# Detecting Fusion Transcripts with Long-Read RNA-Seq: A Practical Workflow Using JAFFAL and LongGF

Fusion transcripts are chimeric RNA molecules formed when exons from two different genes are joined, often as a consequence of genomic rearrangements such as translocations, inversions, or deletions. These transcripts are clinically significant because they can encode oncogenic proteins that drive cancer development and progression. Long-read RNA sequencing technologies, including PacBio Iso-Seq and Oxford Nanopore direct RNA sequencing, produce full-length transcript reads that can span entire fusion junctions, providing an advantage over short-read platforms that must infer chimeric junctions from fragmented sequence data. This article presents a practical workflow for detecting fusion transcripts from long-read RNA-seq data using two complementary tools: JAFFAL and LongGF. The workflow covers data preparation, alignment, fusion calling, filtering, validation, and interpretation, with concrete commands and decision criteria for each stage.

## Understanding the Long-Read Advantage for Fusion Detection

Long-read RNA-seq has matured substantially over the past five years, with improvements in both instrumentation and analytical methods enabling routine application to transcriptome studies. Unlike short-read sequencing, which generates fragments typically 150 base pairs in length, long-read platforms can sequence entire mRNA molecules in a single read. This capability is particularly valuable for fusion transcript detection because the chimeric junction, where two genes are joined, can be captured within a single continuous read. Short-read approaches must rely on paired-end information and split-read mapping to infer that two genomic regions are connected, a process that becomes increasingly difficult when fusion breakpoints fall within repetitive or homologous sequences.

The practical consequence of this technological difference is that long-read data can resolve fusion transcripts at isoform resolution. A single long read can reveal also that a fusion exists but also which exons of each partner gene are included, whether the fusion preserves the reading frame, and whether alternative splicing produces multiple fusion isoforms from the same genomic rearrangement. This information is directly relevant for clinical interpretation because different fusion isoforms can have different functional consequences. For example, a fusion that preserves the kinase domain of one partner while truncating the other may produce a constitutively active protein, whereas an out-of-frame fusion may be subject to nonsense-mediated decay and have no oncogenic effect.

The maturation of long-read RNA-seq has been accompanied by benchmarking studies that identify strengths and limitations of different analytical approaches. These studies consistently show that long-read data improve fusion detection sensitivity in many contexts, though short-read data still contribute valuable information for validating candidate fusions and for detecting fusions expressed at very low levels. The most effective strategies often combine both data types, using long reads to resolve full-length fusion isoforms and short reads to provide orthogonal confirmation and quantitative support.

## Core Principles of Fusion Transcript Detection

Fusion transcript detection from RNA-seq data rests on a straightforward principle: identify reads that contain sequence from two different genes and determine the precise junction between them. In practice, this task is complicated by several factors that must be addressed at each stage of the analysis.

The first complication is sequencing error. Long-read platforms have higher error rates than short-read platforms, with error rates that vary by platform and chemistry. These errors can create false chimeric junctions when a sequencing error causes a read to misalign to a different genomic location. The second complication is alignment ambiguity. Many genes share sequence similarity due to common evolutionary origins, and reads from paralogous genes can align to multiple genomic locations with similar scores. The third complication is biological noise. Trans-splicing, template switching during library preparation, and read chimeras formed during sequencing can all produce reads that appear to be fusions but do not represent genuine genomic rearrangements.

Effective fusion detection tools address these complications through multiple strategies. One common approach is to anchor breakpoints to exon boundaries, recognizing that most genuine fusions in cancer arise from genomic rearrangements that join exons from different genes. Another approach is to realign unaligned regions of reads to identify chimeric junctions that initial alignment missed. A third approach is to cluster breakpoints across multiple reads, requiring that a candidate fusion be supported by more than one read before it is reported. These strategies reduce false positives while maintaining sensitivity for genuine fusions.

The choice of alignment tool and parameters has a substantial impact on fusion detection. Most fusion detection tools accept alignments in BAM format and expect reads to be aligned with splice-aware aligners that can handle reads spanning exon junctions. For long-read data, aligners such as minimap2 are commonly used because they are designed to handle the higher error rates and longer read lengths of PacBio and Nanopore data. The alignment parameters should be set to allow for the expected error rate of the sequencing platform, and the output should include the supplementary alignments that represent chimeric read mappings.

