# A Practical Guide to Filtering Structural Variant Calls: Removing Artifacts and Reducing False Positives


## Key Takeaways

- Raw structural variant (SV) call sets from next-generation sequencing are inherently noisy, with a substantial proportion being artifacts from alignment, library preparation, or reference biases, necessitating robust filtering for biological interpretation.
- Effective SV filtering relies on integrating multiple evidence types (read-pair, split-read, depth, assembly), as each has characteristic failure modes; calls supported by independent evidence types are demonstrably more reliable.
- Population frequency filters are critical for removing common benign polymorphisms and systematic reference artifacts, but must be applied with awareness of reference composition and ancestry-specific allele frequencies.
- Visual validation of read depth and sequence alignments remains an indispensable step to adjudicate candidate SVs, even with advanced automated filtering, particularly for complex or clinically significant calls.
- A tiered decision framework, assigning calls to high, medium, or low confidence based on combined evidence strength, genomic context, and population frequency, allows for context-specific filtering stringency, optimizing precision for clinical reporting and recall for research discovery.
- Systematic record-keeping of call set statistics, filtering parameters, and visual validation outcomes is paramount for reproducibility and auditing of SV filtering workflows.

---

Structural variant (SV) calling from next-generation sequencing data produces raw call sets that routinely contain more artifacts than true biological variants. Analysts facing thousands of candidate deletions, duplications, inversions, insertions, and translocations need a systematic filtering strategy that separates real events from alignment noise, library preparation artifacts, and reference biases. This guide provides a step-by-step framework for filtering SV calls using read support, size thresholds, population frequency, and visual validation, with concrete decision criteria for germline and somatic workflows.

## The False Positive Problem in Structural Variant Detection

Structural variants are broadly defined as genomic alterations of 50 base pairs or larger, including deletions, duplications, inversions, insertions, and translocations. These events are a major cause of human genetic disease, and their detection in clinical practice has historically relied on karyotyping and chromosomal microarray analysis. Karyotyping offers very low resolution, while microarrays cannot detect balanced structural variants or indicate the genomic localization and orientation of duplicated segments or insertions. [Optical genome mapping has demonstrated the ability to detect nearly all types of chromosomal aberrations with high concordance compared to standard assays](https://pubmed.ncbi.nlm.nih.gov/34237280), but sequencing-based approaches remain the most widely used method in research and clinical genomics.

The core challenge with sequencing-based SV detection is that every aligner and variant caller produces a substantial number of false positives. These artifacts arise from several sources. Repetitive regions of the genome cause ambiguous read placement, leading to spurious breakpoint calls. Library preparation introduces chimeric molecules that mimic structural variation. GC bias and coverage fluctuations create apparent copy number changes that are not real. Reference errors and misassemblies produce systematic false calls that appear across many samples. Short-read sequencing is particularly prone to these problems because individual reads are far shorter than the structural variants being detected, so evidence must be assembled from discordant read pairs, split reads, and depth signals.

The practical consequence is that raw SV call sets are not directly interpretable. A typical whole-genome sequencing experiment may produce tens of thousands of candidate variants, and the majority of these will be artifacts or common benign polymorphisms. Filtering is not an optional quality step. It is the process that converts raw computational output into a biologically meaningful variant list suitable for downstream analysis, clinical interpretation, or functional validation.

## Core Principles of SV Filtering

Effective SV filtering rests on several principles that apply across calling tools and sequencing platforms. Understanding these principles helps analysts make informed decisions instead of applying arbitrary thresholds.

### Evidence Types and Their Limitations

SV callers integrate multiple lines of evidence to predict variants. Read-pair methods detect insert sizes and orientations that deviate from expectation, identifying clusters of discordant reads that support a breakpoint. Split-read methods identify reads that align discontinuously across a breakpoint, providing base-pair resolution of the junction. Depth-based methods compare read coverage across genomic intervals to identify deletions and duplications. Assembly-based methods construct local contigs and compare them to the reference to identify structural differences.

Each evidence type has characteristic failure modes. Read-pair evidence is sensitive to library preparation artifacts and fails in repetitive regions where insert sizes are inherently variable. Split-read evidence is precise but requires reads that span the breakpoint, which is rare for large variants. Depth evidence detects copy number changes but cannot resolve breakpoints or distinguish tandem duplications from dispersed duplications. Assembly evidence is computationally expensive and can produce chimeric contigs that create false variants.

A filtering strategy should therefore consider which evidence types support each call. Calls supported by multiple independent evidence types are more reliable than calls supported by a single type. Many callers provide a quality score or flag that indicates the evidence combination, and these flags should be used as a primary filtering criterion.

