containing 12,745 isolated insertions and deletions of at least 50 base pairs, with Tier 1 regions covering 2.51 Gbp and containing 5,262 insertions and 4,095 deletions supported by at least one diploid assembly.

If long-read data is not available, examine the raw sequencing data manually. Load the BAM file in a genome browser and visually inspect the region of interest. Look for evidence of the deletion, including reduced read depth, discordant read pairs, and split reads. If none of these evidence types are present in the BAM file, the deletion may not be detectable with the available data.

Step 2: Check Coverage at the Deletion Locus

Coverage is the first parameter to examine when a deletion is missed. Calculate the mean coverage across the entire genome and then examine coverage specifically at the deletion locus and its flanking regions.

Use a coverage profiling tool to generate per-base or per-window coverage across the region of interest. Compare the coverage at the deletion locus to the genome-wide average. If the coverage at the deletion locus is substantially lower than the genome-wide average, the deletion may have been missed because there were not enough reads to provide statistical evidence for a copy number change.

For read depth based callers, the sensitivity for detecting a heterozygous deletion depends on the number of reads in the region. With 30x coverage, a 1 kilobase deletion would contain approximately 30 reads in a diploid sample, with 15 reads expected from the deleted allele and 15 from the non-deleted allele. The statistical power to detect this reduction depends on the variance in coverage, which is influenced by GC content, mappability, and other factors.

If coverage is low at the deletion locus, consider whether the low coverage is specific to this region or reflects a genome-wide problem. Regional low coverage may be caused by GC bias, repetitive sequence, or poor mappability. Genome-wide low coverage indicates a problem with sequencing depth or library preparation that affects the entire sample.

Step 3: Examine Library Insert Size Statistics

The insert size distribution is a critical parameter for discordant read pair detection. Most SV callers estimate insert size from the aligned reads, but this estimation can be biased if the alignment contains many discordant pairs or if the library has an unusual fragment size distribution.

Calculate the insert size distribution from the BAM file using a tool such as Picard CollectInsertSizeMetrics. Examine the mean, median, standard deviation, and the shape of the distribution. Compare these statistics to the values assumed by your SV caller.

If the insert size distribution is broader than expected, discordant read pair detection will have reduced sensitivity. A broad insert size distribution means that the range of normal insert sizes is larger, and deletions must be larger to produce insert sizes that fall outside the normal range. If the insert size distribution has multiple peaks, this may indicate a problem with library preparation, such as incomplete adapter ligation or template switching.

The insert size distribution also affects the minimum deletion size detectable by discordant read pair methods. If the insert size standard deviation is 50 base pairs and the caller uses a threshold of 5 standard deviations, the minimum detectable deletion size is approximately 250 base pairs. Smaller deletions will not produce discordant read pairs that exceed this threshold.

Step 4: Review Caller Parameters and Filter Settings

Each SV caller has parameters that control sensitivity and specificity. Review the parameters used in your analysis and compare them to the recommended settings for your data type and research question.

The minimum and maximum SV size parameters determine the range of deletion sizes that the caller will report. If the caller was configured with a minimum deletion size of 500 base pairs, deletions smaller than this threshold will not be reported regardless of the evidence. Check whether the missed deletion falls within the configured size range.

Quality score thresholds filter out candidate variants with low confidence. SV callers assign quality scores based on the number of supporting reads, the consistency of the evidence, and other factors. If the quality score threshold is too high, genuine deletions with moderate support will be filtered out. Review the distribution of quality scores for called deletions and compare the quality score of the missed deletion to the threshold.

Filter flags in the VCF file indicate why a variant was filtered out. Examine the FILTER column for the missed deletion. Common filter flags include low quality, insufficient supporting reads, strand bias, and mapping quality issues. Each filter flag points to a specific aspect of the evidence that failed to meet the caller's criteria.

Step 5: Evaluate Alignment Quality at the Breakpoints

The alignment of reads at the deletion breakpoints determines whether split read and discordant read evidence can be generated. Examine the alignment in the BAM file at the predicted breakpoint positions.

Check the mapping quality of reads that span or flank the breakpoints. Low mapping quality reads may be filtered out by the SV caller, even if they provide genuine evidence for the deletion. Reads that map to repetitive regions often have low mapping quality because they cannot be uniquely placed.

