TL;DR
This paper introduces CTAT-LR-Fusion, a new computational tool that significantly improves the detection of gene fusions, particularly from long-read RNA sequencing data. By leveraging the full-length transcript information from long reads, CTAT-LR-Fusion offers higher sensitivity and accuracy than existing methods, especially for single-cell applications, and can be further enhanced by integrating short-read data.
What Problem Does This Paper Address?
Gene fusions are critical drivers of cancer, and their accurate identification is essential for diagnosis, prognosis, and targeted therapies. While short-read RNA sequencing (RNA-seq) has been the standard for fusion detection, it struggles to resolve complex fusion isoforms and can be less sensitive, especially in single-cell samples where read lengths are often limited. Long-read RNA sequencing offers the potential for higher resolution and sensitivity by capturing full-length transcripts, but dedicated computational tools and robust benchmarking were lacking. This paper addresses this gap by developing and validating CTAT-LR-Fusion, a tool specifically designed for long-read fusion detection, and demonstrating its superiority over existing methods.
Methods
The core of this work is the development of CTAT-LR-Fusion, a two-phase computational pipeline:
- Phase 1: Candidate Identification: This phase uses a modified minimap2 aligner (ctat-minimap2) to rapidly identify potential chimeric long reads by only reporting alignments to multiple genomic loci. These candidate reads are then filtered based on proximity to reference gene annotations and alignment identity.
- Phase 2: Vetting and Quantification: This phase models fusion gene contigs and realigns candidate reads to these contigs using minimap2. It then quantifies fusion transcript breakpoints and expression levels. For samples with both long and short reads, FusionInspector is integrated to leverage short-read evidence for further validation and quantification.
The authors also developed a comprehensive benchmarking framework that includes simulated data with varying error rates and sequence identities, as well as real-world datasets from control samples, cancer cell lines, and single-cell transcriptomes.
Datasets
The authors utilized several datasets for benchmarking and validation:
- Simulated Data:
- PacBio and ONT long reads simulated with varying error rates (75%-95% sequence identity) using Badread.
- High-fidelity PacBio HiFi and ONT R10.4.1 reads simulated using PBSIM3 at 5x and 50x coverage.
- Real Data:
- A commercial reference RNA sample (SeraCare Fusion RNA Mix v4) containing 16 spiked-in oncogenic fusion transcripts.
- Bulk transcriptomes from nine cancer cell lines.
- Single-cell RNA-seq data from a T cell-infiltrated melanoma tumor sample.
- Single-cell RNA-seq data from three metastatic high-grade serous ovarian carcinoma (HGSOC) samples.
- Matched Illumina short-read RNA-seq data for most of the real datasets, used for comparative analysis and integration.
Key Insights and Ideas
- Long reads unlock higher sensitivity: The ability of long reads to capture full-length transcripts dramatically improves fusion detection sensitivity and isoform resolution compared to short reads, especially in single-cell contexts.
- CTAT-LR-Fusion excels: The new tool demonstrates superior accuracy and recall across various simulated and real-world datasets, outperforming existing long-read fusion callers.
- Order matters: Accurately identifying the order and orientation of fused genes is crucial for distinguishing true oncogenic fusions from artifacts, a feature well-handled by CTAT-LR-Fusion.
- Short reads still valuable: While long reads offer higher sensitivity, combining them with short reads in CTAT-LR-Fusion can further maximize detection and provide complementary evidence, especially for certain fusion isoforms.
- Single-cell fusion detection is feasible: The study demonstrates the power of long-read sequencing for identifying fusion transcripts at the single-cell level, revealing tumor heterogeneity and potential driver fusions.
Key Results
- Benchmarking Accuracy: CTAT-LR-Fusion consistently showed high precision and recall, often outperforming other long-read fusion callers like JAFFAL, LongGF, and FusionSeeker in simulated data, especially when considering gene pair ordering.
- SeraCare Control Sample: CTAT-LR-Fusion successfully detected all 16 spiked-in control fusions, while short-read methods detected few or no supporting reads in down-sampled data, highlighting the advantage of long reads.
- Cancer Cell Lines: CTAT-LR-Fusion demonstrated robust fusion detection across nine cancer cell lines, identifying known oncogenic fusions and showing high agreement with other methods when using a "wisdom of the crowds" approach for truth set definition.
- Single-Cell Applications: In melanoma and HGSOC single-cell data, CTAT-LR-Fusion identified more fusion-expressing tumor cells and fusion isoforms than short-read methods, revealing tumor heterogeneity. For instance, in one HGSOC sample, CTAT-LR-Fusion detected four somatic fusion transcripts, with some detected solely by long reads.
- ONT Data: CTAT-LR-Fusion also showed promise for detecting fusions from ONT long-read sequencing data.
Limitations and Future Work
- Reference Dependence: CTAT-LR-Fusion requires a reference genome annotation and primarily identifies fusions involving known exons, potentially missing fusions with novel exon usage or entirely intergenic breakpoints.
- Paralog Filtering: By default, the tool filters out fusions involving paralogous genes, which can be a limitation if such fusions are of interest.
- Computational Cost: CTAT-LR-Fusion is computationally more intensive than some short-read methods, taking approximately 50% longer per transcriptome.
- Simulated Data Limitations: Simulated datasets, while useful, may not fully capture the complexity of real-world sequencing data, including background noise and novel fusion structures.
- Future Work: The authors suggest further optimization of the tool and continued evaluation with more diverse datasets. They also highlight the potential for long-read sequencing to uncover novel fusion isoforms and their role in cancer immunotherapy.
Summary updated