### Reference Bias and Pangenome Considerations

Single-reference alignment introduces systematic bias because the reference genome represents only one haplotype. Variants in the reference that are not present in the sample can cause false read mappings, and variants in the sample that are absent from the reference may be missed. [Pangenome references reduce this bias by representing genetic diversity more accurately than a single reference sequence](https://pubmed.ncbi.nlm.nih.gov/39261641). However, comparing a sample to a pangenome introduces a different problem: variants in the pangenome that are not part of the sample can be misleading and cause false read mappings. These irrelevant variants are generally rarer in terms of allele frequency, and filtering rare variants has been a common heuristic. This blunt approach both fails to remove some irrelevant variants and removes many relevant ones. [Personalized pangenome approaches that impute a sample-specific subgraph based on k-mer counts in the reads have been shown to reduce small variant genotyping errors substantially and make short-read structural variant genotyping of known variants competitive with long-read discovery methods](https://pubmed.ncbi.nlm.nih.gov/39261641).

For filtering purposes, this means that population frequency filters must be applied with awareness of reference composition. A variant that appears rare because it is absent from the reference but common in the population may be a reference artifact instead of a true rare variant. Conversely, a variant that is common in population databases may be a benign polymorphism or a systematic artifact of the reference. Analysts should check whether a candidate variant falls in a region known to have reference assembly errors or segmental duplications.

### The Role of Visual Validation

Automated filtering reduces the number of false positives, but [visual validation remains an important step for minimizing false predictions from structural variant detection](https://pubmed.ncbi.nlm.nih.gov/34034781). Tools that display read depth and sequence alignments allow analysts to adjudicate purported variants across samples and sequencing technologies. These images can be rapidly reviewed to curate large variant call sets. Visual validation is applicable to many biological problems, including variant prioritization in disease studies, analysis of inherited variation, and de novo variant review. [Machine learning packages that operate on these visual representations can dramatically decrease the number of false positives without human review](https://pubmed.ncbi.nlm.nih.gov/34034781), but human inspection remains valuable for complex or clinically significant calls.

## At a Glance: SV Filtering Decision Framework

The following table summarizes the primary filtering criteria and their application across different analysis contexts. These criteria should be applied in sequence, with each step removing a distinct class of artifacts.

| Filtering Criterion | Germline Default | Somatic Default | Primary Artifact Removed |
|---|---|---|---|
| Read support | Minimum 2 split reads or 3 discordant pairs | Minimum 3 split reads or 5 discordant pairs, plus tumor-specific evidence | Alignment noise and single-molecule artifacts |
| Variant size | 50 bp to 1 Mb for short reads | 50 bp to 1 Mb, with upper limit adjusted for tumor purity | Calls outside the detectable range of the platform |
| Population frequency | Remove variants at allele frequency above 1% in matched population | Remove variants above 1% in population, then apply tumor-normal subtraction | Common polymorphisms and reference artifacts |
| Evidence type | Require at least 2 independent evidence types | Require at least 2 evidence types, with breakpoint support preferred | Calls supported by a single noisy signal |
| Visual validation | Review all candidate pathogenic or novel calls | Review all calls used for clinical reporting | Residual artifacts that pass automated filters |

## Step-by-Step SV Filtering Workflow

The filtering workflow proceeds through distinct stages, each with specific inputs, outputs, and decision criteria. The workflow assumes that variant calling has already been completed and that the analyst has access to the raw alignment files for validation.

### Step 1: Assess Caller Output and Quality Metrics

Begin by examining the distribution of calls produced by the variant caller. Generate summary statistics including the total number of calls, the distribution of variant sizes, the distribution of quality scores, and the proportion of calls supported by each evidence type. This initial assessment identifies obvious problems such as an excessive number of calls in repetitive regions or a bimodal quality distribution that suggests two distinct call populations.

Record the caller version, parameters, and reference genome build. These details are essential for reproducibility and for comparing results across samples or experiments. The [Bioconductor project](https://bioconductor.org/) provides reproducible genomic analysis workflows that can help standardize this process, and the [nf-core documentation](https://nf-co.re/docs) describes community standards for pipeline configuration and usage that support consistent variant calling and filtering.

### Step 2: Apply Read Support Thresholds

Read support is the most direct measure of evidence for a structural variant. A true variant should be supported by multiple independent reads or read pairs. The specific thresholds depend on the sequencing depth, read length, and library preparation method.

For germline analysis at 30x whole-genome coverage, a minimum of two split reads or three discordant read pairs is a reasonable starting threshold. For somatic analysis, where tumor purity reduces the variant allele fraction, higher support is needed. A minimum of three split reads or five discordant pairs in the tumor, with no evidence in the matched normal, provides greater specificity.

Low-support calls should not be discarded without examination. They may represent genuine variants in regions of low mappability or variants present at low allele fraction. However, for most applications, low-support calls are more likely to be artifacts, and removing them substantially improves precision.

### Step 3: Apply Size Filters

The size distribution of structural variants detectable by a given platform is bounded by read length and library insert size. Short-read sequencing reliably detects deletions and duplications from approximately 50 base pairs to several megabases. Smaller events are better detected by small-variant callers, and larger events may be missed due to insufficient read-pair span.

Apply a minimum size threshold of 50 base pairs to match the standard definition of structural variation. Apply a maximum size threshold based on the library insert size and the expected detection limit of the caller. For most short-read whole-genome experiments, an upper limit of 1 megabase is appropriate, though this should be adjusted based on the specific library preparation and sequencing platform.