## At a Glance: Workflow Overview

| Workflow Stage | Primary Tool | Input | Key Output | Decision Point |
| --- | --- | --- | --- | --- |
| Quality assessment | FastQC, NanoPlot | Raw FASTQ files | Per-read quality metrics, read length distribution | Proceed if median read quality meets platform expectations |
| Alignment | minimap2 | Filtered FASTQ, reference transcriptome | Sorted BAM with supplementary alignments | Verify alignment rate exceeds 70 percent |
| Fusion calling | JAFFAL | Sorted BAM | BEDPE file with candidate fusions | Review candidates with split-read support |
| Fusion calling | LongGF | Sorted BAM | Text output with fusion coordinates | Cross-check candidates against JAFFAL results |
| Filtering | Custom scripts, BEDTools | Fusion calls from both tools | Filtered candidate list | Apply read support and breakpoint criteria |
| Validation | IGV, PCR | Filtered candidates, BAM files | Visual confirmation, experimental validation | Confirm junction sequence and breakpoint |

## Preparing Long-Read RNA-Seq Data for Fusion Analysis

### Input Data Requirements

The workflow begins with raw sequencing data in FASTQ format from either PacBio or Oxford Nanopore platforms. Both platforms produce full-length transcript reads, but the data characteristics differ in ways that affect downstream analysis. PacBio Iso-Seq data typically have lower error rates and more uniform read lengths, while Nanopore direct RNA-seq data can be produced in real time and can capture base modifications in addition to sequence information. The choice of platform affects the expected error profile and therefore the alignment parameters and filtering thresholds used in the analysis.

Before beginning fusion detection, verify that the sequencing library was prepared from RNA of adequate quality. Degraded RNA produces truncated transcripts that may not span fusion junctions, reducing detection sensitivity. The library preparation method also matters. Some protocols include a PCR amplification step that can introduce template switching artifacts, while PCR-free protocols reduce this risk but require more input RNA. For clinical samples with limited material, the tradeoff between input requirements and artifact rates must be considered.

### Quality Assessment and Read Filtering

Quality assessment of long-read data should examine read length distribution, per-read quality scores, and the proportion of reads that align to the reference. Tools such as NanoPlot provide summary statistics and visualizations that help identify problems early in the workflow. Reads that are substantially shorter than the expected transcript length may represent degraded RNA or incomplete reverse transcription, and these reads are less likely to span fusion junctions.

Read filtering decisions should be based on the quality metrics observed in the specific dataset instead of on universal thresholds. A common approach is to remove reads shorter than a minimum length threshold, typically 500 base pairs for transcriptome analysis, and to remove reads with average quality scores below a platform-specific cutoff. However, these thresholds should be adjusted based on the read length distribution of the dataset. If the library contains many short transcripts, a higher minimum length threshold will discard legitimate data.

The filtering step also provides an opportunity to remove adapter sequences and poly-A tails, which can interfere with alignment. Most long-read library preparation protocols include adapter sequences that must be trimmed before alignment. Poly-A tails are biologically informative for some analyses but can cause alignment artifacts when they occur at the ends of reads, so they are often removed before fusion detection.

### Reference Preparation

Fusion detection requires a reference to which reads can be aligned. The choice of reference depends on the analysis goals. For fusion detection, the reference transcriptome is typically derived from the genome annotation, with one transcript sequence per gene or per isoform. Using the transcriptome instead of the genome as the alignment reference simplifies the detection of chimeric reads because reads are expected to align to transcript sequences, and fusions appear as reads that align to two different transcripts.

The reference transcriptome should be prepared from a current genome annotation that includes all known isoforms. The NCBI provides access to reference genome and annotation resources that can be used for this purpose. The annotation should be filtered to include only transcripts from the species of interest, and the transcript sequences should be extracted in FASTA format. Some fusion detection tools require the reference in specific formats, so the preparation step should follow the tool documentation.

An alternative approach is to align reads to the genome instead of the transcriptome. Genome alignment allows fusion breakpoints to be mapped to genomic coordinates, which is useful for correlating fusion transcripts with genomic rearrangements. However, genome alignment is more computationally intensive and requires splice-aware aligners that can handle reads spanning introns. The choice between transcriptome and genome alignment should be made based on the downstream analysis goals and the capabilities of the fusion detection tools being used.