Examine whether reads are clipped at the breakpoint. Split read detection relies on reads that are clipped at the breakpoint, with one portion mapping to one side and the other portion mapping to the other side. If reads are not clipped at the breakpoint, the aligner may have mapped them entirely to one side with mismatches, or may have failed to map them at all.

Check whether the breakpoints fall in regions of low mappability. Some genomic regions cannot be uniquely mapped with short reads, and these regions are effectively invisible to short-read sequencing. Deletions with breakpoints in these regions will be missed regardless of coverage or caller parameters.

Step 6: Consider the Variant Caller Choice

Different SV callers use different algorithms and evidence types, and their sensitivity varies across different types of structural variants. If a deletion is consistently missed by one caller, try a different caller that uses a different evidence type or algorithm.

Some callers specialize in read depth based detection, while others focus on discordant read pairs or split reads. A deletion that is missed by a split read based caller may be detected by a read depth based caller, and vice versa. Running multiple callers and combining their results can improve sensitivity, although this approach also increases the number of false positives that must be filtered.

The 2021 Genome Biology study found that SV calling variability persists even in a genotyping approach, indicating that the underlying sequencing and preparation approaches have a substantial impact on detection. This finding suggests that changing the variant caller alone may not be sufficient if the sequencing data has inherent limitations.

Options and Tradeoffs in SV Detection Strategies

Increasing Sequencing Coverage

Increasing sequencing coverage is the most direct way to improve SV detection sensitivity, but it comes with substantial cost. Each increase in coverage requires additional sequencing reagents and computational resources.

For read depth based detection, the benefit of increased coverage is clear. Higher coverage reduces the variance in read depth, making it easier to distinguish genuine copy number changes from stochastic variation. For split read detection, higher coverage increases the probability that reads will span breakpoints.

The tradeoff is cost. Doubling coverage doubles the sequencing cost for the same number of samples. For large cohort studies, this cost increase may be prohibitive. An alternative is to use targeted coverage increases for regions of interest, although this approach requires prior knowledge of which regions to target.

Adjusting Library Preparation

Library preparation choices affect SV detection sensitivity in several ways. Fragment size, read length, and PCR amplification all influence the types of evidence that can be generated.

Longer fragments improve discordant read pair detection because they provide a larger window for detecting insert size abnormalities. However, longer fragments may reduce the efficiency of cluster generation on some sequencing platforms. The optimal fragment size depends on the sequencing platform and the SV caller being used.

Longer reads improve split read detection because they are more likely to span breakpoints and provide unique mapping across repetitive regions. Read length is determined by the sequencing platform and the number of cycles used. Some platforms offer options for longer reads at the cost of lower throughput.

PCR amplification introduces biases that can affect read depth uniformity. PCR-free library preparation methods reduce these biases and improve read depth based SV detection. The tradeoff is that PCR-free methods require more input DNA, which may not be available for all samples.

Using Multiple Callers and Ensemble Approaches

Running multiple SV callers and combining their results can improve sensitivity, but this approach requires careful consideration of how to handle discordant calls.

The simplest approach is to take the union of calls from multiple callers, which maximizes sensitivity but also increases false positives. A more sophisticated approach is to require support from multiple callers, which improves specificity but may miss variants that are only detected by one caller.

The 2021 Genome Biology study found that structural variants supported by only one center or replicate often represented true positives, with approximately 45% to 47% overlapping with long-read calls. This finding suggests that requiring support from multiple callers or replicates would miss a substantial number of genuine variants.

Ensemble methods that weight evidence from multiple callers based on their known performance characteristics can provide a balance between sensitivity and specificity. These methods require benchmark data to calibrate the weights, which may not be available for all sample types or variant classes.

Incorporating Long-Read Sequencing

Long-read sequencing platforms can detect structural variants that are missed by short-read sequencing, particularly in repetitive regions and for larger variants. However, long-read sequencing is more expensive and has lower throughput than short-read sequencing.

A hybrid approach uses short-read sequencing for genome-wide coverage and long-read sequencing for targeted regions or for validation of candidate variants. This approach can improve SV detection sensitivity while controlling costs.