Calls outside this size range should be flagged and examined separately. Very large events may be real but require orthogonal confirmation. Very small events may be real but are better characterized by other methods.

### Step 4: Apply Population Frequency Filters

Population frequency is a powerful filter for removing common polymorphisms and systematic artifacts. A variant that appears at high frequency in population databases is unlikely to be a rare disease-causing event. Conversely, a variant that is absent from population databases is more likely to be a true rare variant or a sample-specific artifact.

The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide access to population variant databases, sequence resources, and analysis services that support frequency-based filtering. Query candidate variants against these resources to obtain allele frequencies across global and ancestry-matched populations.

For germline analysis, remove variants with population allele frequency above 1%. This threshold removes common polymorphisms while retaining rare variants that may be clinically relevant. For somatic analysis, apply the same population frequency filter, then additionally require that the variant is absent or present at low allele fraction in the matched normal sample.

Population frequency filters must be applied with awareness of the reference genome build and the population composition of the database. A variant that is common in one ancestry group may be rare in another, and using a global frequency threshold may incorrectly remove ancestry-specific variants.

### Step 5: Evaluate Evidence Type Combinations

Calls supported by multiple independent evidence types are more reliable than calls supported by a single type. Examine the evidence flags produced by the caller and require that retained calls have support from at least two evidence categories.

For example, a deletion supported by both discordant read pairs and split reads is more reliable than a deletion supported only by depth changes. A duplication supported by depth changes and read-pair orientation is more reliable than a duplication supported only by depth.

Calls supported by a single evidence type should be flagged for additional scrutiny. They may be retained if they have strong support within that evidence type and pass visual validation, but they should not be treated as high-confidence calls.

### Step 6: Perform Visual Validation

Visual validation is the final quality gate before a call is considered high confidence. [Tools that display read depth and sequence alignments allow rapid review of candidate variants](https://pubmed.ncbi.nlm.nih.gov/34034781). Generate images for all calls that pass the automated filters and review them systematically.

The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible workflow training that includes visual validation of genomic variants. These tutorials demonstrate how to generate and interpret alignment images and how to incorporate visual review into reproducible analysis workflows.

During visual review, confirm that the read depth and alignment patterns are consistent with the predicted variant type. A deletion should show reduced depth across the deleted interval with split reads spanning the breakpoints. A duplication should show increased depth across the duplicated interval with read pairs in the expected orientation. An inversion should show read pairs with inverted orientation at the breakpoints.

Calls that do not show convincing visual evidence should be removed or flagged as low confidence. Calls that show clear evidence should be retained. For large call sets, [machine learning approaches that operate on visual representations can reduce the burden of manual review while maintaining high specificity](https://pubmed.ncbi.nlm.nih.gov/34034781).

### Step 7: Apply Context-Specific Filters

The final filtering stage applies filters specific to the biological context of the analysis. For germline analysis, this may include filters for inheritance patterns, linkage to known phenotypes, or overlap with known disease-associated regions. For somatic analysis, this may include filters for tumor purity, clonality, or recurrence across multiple samples from the same patient.

Context-specific filters should be documented and applied consistently across samples. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide learning pathways for bioinformatics analysis that include guidance on context-specific variant interpretation and filtering.

## Options and Tradeoffs in Filtering Stringency

The choice of filtering thresholds involves a tradeoff between precision and recall. Stringent filters remove more false positives but also remove more true variants. Relaxed filters retain more true variants but also retain more artifacts. The optimal balance depends on the downstream application.

### Research Discovery Applications

For research discovery, where the goal is to identify candidate variants for further study, relaxed filtering is appropriate. Retaining more calls allows the analyst to examine a broader set of candidates and apply functional prioritization downstream. The cost is additional time spent reviewing false positives.

A reasonable approach for research discovery is to apply read support and size filters to remove obvious artifacts, then retain all remaining calls for downstream analysis. Population frequency filters can be applied as an annotation instead of a hard filter, allowing the analyst to see the frequency of each candidate variant.

### Clinical Reporting Applications

For clinical reporting, where the goal is to identify variants that may explain a patient's phenotype, stringent filtering is required. False positives in a clinical report can lead to incorrect diagnoses and inappropriate management. False negatives can lead to missed diagnoses.

A reasonable approach for clinical reporting is to apply all automated filters, then require visual validation for every retained call. Calls that pass visual review should be confirmed by an orthogonal method before inclusion in a clinical report. The [PubMed record for Fanconi Anemia](https://pubmed.ncbi.nlm.nih.gov/20301575) illustrates the clinical importance of accurate variant interpretation in a disorder characterized by increased chromosome breakage and a high burden of structural variation.

### Somatic Analysis Applications

Somatic analysis requires additional considerations because tumor samples contain a mixture of tumor and normal cells. The variant allele fraction of a true somatic variant depends on tumor purity and clonality. Low-purity tumors produce low variant allele fractions that may fall below the detection threshold of the caller.