## Alignment Strategies for Chimeric Read Detection

### Selecting an Aligner for Long Reads

The alignment step is critical for fusion detection because the aligner must correctly identify reads that span exon junctions within a single gene as well as reads that span junctions between two different genes. For long-read data, the aligner must also tolerate the higher error rates of PacBio and Nanopore platforms. Minimap2 is widely used for this purpose because it is designed for long reads and can produce the supplementary alignments that represent chimeric mappings.

The alignment parameters should be set based on the sequencing platform and the expected error rate. For PacBio data, the error rate is typically lower, and the alignment parameters can be more stringent. For Nanopore data, the error rate is higher, and the alignment parameters should be more permissive. The aligner documentation provides recommended parameters for different platforms, and these recommendations should be used as starting points that are adjusted based on the observed alignment statistics.

The output of the alignment step is a BAM file that contains the primary alignment for each read as well as supplementary alignments for reads that map to multiple locations. Fusion detection tools use these supplementary alignments to identify chimeric reads. It is important to verify that the aligner is configured to output supplementary alignments, as some alignment modes suppress them by default.

### Verifying Alignment Quality

After alignment, examine the alignment statistics to verify that the data are suitable for fusion detection. The alignment rate, defined as the proportion of reads that align to the reference, should be high for good quality data. Low alignment rates may indicate problems with the reference, the library preparation, or the sequencing run. The alignment rate should be interpreted in the context of the expected transcriptome composition. If the sample contains a high proportion of reads from genes not represented in the reference annotation, the alignment rate will be lower.

The distribution of alignment scores and the proportion of reads with supplementary alignments are also informative. A high proportion of reads with supplementary alignments may indicate that the reference contains many similar sequences, or it may indicate that the library contains many chimeric molecules. Both situations require careful interpretation in the fusion detection step.

### Handling Multi-Mapping Reads

Reads that align to multiple locations in the reference present a challenge for fusion detection. These multi-mapping reads can arise from paralogous genes, from repetitive elements within transcripts, or from genuine fusions where one partner gene has a close relative elsewhere in the genome. Fusion detection tools handle multi-mapping reads differently, and the choice of strategy affects both sensitivity and specificity.

One approach is to require that a read have a unique primary alignment before it is considered for fusion detection. This approach reduces false positives but may miss fusions involving genes with close paralogs. Another approach is to consider all possible alignments for each read and to report fusions that are supported by any alignment. This approach increases sensitivity but also increases the false positive rate. The choice between these strategies should be based on the analysis goals. For discovery-oriented research, the more permissive approach may be preferred, while for clinical applications, the more stringent approach may be necessary.

## Fusion Calling with JAFFAL

### How JAFFAL Works

JAFFAL is a fusion detection tool designed specifically for long-read RNA-seq data. The tool takes aligned reads in BAM format and identifies chimeric reads that align to two different genes. JAFFAL applies several filtering strategies to reduce false positives, including requiring that the breakpoint be anchored to exon boundaries and that the fusion be supported by a minimum number of reads.

The JAFFAL algorithm begins by scanning the BAM file for reads with supplementary alignments. For each such read, the tool determines whether the two aligned segments correspond to different genes. If they do, the read is considered a candidate fusion read. The tool then clusters candidate reads that support the same fusion and applies filtering criteria to the clustered candidates.

One of the key filtering criteria used by JAFFAL is the requirement that breakpoints be anchored to exon boundaries. This criterion is based on the observation that most genuine fusions in cancer arise from genomic rearrangements that join exons, and that fusions with breakpoints in intronic or intergenic regions are more likely to be artifacts. The exon boundary anchoring criterion reduces false positives but may miss fusions with breakpoints in intronic regions, which can occur in some cancer types.

### Running JAFFAL

The JAFFAL tool is run from the command line with the aligned BAM file as input. The tool requires that the BAM file be sorted and indexed, and it uses the reference annotation to determine gene boundaries. The command syntax and parameters are documented in the tool repository, and the parameters should be adjusted based on the sequencing platform and the desired sensitivity.