The 2020 Nature Biotechnology benchmark paper demonstrates that integrating multiple sequencing technologies can produce a more complete SV call set. The benchmark set was created by integrating 19 sequence-resolved variant calling methods from diverse technologies, including short-read, linked-read, and long-read sequencing and optical mapping.

Records and Measurements for SV Detection Troubleshooting

Essential Metrics to Track

Maintain systematic records of the metrics that affect SV detection sensitivity. These records enable you to diagnose problems when a deletion is missed and to compare performance across samples and runs.

Coverage metrics should include mean coverage, coverage uniformity, and the fraction of the genome covered at various thresholds. The percentage of bases covered at 10x, 20x, and 30x provides a more complete picture than mean coverage alone, because mean coverage can be inflated by regions with very high coverage.

Insert size metrics should include the mean, median, standard deviation, and the shape of the distribution. Record these metrics for each library preparation batch, because batch effects can introduce systematic differences in insert size.

Alignment metrics should include the overall mapping rate, the fraction of reads with high mapping quality, and the fraction of reads that are properly paired. Low mapping rates or high fractions of improperly paired reads indicate problems that will affect SV detection.

Documentation of Caller Parameters

Document the exact parameters used for each SV calling run. This documentation is essential for reproducing results and for diagnosing why a particular deletion was missed.

Record the software version, the reference genome version, and all parameter settings. Include the minimum and maximum SV size, the quality score thresholds, and any filter settings. Note any deviations from the default parameters and the rationale for those deviations.

Version control for analysis scripts and pipelines is essential. The Carpentries lessons provide foundational training in version control with Git, which enables you to track changes to analysis code and reproduce previous results. The nf-core documentation describes community standards for reproducible workflow configuration that can be applied to SV calling pipelines.

Benchmarking Against Known Variants

Benchmarking your SV calling pipeline against samples with known variants provides a quantitative measure of sensitivity and specificity. The Genome in a Bottle benchmark set described in the 2020 Nature Biotechnology paper provides a reference for germline large deletions and insertions.

Run your pipeline on the benchmark sample and compare your calls to the benchmark set. Calculate the true positive rate, false positive rate, and false negative rate. Examine the characteristics of false negatives to identify systematic biases in your pipeline.

The Galaxy Training Network provides accessible workflow training that includes guidance on running benchmarking analyses and evaluating variant calling performance. The EMBL-EBI Training portal offers learning pathways for bioinformatics data analysis that cover quality assessment and benchmarking approaches.

Common Failure Patterns in SV Detection

Pattern 1: Deletions in Repetitive Regions

Deletions that overlap or are flanked by repetitive elements are frequently missed by short-read SV callers. The reads in these regions cannot be uniquely mapped, resulting in low mapping quality and ambiguous alignment.

The diagnostic signature of this failure pattern is a deletion that is detected by long-read sequencing but absent from short-read calls, with the breakpoints mapping to repetitive elements. The BAM file may show reads with low mapping quality or reads that are randomly placed within the repetitive region.

The solution depends on the specific repetitive element. Some repetitive elements can be handled by using a more sensitive aligner or by adjusting mapping quality filters. Other repetitive elements are effectively invisible to short-read sequencing, and long-read sequencing or optical mapping is required for detection.

Pattern 2: Small Deletions Below the Detection Threshold

Deletions smaller than the caller's minimum size threshold are systematically missed. This failure pattern is straightforward to diagnose by comparing the deletion size to the caller's configured minimum.

The diagnostic signature is a deletion that is smaller than the minimum SV size parameter. The deletion may be detected by small variant callers if it is smaller than 50 base pairs, or may be missed entirely if it falls in the gap between small variant and structural variant detection ranges.

The solution is to adjust the minimum size parameter or to use a caller that can detect smaller deletions. Some SV callers can detect deletions as small as 30 to 50 base pairs, while others have minimum thresholds of several hundred base pairs.

Pattern 3: Low Coverage Regions

Deletions in regions with below-average coverage are frequently missed because the statistical evidence for a copy number change is weak. This failure pattern is common in GC-rich or GC-poor regions, which are underrepresented in standard sequencing libraries.