For somatic analysis, filtering should be performed in the context of the matched normal sample. Variants present in the normal sample at comparable allele fraction are germline and should be removed. Variants present only in the tumor are candidate somatic events. The [Prenatal exome sequencing study](https://pubmed.ncbi.nlm.nih.gov/30712880) demonstrates the importance of distinguishing inherited from de novo variants in a clinical context, a distinction that requires careful filtering of germline events.

## Records and Measurements for SV Filtering

Systematic record keeping is essential for reproducible SV filtering. The following records should be maintained for each analysis.

### Call Set Statistics

Record the number of calls at each filtering stage. This includes the raw call count, the number of calls passing read support filters, the number passing size filters, the number passing population frequency filters, and the number passing visual validation. These numbers provide a quantitative measure of the filtering stringency and allow comparison across samples and experiments.

### Filtering Parameters

Record all filtering parameters, including the specific thresholds for read support, size, population frequency, and evidence type. Record the versions of all software tools and the reference genome build. This information is essential for reproducing the analysis and for understanding why specific calls were retained or removed.

### Visual Validation Records

Record the outcome of visual validation for each call. This includes whether the call was confirmed, rejected, or flagged as uncertain, and the reason for the decision. These records provide a basis for auditing the filtering process and for improving filtering criteria over time.

The [The Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in data management and reproducible analysis that supports systematic record keeping in bioinformatics workflows.

## Common Failure Patterns in SV Filtering

Several failure patterns recur across SV filtering analyses. Recognizing these patterns helps analysts avoid common mistakes.

### Overfiltering True Variants

The most common failure is applying thresholds that are too stringent for the data. This occurs when analysts use thresholds from published studies without adjusting for their sequencing depth, library preparation, or sample type. A threshold that works well for 30x whole-genome sequencing may remove most true variants from 10x sequencing data.

The solution is to calibrate thresholds using positive controls. If the analysis includes samples with known structural variants, verify that these variants pass the filtering thresholds. If no positive controls are available, compare the size distribution and genomic distribution of retained calls to expected patterns.

### Underfiltering Artifacts

The opposite failure is applying thresholds that are too relaxed, retaining large numbers of false positives. This occurs when analysts are concerned about missing true variants and therefore retain all calls with minimal evidence.

The solution is to perform visual validation on a random sample of retained calls to estimate the false positive rate. If the false positive rate is high, increase the stringency of read support and evidence type filters.

### Ignoring Reference Composition

Analysts who apply population frequency filters without considering the reference genome composition may incorrectly remove or retain variants in problematic regions. Segmental duplications, low-complexity regions, and known reference errors produce systematic artifacts that appear at high frequency in population databases.

The solution is to annotate calls with genomic context information and to apply region-specific filtering. Calls in known problematic regions should be flagged for additional scrutiny regardless of their population frequency.

### Failing to Validate with Orthogonal Methods

Automated filtering and visual validation reduce false positives but do not eliminate them. For clinically significant calls, orthogonal confirmation is essential. This may include PCR validation, optical genome mapping, or long-read sequencing.

The [optical genome mapping study](https://pubmed.ncbi.nlm.nih.gov/34237280) demonstrated high concordance between optical genome mapping and standard cytogenetic assays for chromosomal aberrations, suggesting that orthogonal methods can provide independent confirmation of sequencing-based calls.

## Limitations of SV Filtering Approaches

SV filtering has inherent limitations that cannot be overcome by threshold adjustment alone. Understanding these limitations is essential for interpreting filtered call sets.

### Detection Limits of Short-Read Sequencing

Short-read sequencing has fundamental limitations for structural variant detection. Variants in repetitive regions are difficult or impossible to detect because reads cannot be uniquely placed. Variants smaller than the read length may be missed by read-pair methods. Variants larger than the library insert size may be missed by depth-based methods.

Filtering cannot recover variants that were never detected by the caller. Analysts should be aware of the detection limits of their platform and should not interpret the absence of a variant as evidence that the variant is not present.

### Population Database Limitations

Population frequency databases are incomplete and biased toward certain ancestry groups. A variant that is absent from a population database may be a rare variant or may be common in an underrepresented population. Frequency-based filtering should be applied with awareness of these limitations.

The [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/) provide access to multiple population databases with different compositions and ascertainment strategies. Comparing frequencies across databases provides a more complete picture than relying on a single database.

### Reference Genome Limitations

The reference genome is not a complete representation of human genetic diversity. Regions that are absent from the reference or misassembled in the reference produce systematic artifacts. [Pangenome references reduce but do not eliminate this problem](https://pubmed.ncbi.nlm.nih.gov/39261641).

Analysts should be aware of the reference genome build and its known limitations. Variants in regions with known reference errors should be interpreted with caution regardless of their filtering status.

## Safety and Regulatory Context for SV Filtering

SV filtering in clinical contexts is subject to regulatory oversight and professional standards. Analysts working in clinical laboratories should be aware of the applicable regulations and should document their filtering procedures accordingly.