A typical JAFFAL command specifies the input BAM file, the reference annotation file, and the output directory. The tool produces a BEDPE file that lists the candidate fusions with their genomic coordinates, the genes involved, and the number of supporting reads. The output also includes information about the breakpoint positions and the strand orientation of the fusion.

The runtime for JAFFAL depends on the size of the BAM file and the number of chimeric reads. For typical long-read RNA-seq datasets, the runtime is on the order of hours, and the tool can be run on a standard workstation with sufficient memory. The tool documentation provides guidance on memory requirements and runtime expectations.

### Interpreting JAFFAL Output

The JAFFAL output should be examined carefully before proceeding to downstream analysis. The BEDPE file contains one line per candidate fusion, with columns for the genomic coordinates of the two breakpoints, the names of the two genes, and the number of supporting reads. Candidates with a single supporting read should be treated with caution, as they may represent sequencing artifacts or template switching during library preparation.

The strand orientation of the fusion is an important consideration. Fusions can be in the same orientation or in opposite orientations, depending on the underlying genomic rearrangement. In-frame fusions that preserve the reading frame of both partner genes are more likely to produce functional proteins and are therefore of greater interest for cancer research. The JAFFAL output includes information about the strand orientation, which can be used to prioritize candidates for validation.

## Fusion Calling with LongGF

### How LongGF Works

LongGF is another fusion detection tool for long-read RNA-seq data, with a different algorithmic approach than JAFFAL. LongGF uses a graph-based method to identify fusion transcripts, constructing a graph of aligned segments and searching for paths that connect two different genes. This approach can identify fusions that involve more than two genes, such as complex rearrangements that join three or more genomic regions.

LongGF also incorporates error correction strategies to improve accuracy. The tool realigns unaligned regions of reads to identify chimeric junctions that initial alignment missed, and it clusters breakpoints across multiple reads to reduce the impact of individual sequencing errors. These strategies are particularly important for Nanopore data, which have higher error rates than PacBio data.

The LongGF algorithm is designed to work with both transcriptome and genome alignments. When used with transcriptome alignments, the tool identifies fusions between transcripts. When used with genome alignments, the tool can identify fusions with breakpoints in intronic or intergenic regions, providing a more comprehensive view of the genomic rearrangements underlying fusion transcripts.

### Running LongGF

LongGF is run from the command line with the aligned BAM file and the reference annotation as inputs. The tool requires that the BAM file be sorted and indexed, and it produces a text output file with the fusion coordinates and supporting read information. The command syntax and parameters are documented in the tool repository.

The parameters for LongGF should be adjusted based on the sequencing platform and the desired sensitivity. The tool has parameters for the minimum number of supporting reads, the minimum alignment score, and the maximum distance between breakpoints. These parameters should be set based on the error profile of the sequencing platform and the expected characteristics of fusions in the sample.

LongGF can be more computationally intensive than JAFFAL, particularly for large datasets. The tool documentation provides guidance on memory requirements and runtime expectations, and the tool can be run on a high-performance computing cluster for large datasets.

### Interpreting LongGF Output

The LongGF output contains the fusion coordinates, the genes involved, and the number of supporting reads. The output format is different from JAFFAL, so the results from the two tools must be converted to a common format for comparison. The comparison of results from multiple tools is an important validation step, as fusions detected by multiple independent tools are more likely to be genuine.

LongGF also provides information about the breakpoint positions at single-base resolution. This information is useful for designing validation experiments, such as PCR primers that span the fusion junction. The breakpoint information can also be used to determine whether the fusion is in-frame and whether it preserves functional domains of the partner genes.

## Filtering and Validation of Candidate Fusions

### Applying Read Support Thresholds

The first filtering step is to apply a minimum read support threshold to the candidate fusions. Fusions supported by a single read are more likely to be artifacts, particularly in datasets with high error rates. A common threshold is to require at least two supporting reads, though the appropriate threshold depends on the sequencing depth and the error rate of the platform.

The read support threshold should be applied consistently across all candidates, but the interpretation of the threshold should consider the expression level of the fusion. Fusions expressed at low levels will have fewer supporting reads even if they are genuine, so a strict threshold may miss low-expressed fusions. For clinical applications where sensitivity is critical, a lower threshold may be used, with additional validation steps to confirm the candidates.

