TL;DR
This paper introduces scFusion, a novel computational tool for detecting gene fusions at the single-cell level using scRNA-seq data. By combining a statistical model and a deep learning approach, scFusion significantly improves accuracy and reduces false positives compared to existing methods, enabling the identification of cell subtypes associated with specific gene fusions.
What Problem Does This Paper Address?
Gene fusions are crucial drivers of cancer development and progression, and their detection is vital for diagnosis and targeted therapy. While bulk RNA sequencing has been effective for identifying gene fusions, the advent of single-cell RNA sequencing (scRNA-seq) offers the potential to uncover fusions within individual cells, revealing cellular heterogeneity and identifying rare cell populations with unique fusion profiles. However, scRNA-seq data is inherently noisy and prone to technical artifacts that can lead to a high rate of false positive fusion discoveries, making accurate detection challenging. Existing bulk RNA-seq fusion detection tools are not optimized for the unique characteristics of scRNA-seq data, necessitating a new approach.
Methods
The core of scFusion lies in its two-pronged approach to accurately identify gene fusions from noisy scRNA-seq data:
- Statistical Model: scFusion employs a zero-inflated negative binomial (ZINB) statistical model. This model accounts for the high proportion of "zero" counts (no supporting reads) common in scRNA-seq and models the distribution of supporting reads for both true fusions and technical artifacts. It uses gene expression levels of partner genes and local GC content as factors influencing read counts. A resampling scheme is used to generate p-values for fusion candidates, controlling for false discovery rates.
- Deep Learning Model (bi-LSTM): To further filter out technical artifacts, scFusion utilizes a bidirectional Long Short-Term Memory (bi-LSTM) neural network. This model is trained on subsequences of chimeric reads, with "negative" training data representing known technical artifacts and "positive" training data generated from random read pairs (as a proxy for true fusions). The bi-LSTM learns to distinguish between genuine fusion signals and artifacts based on sequence features.
Datasets
The study utilized several datasets for training and evaluation:
- Simulated Data: scFusion was evaluated on simulated single-cell RNA-seq data generated using FusionSimulatorToolkit. This simulation incorporated chimeric reads from real scRNA-seq data to better mimic real-world technical artifacts.
- Real scRNA-seq Data:
- A spike-in scRNA-seq dataset (126 cells) where known gene fusions were experimentally introduced.
- A T cell dataset (2355 cells) from patients with liver cancer.
- A multiple myeloma (MM) dataset (597 cells) from 15 patients.
- Publicly Available scRNA-seq Datasets: Six publicly available scRNA-seq datasets were used to evaluate the performance of the bi-LSTM model.
- PCAWG Study Data: Gene fusions reported in the Pan-Cancer Analysis of Whole Genomes (PCAWG) studies were used as a gold standard for evaluating the bi-LSTM's ability to distinguish true fusions from artifacts.
Key Insights and Ideas
- Technical artifacts are a major hurdle: The paper strongly emphasizes that technical noise and artifacts are the primary challenge in single-cell fusion detection, often overwhelming true biological signals.
- Combining statistical and deep learning is powerful: The synergistic approach of using a statistical model to filter initial candidates and a deep learning model to refine artifact detection proves highly effective.
- scFusion can identify rare cell populations: The tool's ability to detect fusions in specific cell subtypes, like MAIT cells, opens doors for understanding their unique biological roles.
- Bulk methods over-report fusions in normal cells: In non-malignant cell types (like T cells), bulk methods detected a significantly higher number of fusions, suggesting a high false positive rate, whereas scFusion reported a more biologically plausible number.
Key Results
- High Sensitivity and Low False Discovery Rate (FDR): scFusion demonstrated high sensitivity in detecting both simulated and experimentally introduced fusions while maintaining a low FDR.
- Superior Precision and F-score: In simulations, scFusion achieved significantly higher precision and F-scores compared to popular bulk RNA-seq fusion detection methods. For example, in a simulation with 1000 cells and 3 million reads/cell, scFusion achieved a precision of 0.94 and F-score of 0.93, compared to bulk methods ranging from 0.13-0.70 and 0.23-0.72, respectively.
- Computational Efficiency: scFusion was computationally more efficient than most bulk methods, taking about half the time of Arriba and significantly less time than STAR-Fusion, EricScript, and FusionCatcher.
- Detection of Novel Cell Subtypes: In a T cell dataset, scFusion identified invariant TCR gene recombinations in Mucosal-associated invariant T (MAIT) cells, which are difficult to detect with bulk methods. In a multiple myeloma dataset, it successfully identified the recurrent IgH-WHSC1 fusion.
- Bi-LSTM Performance: The bi-LSTM model achieved a median AUC of 0.884 and AUPR of 0.913 across various scRNA-seq datasets, effectively filtering chimeric artifacts.
Limitations and Future Work
- Simulating Technical Artifacts: The authors acknowledge that accurately simulating the complex patterns of technical chimeric artifacts remains a challenge, and they bypassed this by incorporating real chimeric reads into their simulations.
- Unknown Mechanisms of Artifacts: The precise mechanisms generating many technical artifacts are still unknown, which limits the ability to fully predict and model them.
- Potential for Missing Rare Fusions: While scFusion is sensitive, it might miss fusions that occur in very few cells or have very low expression levels, as seen with the CCDC6-RET fusion in the spike-in experiment.
- Future Applications: The authors envision scFusion's application in defining cellular subtypes, identifying drug-resistant subclones, and understanding tumor heterogeneity and response to therapies.
Summary updated