### Clinical Laboratory Standards

Clinical laboratories performing structural variant analysis should follow established standards for test validation, quality control, and result reporting. Filtering procedures should be validated using samples with known variants, and the performance characteristics of the filtering approach should be documented.

The [Fanconi Anemia review](https://pubmed.ncbi.nlm.nih.gov/20301575) illustrates the clinical importance of accurate structural variant detection in a disorder with a high burden of chromosomal breakage. In such contexts, filtering errors can have direct consequences for patient care.

### Professional Escalation Criteria

Analysts should have clear criteria for escalating uncertain or complex cases to senior personnel. These criteria should include calls that are potentially clinically significant but have ambiguous evidence, calls in regions with known technical challenges, and calls that are inconsistent with the patient's phenotype.

For prenatal applications, the [PAGE study](https://pubmed.ncbi.nlm.nih.gov/30712880) demonstrated that exome sequencing can identify diagnostic variants in a substantial proportion of fetuses with structural anomalies. In this context, filtering decisions directly affect clinical management decisions, and escalation criteria are essential.

## Practical Implementation Steps

The following steps provide a practical implementation path for SV filtering in a research or clinical laboratory.

### Step 1: Establish the Analysis Environment

Install the required software tools and verify that they are functioning correctly. The [Bioconductor project](https://bioconductor.org/) provides reproducible genomic analysis workflows, and the [nf-core documentation](https://nf-co.re/docs) describes community standards for pipeline configuration. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible tutorials for workflow implementation.

### Step 2: Define Filtering Criteria

Define the filtering criteria based on the analysis context, sequencing platform, and downstream application. Document the criteria and the rationale for each threshold. The criteria should be specific enough to be applied consistently but flexible enough to accommodate sample-specific variation.

### Step 3: Apply Automated Filters

Apply the automated filters in the order described in the workflow. Record the number of calls removed at each stage. Review the distribution of retained calls to identify any unexpected patterns.

### Step 4: Perform Visual Validation

Generate visual validation images for all retained calls. Review the images systematically and record the outcome for each call. For large call sets, consider using machine learning approaches to prioritize calls for manual review.

### Step 5: Document and Report

Document the filtering process, including all parameters, software versions, and validation outcomes. Report the number of high-confidence calls and the evidence supporting each call. For clinical applications, follow the applicable reporting standards and escalate uncertain cases according to the established criteria.

## Building a Tiered Filtering Decision Framework for SV Call Sets

Automated filtering thresholds applied uniformly across a call set treat every variant as if it carries the same evidentiary weight and the same downstream consequence. In practice, structural variant calls differ dramatically in their reliability and their relevance to the biological question at hand. A deletion supported by 40 split reads in a unique region of the genome is not comparable to a duplication supported by three discordant pairs in a segmental duplication. A tiered decision framework assigns each call to a confidence class based on the strength and diversity of its evidence, then applies filtering stringency proportional to the intended use of that call. This approach prevents the two most common failures in SV filtering: discarding genuine variants that happen to fall below an arbitrary threshold, and retaining artifacts that happen to exceed it.

### Defining Confidence Tiers for SV Calls

The tiered framework organizes calls into three confidence classes: high confidence, medium confidence, and low confidence. These tiers are defined by explicit criteria that combine read support, evidence type diversity, genomic context, and population frequency. The criteria are applied consistently across all samples in an analysis, and the tier assignment is recorded for every call.

High-confidence calls are supported by at least two independent evidence types, have read support exceeding the minimum thresholds for the sequencing platform, fall outside known problematic genomic regions, and have population frequency consistent with the analysis context. For germline analysis, a high-confidence deletion might have five or more split reads, ten or more discordant read pairs, and a population allele frequency below 1%. For somatic analysis, the same call would additionally require absence from the matched normal sample at comparable allele fraction.

Medium-confidence calls meet some but not all of the high-confidence criteria. A call supported by strong depth evidence but weak split-read support would fall into this tier. A call with excellent read support but located in a segmental duplication would also be classified as medium confidence. These calls are retained for research discovery but are flagged for additional scrutiny before any clinical or functional interpretation.

Low-confidence calls fail multiple criteria or have evidence that is internally inconsistent. A call supported only by depth changes in a region of variable coverage, or a call with discordant read pairs but no split-read support and no depth change, would be classified as low confidence. These calls are typically removed from the final call set but are recorded in the analysis log for reference.

The tier assignment is not a substitute for visual validation. [Visual validation remains an important step to minimize false-positive predictions from structural variant detection](https://pubmed.ncbi.nlm.nih.gov/34034781), and the tiered framework determines which calls receive visual review and how much review effort is allocated to each call. High-confidence calls receive a rapid confirmation review. Medium-confidence calls receive a detailed review with attention to the specific evidence weakness that prevented high-confidence classification. Low-confidence calls are reviewed only if they fall in a genomic region of particular interest to the analysis.

### Applying Tier-Specific Filtering Stringency

The tiered framework changes the filtering decision from a binary pass or fail to a graduated response. Instead of applying a single read support threshold to all calls, the analyst applies different thresholds depending on the downstream use of the call.