### Cross-Tool Comparison

Comparing the results from JAFFAL and LongGF provides a powerful validation strategy. Fusions detected by both tools are more likely to be genuine, as the two tools use different algorithms and are unlikely to produce the same false positive. The comparison should be done at the level of the fusion junction, requiring that the breakpoint coordinates and the gene identities match between the two tools.

The cross-tool comparison can be performed with custom scripts or with existing tools for BED file manipulation. The comparison should account for small differences in breakpoint coordinates, as the two tools may report slightly different positions for the same fusion. A common approach is to allow a small window of tolerance, such as 10 base pairs, when matching breakpoints between tools.

### Visual Inspection with IGV

Visual inspection of candidate fusions in a genome browser provides additional confidence. The Integrative Genomics Viewer (IGV) can display the aligned reads at the fusion breakpoint, allowing the analyst to verify that the reads genuinely span the junction between the two genes. Visual inspection can identify artifacts that pass automated filters, such as reads with alignment errors that create false chimeric junctions.

The visual inspection should examine the read alignments at both breakpoints, verifying that the reads have high-quality alignments to both genes and that the junction is clean. Reads with soft-clipped ends at the junction may indicate alignment errors, while reads with clear split alignments to both genes provide strong support for the fusion.

### Experimental Validation

For fusions that will be reported or used for downstream analysis, experimental validation is recommended. RT-PCR with primers spanning the fusion junction can confirm that the fusion transcript is present in the sample. Sanger sequencing of the PCR product can confirm the exact junction sequence. For fusions with clinical implications, validation in an independent sample or with an orthogonal method is essential.

The design of validation primers requires knowledge of the fusion junction sequence. The long-read data can provide this sequence directly, as the reads that support the fusion contain the full junction sequence. The primer design should account for the possibility of multiple fusion isoforms, and the PCR conditions should be optimized to amplify the expected product.

## Records and Measurements for Reproducible Analysis

### Documenting the Analysis Environment

Reproducibility requires careful documentation of the analysis environment, including the versions of all software tools, the parameters used, and the reference files. The version of each tool should be recorded, as different versions may produce different results. The parameters should be recorded in a configuration file that is stored with the analysis outputs.

Containerization tools can improve reproducibility by packaging the analysis environment in a single image. Containers ensure that the software versions and dependencies are consistent across different computing environments. The use of containers is particularly important for long-term reproducibility, as software versions change over time and older versions may become unavailable.

### Recording Quality Metrics

The quality metrics from each stage of the workflow should be recorded in a structured format. The metrics should include the number of reads at each stage, the alignment rate, the number of candidate fusions, and the number of fusions passing each filter. These metrics provide a record of the analysis that can be used to troubleshoot problems and to compare results across samples.

The quality metrics should be recorded in a table or spreadsheet that is stored with the analysis outputs. The table should include the sample identifier, the stage of the analysis, the metric name, and the metric value. This format allows the metrics to be queried and compared across samples.

### Version Control for Analysis Scripts

Analysis scripts should be maintained under version control, allowing the history of changes to be tracked. Version control systems such as Git provide a record of when and why each change was made, which is valuable for troubleshooting and for reproducing results. The version control repository should include the analysis scripts, the configuration files, and the documentation of the analysis.

The Carpentries provides lessons on version control with Git that are relevant for researchers who need to manage analysis scripts. These lessons cover the basic commands for tracking changes, committing versions, and collaborating with others. The skills taught in these lessons are directly applicable to managing bioinformatics analysis scripts.

## Common Failure Patterns and Troubleshooting

### Low Alignment Rates

A low alignment rate, defined as the proportion of reads that align to the reference, can indicate several problems. The reference may not include the species or the transcript isoforms present in the sample. The library preparation may have produced reads with adapter contamination or other artifacts that prevent alignment. The sequencing run may have produced low-quality reads that fail to align.

Troubleshooting a low alignment rate begins with examining the quality metrics of the raw reads. If the reads have low quality scores, the problem is likely in the sequencing run. If the reads have high quality but fail to align, the problem is likely in the reference or the alignment parameters. The alignment parameters should be checked to ensure that they are appropriate for the sequencing platform and the expected error rate.