The diagnostic signature is a deletion in a region where the coverage is substantially below the genome-wide average. The BAM file may show few or no reads in the deletion region, making it impossible to distinguish a genuine deletion from a coverage gap.

The solution is to improve coverage in the affected region. This may involve using a PCR-free library preparation method, adjusting the fragmentation conditions, or using a coverage normalization approach. In some cases, the region may be inherently difficult to sequence, and alternative technologies may be required.

Pattern 4: Heterozygous Deletions with Weak Support

Heterozygous deletions produce weaker signals than homozygous deletions because only one allele is deleted. The read depth reduction is 50% instead of 100%, and the number of discordant read pairs and split reads is correspondingly lower.

The diagnostic signature is a deletion that is present in the sample but has borderline statistical support. The quality score may be just below the threshold, or the number of supporting reads may be just below the caller's minimum.

The solution is to lower the quality score threshold or the minimum supporting read requirement, accepting that this will also increase the number of false positives. Alternatively, use a caller that is specifically designed for heterozygous SV detection or that uses a probabilistic model that accounts for allele frequency.

Pattern 5: Alignment Artifacts Masking the Deletion

Alignment errors can mask genuine deletions by placing reads incorrectly. This failure pattern is common in regions with segmental duplications or other complex genomic structures.

The diagnostic signature is a deletion that is present in the sample but the reads in the region show an unusual alignment pattern. Reads may be clipped at unexpected positions, or may map with many mismatches, suggesting that the aligner has placed them incorrectly.

The solution is to use a more sensitive aligner or to adjust alignment parameters. Some aligners have specific modes for handling structural variation, such as allowing longer indels or using a more sensitive scoring scheme. In complex regions, a combination of alignment strategies may be required.

Limitations of Short-Read SV Detection

Inherent Blind Spots

Short-read sequencing has inherent limitations for structural variant detection that cannot be overcome by adjusting parameters or using different callers. These limitations are fundamental to the technology.

Repetitive regions are the most significant blind spot. Short reads cannot uniquely map to repetitive elements, making it impossible to determine the exact genomic location of reads in these regions. Deletions that are entirely within repetitive regions, or that have breakpoints in repetitive regions, cannot be detected with short-read data.

Large insertions are also difficult to detect with short-read sequencing. While deletions produce a reduction in read depth and discordant read pairs with larger insert sizes, insertions produce an increase in read depth and discordant read pairs with smaller insert sizes. However, the inserted sequence itself cannot be assembled from short reads, so the exact breakpoint and sequence of the insertion cannot be determined.

The Impact of Reference Genome Quality

The quality of the reference genome affects SV detection sensitivity. Regions where the reference genome is misassembled or contains errors will produce spurious SV calls and may mask genuine variants.

The 2021 Genome Biology study found that mapping methods provide the major contribution to variability in SV detection. This finding reflects the impact of reference genome quality and alignment algorithm choices on the ability to detect structural variants.

Reference genome improvements can improve SV detection sensitivity. The availability of complete telomere-to-telomere reference genomes and the incorporation of diverse population sequences into reference panels can reduce the number of false negatives caused by reference genome errors.

The Challenge of Somatic SV Detection

Somatic structural variant detection in cancer samples presents additional challenges beyond those of germline SV detection. Tumor samples are often heterogeneous, with a mixture of tumor and normal cells, and the tumor cells themselves may be genetically heterogeneous.

The 2015 Blood study examined somatic variant calling in acute myeloid leukemia samples and found that two different bioinformatics pipelines gave discrepant results. The researchers observed significantly different calling efficiencies, with poor sensitivity in both pipelines resulting in a high rate of false negatives. They found that the leukemia genomes might be more complex than previously reported, with hundreds of genes mutated at low variant allele frequency.

For somatic SV detection, the variant allele frequency is a critical parameter. Somatic deletions may be present in only a fraction of the cells in the sample, producing weak signals that are difficult to distinguish from noise. The choice of caller and the parameter settings must account for the expected variant allele frequency, which depends on the tumor purity and the clonal structure of the sample.