For research discovery applications, high-confidence and medium-confidence calls are both retained in the final call set. Low-confidence calls are removed unless they fall in a region of specific biological interest, in which case they are flagged for orthogonal validation. This approach maximizes recall while maintaining a manageable false positive rate.

For clinical reporting applications, only high-confidence calls are eligible for inclusion in a clinical report. Medium-confidence calls are retained in the analysis file but are not reported without orthogonal confirmation. Low-confidence calls are excluded from clinical consideration. This approach prioritizes precision over recall, recognizing that [false positives in a clinical report can lead to incorrect diagnoses and inappropriate management](https://pubmed.ncbi.nlm.nih.gov/30712880).

For somatic analysis, the tiered framework incorporates the matched normal comparison as a tier criterion instead of a separate filter. A variant present in the tumor with strong evidence but also present in the normal sample at comparable allele fraction is classified as low confidence for somatic significance regardless of its technical evidence quality. A variant present only in the tumor with strong evidence is classified as high confidence. This integration prevents the common error of applying technical filters and germline filters independently, which can produce contradictory results for variants with mixed evidence.

### Implementing the Tiered Framework in Practice

Implementation begins with defining the tier criteria in a written analysis plan before examining the call set. The criteria should specify the exact read support thresholds, evidence type requirements, genomic context annotations, and population frequency cutoffs for each tier. The [nf-core documentation](https://nf-co.re/docs) describes community standards for pipeline configuration that support consistent implementation of such criteria across samples and experiments.

The analysis plan should also specify how tier assignments are recorded. A common approach is to add a tier column to the variant call format file, with the tier assignment populated by a script that applies the criteria automatically. The [Bioconductor project](https://bioconductor.org/) provides packages for manipulating and annotating genomic ranges that can be used to implement tier assignment in a reproducible workflow.

After the automated tier assignment, the analyst reviews a sample of calls from each tier to verify that the criteria are performing as intended. If the high-confidence tier contains an unexpectedly large proportion of calls that fail visual review, the criteria are too relaxed. If the medium-confidence tier contains calls with excellent visual evidence, the criteria may be too stringent. The [Galaxy Training Network](https://training.galaxyproject.org/) provides accessible tutorials for implementing and validating such filtering workflows.

### Recording Tier Assignments and Filtering Outcomes

The tiered framework generates records that support both reproducibility and continuous improvement of the filtering process. For each analysis, record the number of calls assigned to each tier, the number of calls removed at each filtering stage, and the outcomes of visual validation for each tier. These records allow the analyst to quantify the false positive rate within each tier and to adjust the tier criteria if the observed rates deviate from expectations.

The records should also capture the specific evidence that determined the tier assignment for each call. This includes the read support counts, the evidence types present, the genomic context annotations, and the population frequency values. The [The Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in data management and reproducible analysis that supports systematic record keeping in bioinformatics workflows.

For clinical applications, the tier assignment and the evidence supporting it become part of the laboratory record. The [EMBL-EBI Training](https://www.ebi.ac.uk/training) resources provide learning pathways that include guidance on documenting variant interpretation and filtering decisions in clinical contexts.

### Troubleshooting Tier Assignment Problems

Several failure patterns emerge when implementing a tiered framework. Recognizing these patterns allows the analyst to correct the criteria before they propagate through the analysis.

The first pattern is tier inflation, where the majority of calls are assigned to the high-confidence tier. This occurs when the criteria are too permissive, often because the read support thresholds are set too low for the sequencing depth or because the evidence type requirement is not enforced strictly. The correction is to increase the read support thresholds or to require a specific combination of evidence types instead of any two types.

The second pattern is tier deflation, where almost all calls are assigned to the low-confidence tier. This occurs when the criteria are too stringent for the data, often because the thresholds were borrowed from a study using a different sequencing platform or coverage level. The correction is to calibrate the thresholds using positive controls, such as samples with known structural variants that should pass the high-confidence criteria.

The third pattern is tier inconsistency, where calls with similar evidence receive different tier assignments. This occurs when the criteria are not applied uniformly, often because the automated script has a bug or because some calls are manually overridden without documentation. The correction is to audit the tier assignment script and to require documentation for any manual tier changes.

### Integrating the Tiered Framework with Orthogonal Validation

The tiered framework identifies which calls warrant orthogonal validation and prioritizes the order in which validation is performed. High-confidence calls that will be used for clinical reporting should be validated first. Medium-confidence calls that fall in regions of biological interest should be validated next. Low-confidence calls are validated only if they are the only evidence for a variant in a region of critical importance.

[Optical genome mapping has demonstrated the ability to detect nearly all types of chromosomal aberrations with high concordance compared to standard assays](https://pubmed.ncbi.nlm.nih.gov/34237280), making it a suitable orthogonal method for confirming sequencing-based calls. Long-read sequencing provides another orthogonal approach, particularly for variants in repetitive regions where short-read evidence is ambiguous.