### Excessive Candidate Fusions

A large number of candidate fusions may indicate that the filtering criteria are too permissive or that the library contains many artifacts. Template switching during library preparation can produce chimeric molecules that appear to be fusions. Sequencing errors can create false chimeric junctions when reads misalign to the reference.

The first step in troubleshooting excessive candidates is to examine the read support distribution. If most candidates are supported by a single read, the read support threshold should be increased. If many candidates have breakpoints in intronic or intergenic regions, the exon boundary anchoring criterion should be applied more strictly. The cross-tool comparison can also help identify artifacts, as genuine fusions are more likely to be detected by multiple tools.

### Discordant Results Between Tools

When JAFFAL and LongGF produce different results, the discordance should be investigated instead of ignored. The tools use different algorithms and may have different sensitivities and specificities for different types of fusions. A fusion detected by only one tool may be genuine but may have characteristics that make it difficult for the other tool to detect.

The investigation should examine the supporting reads for the discordant fusion, verifying that the reads genuinely span the junction between the two genes. The breakpoint coordinates should be compared between the tools, and the alignment quality at the breakpoint should be examined. If the fusion is supported by high-quality reads and the breakpoint is clean, the fusion may be genuine even if only one tool detected it.

## Limitations and Interpretation Boundaries

### Error Rates and False Positives

The error rates of long-read sequencing platforms impose fundamental limits on fusion detection accuracy. Even with error correction and filtering, some false positives will remain in the candidate list. The false positive rate depends on the sequencing platform, the error profile of the specific run, and the filtering criteria applied.

The interpretation of fusion detection results should account for the expected false positive rate. Fusions that are detected by multiple tools, supported by multiple reads, and confirmed by visual inspection are more likely to be genuine. Fusions that fail these criteria should be treated as candidates requiring experimental validation before they are reported.

### Detection Sensitivity for Low-Expressed Fusions

Fusions expressed at low levels may not be detected even with deep sequencing. The detection sensitivity depends on the sequencing depth, the expression level of the fusion, and the efficiency of the library preparation. Fusions expressed in a small proportion of cells in a heterogeneous sample are particularly difficult to detect.

The sensitivity limitations should be considered when interpreting negative results. The absence of a fusion in the candidate list does not prove that the fusion is absent from the sample. For clinical applications where the presence or absence of a specific fusion has diagnostic or therapeutic implications, orthogonal methods such as targeted PCR or FISH should be used to confirm negative results.

### Isoform Complexity

Long-read data can reveal that a single genomic rearrangement produces multiple fusion isoforms through alternative splicing. These isoforms may have different functional consequences, and the interpretation of the fusion should consider the full isoform repertoire. The isoform complexity can complicate the interpretation of read support, as reads supporting different isoforms may be counted separately or together depending on the analysis.

The isoform-level analysis of fusions requires careful examination of the read alignments at the fusion junction. Reads that support different isoforms will have different junction sequences, and these differences can be used to reconstruct the isoform repertoire. The functional interpretation of the fusion should consider the predominant isoform and the potential consequences of each isoform.

## Safety and Regulatory Context for Clinical Applications

### Data Privacy and Security

Fusion detection from clinical samples involves the analysis of human genetic data, which is subject to privacy and security requirements. The raw sequencing data and the analysis results must be stored securely, and access must be restricted to authorized personnel. The specific requirements depend on the jurisdiction and the nature of the research or clinical activity.

The NCBI provides resources for the secure storage and sharing of genomic data, including databases that are designed for controlled access. Researchers working with clinical samples should be familiar with the data sharing policies of their institution and the relevant regulatory requirements. The data management plan should address the storage, access, and sharing of the sequencing data and the analysis results.

### Reporting Requirements for Clinical Findings

Fusions detected in clinical samples may have diagnostic, prognostic, or therapeutic implications. The reporting of these findings is subject to regulatory requirements that vary by jurisdiction. In the United States, laboratories that report clinical test results must comply with the Clinical Laboratory Improvement Amendments (CLIA) and may need to obtain approval from the Food and Drug Administration for the test.