Quality Control and Reproducibility Considerations

Establishing Quality Control Thresholds

Quality control thresholds for SV detection should be established based on the specific research question and the acceptable balance between sensitivity and specificity. These thresholds should be documented and applied consistently across samples.

For clinical applications, the sensitivity of SV detection is critical because a missed deletion can lead to a missed diagnosis. The 2024 Genome Medicine study found that structural variant detection rates were lower than small variant detection rates, suggesting that additional quality control measures may be needed for SV calls in clinical settings.

For research applications, the balance between sensitivity and specificity depends on the downstream analyses. If SV calls are being used for association studies, false positives can create spurious associations. If SV calls are being used for discovery, false negatives can cause genuine variants to be missed.

Reproducibility Across Runs and Centers

The 2021 Genome Biology study demonstrated substantial variability in SV detection across sequencing centers and replicates. This variability has important implications for multi-center studies and for the reproducibility of SV calls.

To improve reproducibility, standardize the sequencing protocol, library preparation method, and analysis pipeline across all samples. Document any deviations from the standard protocol and assess their impact on SV detection.

The nf-core documentation describes community standards for reproducible workflow configuration that can be applied to SV calling pipelines. Using containerized workflows and version-controlled analysis scripts ensures that the same analysis is performed consistently across samples and centers.

Validation of SV Calls

Validation of SV calls using an orthogonal method is essential for confirming the accuracy of the calls and for identifying systematic biases in the calling pipeline.

PCR validation can confirm the presence of a deletion by amplifying across the predicted breakpoint. This approach is cost-effective for validating a small number of calls but is not feasible for validating large numbers of calls.

Long-read sequencing validation provides sequence-resolved confirmation of structural variants. The 2020 Nature Biotechnology benchmark paper demonstrates the value of integrating multiple sequencing technologies for SV detection and validation.

For clinical applications, validation of SV calls is essential before reporting results. The specific validation requirements depend on the regulatory framework and the clinical context.

Professional Escalation Criteria

When to Seek Additional Expertise

Some SV detection problems require specialized expertise beyond what is available in a typical bioinformatics laboratory. Recognizing when to escalate a problem can save time and prevent incorrect conclusions.

If a deletion is consistently missed across multiple callers and parameter settings, and the region is not in a known difficult-to-sequence area, the problem may require specialized expertise in structural variant detection. Consider consulting with a bioinformatics core facility or a collaborator with experience in SV calling.

If the deletion is in a clinically relevant gene and the sample is from a patient, the problem may require clinical genetics expertise. A missed deletion in a clinical context can have serious consequences, and the case should be escalated to a clinical molecular geneticist or genetic counselor.

Documentation for Escalation

When escalating an SV detection problem, provide comprehensive documentation of the analysis performed and the evidence for the missed deletion.

Include the sample information, sequencing platform, coverage metrics, insert size statistics, and alignment metrics. Document the SV caller version, parameters, and filter settings. Provide the BAM file or a screenshot of the region of interest showing the evidence for the deletion.

Describe the steps already taken to diagnose the problem, including any alternative callers or parameter settings that were tried. This documentation enables the consultant to focus on the remaining possibilities instead of repeating the same analyses.

When to Consider Alternative Technologies

If short-read sequencing consistently fails to detect structural variants in a particular sample or genomic region, alternative technologies may be required.

Long-read sequencing can detect structural variants that are missed by short-read sequencing, particularly in repetitive regions and for larger variants. The cost of long-read sequencing has decreased substantially, making it more accessible for validation and for samples where short-read SV detection is inadequate.

Optical mapping provides a genome-wide view of structural variants without the biases of sequencing-based approaches. This technology can detect large structural variants and can provide information about the genomic context of the variants.

The decision to use alternative technologies should be based on the specific research question, the characteristics of the sample, and the resources available. For clinical applications, the potential impact of a missed structural variant on patient care should be a primary consideration.

Frequently Asked Questions

Why does my variant caller detect some deletions but miss others in the same sample?

Different deletions present different levels of evidence to the variant caller. A deletion in a unique, well-covered region with clear breakpoints will be detected with high confidence. A deletion in a repetitive region, a region with low coverage, or a region where the breakpoints are ambiguous may not produce enough evidence to exceed the caller's detection threshold. The same caller can detect one deletion and miss another because the evidence quality differs between the two loci.