The tier assignment should be updated after orthogonal validation. A medium-confidence call that is confirmed by optical genome mapping or long-read sequencing can be upgraded to high confidence. A high-confidence call that fails orthogonal validation should be downgraded and investigated for systematic causes, such as a reference error or an alignment artifact that produces consistent false evidence.

### Using the Tiered Framework for Multi-Sample Analyses

The tiered framework is particularly valuable for analyses that compare structural variants across multiple samples, such as family studies, cohort studies, or tumor-normal comparisons. In these contexts, the tier assignment provides a consistent basis for comparing calls across samples and for identifying variants that are shared or unique.

For family studies, the tier assignment helps distinguish inherited variants from de novo variants. A variant that is high confidence in the proband and high confidence in one parent is likely inherited. A variant that is high confidence in the proband but absent from both parents requires careful review, as it may be a de novo event or a false positive in the proband. The [Prenatal exome sequencing study](https://pubmed.ncbi.nlm.nih.gov/30712880) demonstrated the clinical importance of distinguishing inherited from de novo variants in a prenatal context, a distinction that requires careful filtering of germline events.

For cohort studies, the tier assignment supports the calculation of variant frequencies across the cohort. Variants that are high confidence in multiple samples are more likely to be genuine polymorphisms or recurrent artifacts. Variants that are high confidence in a single sample require additional scrutiny, as they may be sample-specific artifacts or rare genuine variants.

For tumor-normal comparisons, the tier assignment integrates the matched normal evidence into the confidence classification. A variant that is high confidence in the tumor and absent from the normal is classified as high confidence for somatic significance. A variant that is high confidence in both tumor and normal is classified as germline and removed from the somatic call set. This integration prevents the common error of applying technical filters and germline filters independently.

### Escalation Criteria within the Tiered Framework

The tiered framework provides a structured basis for escalation decisions. Calls that are classified as medium confidence but have potential clinical significance should be escalated for senior review. Calls that are classified as low confidence but fall in a region of known disease association should also be escalated, as the technical evidence may be limited by the genomic context instead of by the absence of a genuine variant.

The [Fanconi Anemia review](https://pubmed.ncbi.nlm.nih.gov/20301575) illustrates the clinical importance of accurate structural variant detection in a disorder characterized by increased chromosome breakage. In such contexts, a low-confidence call in a Fanconi anemia gene should be escalated for orthogonal validation instead of silently removed by the filtering process.

Escalation should also occur when the tier assignment produces unexpected patterns. If a particular genomic region consistently produces high-confidence calls that fail visual validation, the region should be flagged as problematic and the tier criteria should be adjusted to account for the regional artifact. If a particular sample produces an unusually high proportion of high-confidence calls, the sample should be investigated for library preparation or sequencing problems.

### Practical Steps for Adopting the Tiered Framework

Adopting the tiered framework requires a deliberate implementation process. Begin by reviewing the existing filtering workflow and identifying the decisions that are currently made implicitly. Write down the criteria that are currently used to retain or remove calls, and assess whether these criteria are applied consistently across samples and analysts.

Next, define the tier criteria explicitly using the categories described above. Start with conservative criteria that assign most calls to the medium or low confidence tiers, then relax the criteria based on the observed false positive rates from visual validation. The goal is to achieve a high-confidence tier with a false positive rate below 5% as determined by visual review.

Implement the tier assignment in a script or workflow that can be applied consistently across samples. The [nf-core documentation](https://nf-co.re/docs) provides guidance on implementing reproducible workflows, and the [Bioconductor project](https://bioconductor.org/) provides packages for genomic range manipulation and annotation that support tier assignment.

Finally, validate the tiered framework using samples with known structural variants. Verify that known true variants are assigned to the high-confidence tier and that known artifacts are assigned to the low-confidence tier. Adjust the criteria as needed and document the validation results.

The tiered framework does not eliminate the need for judgment in SV filtering, but it makes the judgment explicit and auditable. By assigning each call to a confidence class and applying filtering stringency proportional to the intended use, the framework reduces the risk of both overfiltering and underfiltering. It also provides a structured basis for escalation, orthogonal validation, and multi-sample comparison that is difficult to achieve with a single set of uniform thresholds.

## Frequently Asked Questions

### What is the minimum read support needed for a confident structural variant call?

The minimum read support depends on the sequencing depth, read length, and variant type. For germline analysis at 30x whole-genome coverage, a minimum of two split reads or three discordant read pairs is a reasonable starting point. For somatic analysis, higher support is needed because tumor purity reduces the variant allele fraction. The optimal threshold should be calibrated using positive controls and should be adjusted based on the observed false positive rate.

### How do I choose between precision and recall when filtering structural variants?

The choice depends on the downstream application. Research discovery applications benefit from relaxed filtering that retains more candidate variants for functional prioritization. Clinical reporting applications require stringent filtering to minimize false positives. A practical approach is to apply relaxed filters for initial analysis, then apply increasingly stringent filters for variants that are prioritized for clinical review or functional validation.