The reporting of fusion detection results should include the methods used, the quality metrics, and the limitations of the analysis. The report should clearly state whether the fusion was confirmed by experimental validation and whether the fusion has known clinical significance. The report should also include the criteria used to classify the fusion as pathogenic, likely pathogenic, or of uncertain significance.

### Professional Escalation Criteria

Researchers and laboratory professionals should escalate findings that have potential clinical implications to the appropriate clinical team. The escalation criteria should be defined in advance and should include fusions that are known to be clinically actionable, fusions that are novel but have features suggesting clinical relevance, and fusions that are detected in samples from patients with known or suspected cancer.

The escalation process should include a review of the fusion detection results by a qualified professional, confirmation of the fusion by an orthogonal method, and communication of the findings to the clinical team. The escalation should be documented, and the documentation should include the evidence supporting the fusion and the rationale for the escalation.

## Practical Implementation Steps

### Step 1: Install Required Tools

The workflow requires several software tools that must be installed and configured before the analysis begins. The tools include the aligner, the fusion detection tools, and the utilities for file manipulation and quality assessment. The installation process varies by tool and by operating system, and the tool documentation should be followed carefully.

Bioconductor provides a framework for installing and managing bioinformatics software in the R programming environment. Many fusion detection tools and supporting utilities are available through Bioconductor, and the installation process is streamlined through the Bioconductor package management system. The Bioconductor documentation provides guidance on installing and using these packages.

### Step 2: Prepare the Reference

The reference transcriptome should be prepared from a current genome annotation. The reference files should be downloaded from a reliable source, such as the NCBI, and the transcript sequences should be extracted in FASTA format. The reference should be indexed for the aligner being used, and the index files should be stored with the reference.

The reference preparation should be documented, including the source of the annotation, the version of the annotation, and the commands used to extract the transcript sequences. This documentation is essential for reproducibility and for troubleshooting alignment problems.

### Step 3: Align the Reads

The alignment step should be run with parameters appropriate for the sequencing platform. The alignment output should be sorted and indexed, and the alignment statistics should be recorded. The alignment should be inspected to verify that the alignment rate is acceptable and that the supplementary alignments are present.

The alignment step is computationally intensive and may require significant memory and processing time. The alignment can be run on a high-performance computing cluster for large datasets, and the alignment parameters should be optimized for the specific dataset.

### Step 4: Run Fusion Detection Tools

The fusion detection tools should be run with the aligned BAM file as input. The parameters should be set based on the sequencing platform and the desired sensitivity. The output from each tool should be recorded, and the results should be compared between tools.

The fusion detection step may produce a large number of candidate fusions, and the filtering criteria should be applied to reduce the candidate list to a manageable size. The filtering criteria should be documented, and the filtered results should be stored for downstream analysis.

### Step 5: Filter and Validate Candidates

The candidate fusions should be filtered based on read support, cross-tool agreement, and visual inspection. The filtered candidates should be examined in a genome browser to verify the junction sequence. Candidates that pass the filtering criteria should be considered for experimental validation.

The validation step should be documented, including the methods used and the results obtained. The validation results should be recorded in a format that allows the findings to be reported and shared.

### Step 6: Document and Report Results

The analysis should be documented in a format that allows the results to be reproduced and reported. The documentation should include the software versions, the parameters, the quality metrics, and the results at each stage of the workflow. The documentation should be stored with the analysis outputs.

The reporting of results should follow the standards of the field and the requirements of the institution. The report should include the methods, the results, and the limitations of the analysis. The report should be reviewed by qualified professionals before it is shared.

## Frequently Asked Questions

### What is the difference between JAFFAL and LongGF?

JAFFAL and LongGF are both fusion detection tools for long-read RNA-seq data, but they use different algorithms. JAFFAL identifies chimeric reads with supplementary alignments and clusters them into candidate fusions, applying filtering criteria such as exon boundary anchoring. LongGF uses a graph-based approach that can identify fusions involving more than two genes and incorporates error correction strategies. The two tools have different strengths and limitations, and using both tools in parallel provides a cross-validation strategy that improves confidence in the results.

### How many supporting reads are needed to call a fusion?

The minimum read support threshold depends on the sequencing depth, the error rate of the platform, and the expression level of the fusion. A common threshold is two supporting reads, but this threshold may miss low-expressed fusions. For clinical applications, a lower threshold may be used with additional validation steps. The read support threshold should be documented and applied consistently across all candidates.