How much coverage do I need for reliable deletion detection?

The coverage required for reliable deletion detection depends on the deletion size, the zygosity, and the detection method. Higher coverage improves sensitivity for all detection methods. The 2024 Genome Medicine study used 37x mean coverage and still detected only 87% of structural variants and aneuploidies, indicating that even moderate-to-high coverage does not guarantee complete detection. For read depth based detection, higher coverage reduces variance and improves the ability to distinguish genuine deletions from coverage noise. For split read detection, higher coverage increases the probability that reads will span breakpoints.

Can I adjust my caller parameters to recover missed deletions?

Adjusting caller parameters can recover some missed deletions, but not all. Lowering the quality score threshold or the minimum supporting read requirement can recover deletions with borderline evidence, but this also increases false positives. Increasing the maximum deletion size can recover large deletions that were filtered out by size limits. However, deletions in repetitive regions or regions with no coverage cannot be recovered by parameter adjustment because the evidence is absent from the data.

What is the difference between false negatives and false positives in SV calling?

False negatives are genuine structural variants that are not reported by the caller. False positives are reported variants that do not actually exist in the sample. The 2021 Genome Biology study found that the false negative rate for SV calling was higher than the false positive rate, with many genuine variants supported by only one center or replicate. False negatives are particularly problematic in clinical applications because they can lead to missed diagnoses.

How do I know if a missed deletion is a real deletion or an artifact?

Confirming that a missed deletion is real requires orthogonal evidence. Long-read sequencing can provide independent confirmation of structural variants. The Genome in a Bottle benchmark set described in the 2020 Nature Biotechnology paper provides a reference for evaluating SV calls. If long-read data is not available, examine the BAM file manually for evidence of the deletion, including reduced read depth, discordant read pairs, and split reads. PCR amplification across the predicted breakpoint can also confirm the deletion.

Why do different SV callers produce different results on the same data?

Different SV callers use different algorithms and evidence types. Some callers rely primarily on read depth, while others use discordant read pairs, split reads, or assembly-based approaches. Each approach has different strengths and weaknesses, and the sensitivity for different types of deletions varies across callers. The 2021 Genome Biology study found that mapping methods provide the major contribution to variability in SV detection, followed by sequencing centers and replicates.

What should I do if a clinically relevant deletion is missed by my pipeline?

If a clinically relevant deletion is missed, first confirm that the deletion is real using orthogonal evidence. Then determine why the deletion was missed by examining coverage, insert size, alignment quality, and caller parameters at the deletion locus. If the deletion cannot be detected with the available data, consider alternative technologies such as long-read sequencing or optical mapping. Escalate the case to a clinical molecular geneticist or genetic counselor, and document the analysis performed and the evidence for the deletion.

How can I reduce false negatives in my SV calling pipeline?

Reducing false negatives requires a systematic approach. Increase coverage if the budget allows, particularly for regions of interest. Use a library preparation method that produces a narrow insert size distribution and uniform coverage. Choose an aligner that is sensitive to structural variation and evaluate its performance on benchmark data. Run multiple SV callers and combine their results, recognizing that requiring support from multiple callers will miss some genuine variants. Validate a sample of calls using an orthogonal method to assess the false negative rate of your pipeline.

Related Bioinformatics Guides

Related Clinical & Scientific Guides

References and Further Reading

NCBI Data Resources. National Center for Biotechnology Information.
EMBL-EBI Training. European Bioinformatics Institute.
Bioconductor. Bioconductor Project.
Galaxy Training Network. Galaxy Project.
nf-core Documentation. nf-core.
The Carpentries Lessons. The Carpentries.
Genome sequencing as a generic diagnostic strategy for rare disease.. Genome medicine, 2024.
Hidden biases in germline structural variant detection.. Genome biology, 2021.
Copy Number Variation.. Methods in molecular biology (Clifton, N.J.), 2018.
The hidden genomic landscape of acute myeloid leukemia: subclonal structure revealed by undetected mutations.. Blood, 2015.
A robust benchmark for detection of germline large deletions and insertions.. Nature biotechnology, 2020.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.