### Why do population frequency filters remove some true variants?

Population frequency filters remove variants that are common in the population, and some of these variants may be true biological variants that are benign polymorphisms. The goal of frequency filtering is to remove common variants that are unlikely to be disease-causing, not to remove all true variants. Variants that are removed by frequency filters can be recovered by examining the frequency annotation and determining whether the variant is relevant to the specific analysis.

### How does reference bias affect structural variant filtering?

Reference bias occurs when the reference genome does not represent the genetic diversity of the sample being analyzed. Variants in the reference that are not in the sample can cause false read mappings, and variants in the sample that are absent from the reference may be missed. [Pangenome references reduce this bias but introduce new challenges](https://pubmed.ncbi.nlm.nih.gov/39261641). Personalized pangenome approaches that impute sample-specific subgraphs based on read content have been shown to improve variant genotyping accuracy.

### What is the role of visual validation in structural variant filtering?

[Visual validation is an important step for minimizing false positive predictions](https://pubmed.ncbi.nlm.nih.gov/34034781). Tools that display read depth and sequence alignments allow analysts to confirm that the evidence supports the predicted variant type. Visual validation is particularly important for calls that will be used for clinical reporting or functional validation. Machine learning approaches that operate on visual representations can reduce the burden of manual review.

### How do I handle structural variant calls in repetitive regions?

Calls in repetitive regions should be treated with caution because read placement is ambiguous and artifacts are common. These calls should be flagged for additional scrutiny regardless of their read support or population frequency. Orthogonal confirmation with long-read sequencing or optical genome mapping may be necessary to confirm calls in these regions.

### What records should I keep for structural variant filtering?

Keep records of the raw call count, the number of calls removed at each filtering stage, all filtering parameters, software versions, reference genome build, and visual validation outcomes. These records are essential for reproducibility and for auditing the filtering process. The [The Carpentries Lessons](https://carpentries.org/lessons) provide foundational training in data management that supports systematic record keeping.

### When should I escalate a structural variant call for professional review?

Escalate calls that are potentially clinically significant but have ambiguous evidence, calls in regions with known technical challenges, and calls that are inconsistent with the patient's phenotype. For clinical applications, follow the applicable reporting standards and consult with senior personnel before including uncertain calls in a clinical report.

## Related Bioinformatics Guides

- [Single-Cell RNA Sequencing Quality Control: A Practical Guide to Filtering and Metrics](/knowledge/bioinformatics/single-cell-rna-sequencing-quality-control-a-practical-guide-to-filtering-and-metrics)
- [Detecting Structural Variants with Long-Read Sequencing: Methods and Considerations](/knowledge/bioinformatics/detecting-structural-variants-with-long-read-sequencing-methods-and-considerations)
- [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow)
- [Metagenomics Tools: A Practical Guide to Software and Pipelines](/knowledge/bioinformatics/metagenomics-tools-a-practical-guide-to-software-and-pipelines)
- [Microbiome Data Analysis in R: A Practical Guide for Compositional Data](/knowledge/bioinformatics/microbiome-data-analysis-in-r-a-practical-guide-for-compositional-data)

## Related Clinical & Scientific Guides

* [A Practical Guide to Detecting Antimicrobial Resistance Genes in Shotgun Metagenomic Data](/knowledge/bioinformatics/a-practical-guide-to-detecting-antimicrobial-resistance-genes-in-shotgun-metagenomic-data)
* [Computational Immunology: Modeling the Immune System](/knowledge/bioinformatics/computational-immunology-modeling-the-immune-system)
* [How to Set Hard Filters for Germline Variant Calling: A Practical Guide to GATK Best Practices](/knowledge/bioinformatics/how-to-set-hard-filters-for-germline-variant-calling-a-practical-guide-to-gatk-best-practices)


## References and Further Reading

- [NCBI Data Resources](https://www.ncbi.nlm.nih.gov/). National Center for Biotechnology Information.
- [EMBL-EBI Training](https://www.ebi.ac.uk/training). European Bioinformatics Institute.
- [Bioconductor](https://bioconductor.org/). Bioconductor Project.
- [Galaxy Training Network](https://training.galaxyproject.org/). Galaxy Project.
- [nf-core Documentation](https://nf-co.re/docs). nf-core.
- [The Carpentries Lessons](https://carpentries.org/lessons). The Carpentries.
- [Optical genome mapping enables constitutional chromosomal aberration detection.](https://pubmed.ncbi.nlm.nih.gov/34237280). American journal of human genetics, 2021.
- [Samplot: a platform for structural variant visual validation and automated filtering.](https://pubmed.ncbi.nlm.nih.gov/34034781). Genome biology, 2021.
- [Fanconi Anemia.](https://pubmed.ncbi.nlm.nih.gov/20301575). 1993.
- [Personalized pangenome references.](https://pubmed.ncbi.nlm.nih.gov/39261641). Nature methods, 2024.
- [Prenatal exome sequencing analysis in fetal structural anomalies detected by ultrasonography (PAGE): a cohort study.](https://pubmed.ncbi.nlm.nih.gov/30712880). Lancet (London, England), 2019.

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