### Can fusion detection be performed with Nanopore data?

Yes, fusion detection can be performed with Nanopore direct RNA-seq data. Nanopore data have higher error rates than PacBio data, so the alignment parameters and filtering criteria must be adjusted accordingly. The fusion detection tools JAFFAL and LongGF both support Nanopore data, and the error correction strategies in these tools are designed to handle the higher error rates.

### What is the role of short-read data in fusion detection?

Short-read data can complement long-read data in fusion detection. Short reads provide orthogonal confirmation of candidate fusions and can detect fusions expressed at very low levels that may be missed by long-read data. Combining short and long reads can maximize the detection of fusion isoforms and fusion-expressing cells. The integration of short and long read data requires careful consideration of the different error profiles and detection biases of the two platforms.

### How are fusion breakpoints determined from long-read data?

Fusion breakpoints are determined by examining the alignment of reads that span the junction between two genes. The breakpoint is the position where the read alignment switches from one gene to the other. Long-read data can provide breakpoint information at single-base resolution, which is useful for designing validation experiments and for determining whether the fusion is in-frame.

### What causes false positive fusion calls?

False positive fusion calls can arise from sequencing errors, template switching during library preparation, read chimeras formed during sequencing, and alignment errors. The false positive rate depends on the sequencing platform, the library preparation method, and the filtering criteria applied. Cross-tool comparison, visual inspection, and experimental validation are strategies for reducing false positives.

### How should fusion detection results be validated?

Fusion detection results should be validated by multiple methods. Cross-tool comparison, where the same fusion is detected by both JAFFAL and LongGF, provides computational validation. Visual inspection in a genome browser provides additional confidence. Experimental validation with RT-PCR and Sanger sequencing confirms the fusion junction sequence. For clinical applications, validation in an independent sample or with an orthogonal method is essential.

### What are the limitations of long-read fusion detection?

Long-read fusion detection has several limitations. The error rates of long-read platforms can create false positives, and the detection sensitivity for low-expressed fusions is limited by sequencing depth. The isoform complexity of fusions can complicate interpretation, and the absence of a fusion in the candidate list does not prove that the fusion is absent from the sample. These limitations should be considered when interpreting fusion detection results.

## Related Bioinformatics Guides

- [RNA-Seq Data Analysis Workflow: From Raw Reads to Insights](/knowledge/bioinformatics/rna-seq-data-analysis-workflow-from-raw-reads-to-insights)
- [De Novo Genome Assembly with Long Reads: A Practical Workflow](/knowledge/bioinformatics/de-novo-genome-assembly-with-long-reads-a-practical-workflow)
- [Full-Length Transcript Sequencing: Unraveling Isoform Diversity with Long Reads](/knowledge/bioinformatics/full-length-transcript-sequencing-unraveling-isoform-diversity-with-long-reads)
- [Metabolomics Data Analysis in R: A Practical Workflow](/knowledge/bioinformatics/metabolomics-data-analysis-in-r-a-practical-workflow)
- [RNA-Seq Batch Effect Detection and Correction](/knowledge/bioinformatics/rna-seq-batch-effect-detection-and-correction)

## 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.
- [Transcriptomics in the era of long-read sequencing.](https://pubmed.ncbi.nlm.nih.gov/40155769). Nature reviews. Genetics, 2025.
- [Detection of isoforms and genomic alterations by high-throughput full-length single-cell RNA sequencing in ovarian cancer.](https://pubmed.ncbi.nlm.nih.gov/38012143). Nature communications, 2023.
- [Accurate fusion transcript identification from long- and short-read isoform sequencing at bulk or single-cell resolution.](https://pubmed.ncbi.nlm.nih.gov/40086881). Genome research, 2025.
- [Gene Fusion Detection in Long-Read Transcriptome Datasets from Multiple Cancer Cell Lines.](https://pubmed.ncbi.nlm.nih.gov/39735992). Frontiers in bioscience (Landmark edition), 2024.
- [Gene Fusion Detection and Characterization in Long-Read Cancer Transcriptome Sequencing Data with FusionSeeker.](https://pubmed.ncbi.nlm.nih.gov/36318117). Cancer research, 2023.

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