Accuracy assessment of fusion transcript detection via read-mapping and de novo fusion transcript assembly-based methods.

Haas BJ, Dobin A, Li B, Stransky N, Pochet N, Regev A. · 2019

Read Accuracy assessment of fusion transcript detection via read-mapping and de novo fusion transcript assembly-based methods. by Haas BJ, Dobin A, Li B, Stransky N, Pochet N, Regev A. (2019) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6802306.txt.

Introduction

Haas and co-authors benchmarked methods for detecting fusion transcripts in RNA sequencing data. Gene fusions drive many cancers. BCR-ABL1 appears in about 95% of chronic myelogenous leukemia, and TMPRSS2-ERG in about half of prostate cancers. Tools to find fusions split into two families: read-mapping methods and de novo assembly methods. The team tested 23 of them, including two of their own.

Simulated and real cancer data were used. STAR-Fusion, Arriba, and STAR-SEQR proved most accurate and fastest overall. The top three all build on the STAR aligner. Assembly-based methods were less accurate at detection. They still earn their place by reconstructing fusion variants and tumor viruses, both important in cancer research.

Catalog record

CREATORHaas BJ, Dobin A, Li B, Stransky N, Pochet N, Regev A.
YEAR2019
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC6802306

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Accuracy assessment of fusion transcript detection via read-mapping and de novo fusion transcript assembly-based methods By Brian J Haas, Alexander Dobin, Bo Li, Nicolas Stransky, Nathalie Pochet, Aviv Regev Published in: Genome Biology DOI: 10.1186/s13059-019-1842-9 ABSTRACT Abstract Background Accurate fusion transcript detection is essential for comprehensive characterization of cancer transcriptomes. Over the last decade, multiple bioinformatic tools have been developed to predict fusions from RNA-seq, based on either read mapping or de novo fusion transcript assembly. Results We benchmark 23 different methods including applications we develop, STAR-Fusion and TrinityFusion, leveraging both simulated and real RNA-seq. Overall, STAR-Fusion, Arriba, and STAR-SEQR are the most accurate and fastest for fusion detection on cancer transcriptomes. Conclusion The lower accuracy of de novo assembly-based methods notwithstanding, they are useful for reconstructing fusion isoforms and tumor viruses, both of which are important in cancer research. Keywords: Fusion, RNA-seq, Cancer, Benchmarking, STAR-Fusion, TrinityFusion BACKGROUND Chromosomal rearrangements leading to the formation of fusion transcripts are a frequent driver in certain cancer types, including leukemia and prostate cancer [1], and contribute to many others [2]. These include BCR–ABL1, found in ~ 95% of chronic myelogenous leukemia (CML) patients [3]; TMPRSS2–ERG in ~ 50% of prostate cancers [4]; and DNAJB1–PRKACA, the hallmark and likely driver of fibrolamellar carcinoma [5]. Determining the driver of a given tumor is important to inform diagnosis and therapeutic strategies. For example, tyrosine kinase inhibitors have been highly effective in the treatment of tumors harboring kinase fusions in leukemia and other cancers [6–9]. Transcriptome sequencing (RNA-seq) has emerged as an effective method to detect fusion transcripts in the precision medicine pipeline. While point mutations and indels can be readily captured from whole exome sequencing (WES), detecting genome rearrangements typically requires whole genome sequencing (WGS). RNA-seq yields the “expressed exome” of the tumor, capturing only the transcriptionally active regions of the genome, and thus provides a cost-effective means to acquire evidence for both mutations and structural rearrangements involving transcribed sequences, which can reflect on functionally relevant changes in the cancer genome. Over the past decade, multiple bioinformatics methods and software tools have been developed to identify candidate fusion transcripts from RNA-seq (reviewed in [10, 11]), with select methods leveraged in recent efforts to build catalogs of fusions across thousands of tumor samples [12, 13]. Following the two general strategies for RNA-seq analyses [14], RNA-seq-based fusion detection falls into two conceptual classes: (1) mapping-first approaches that align RNA-seq reads to genes and genomes to identify discordantly mapping reads that are suggestive of rearrangements and (2) assembly-first approaches that directly assemble reads into longer transcript sequences followed by identification of chimeric transcripts consistent with chromosomal rearrangements (Fig. 1a). Evidence supporting predicted fusions is typically measured by the number of RNA-seq fragments found as chimeric (split or junction) reads that directly overlap the fusion transcript chimeric junction, or as discordant read pairs (bridging read pairs or fusion spanning reads) where each pair of reads maps to opposite sides of the chimeric junction without directly overlapping the chimeric junction itself (Fig. 1a). [Figure: Methods for fusion transcript prediction and accuracy evaluation. a The two general paradigms for fusion transcript identification include (left) mapping reads to the genome and capturing discordant read pairs and chimeric read alignments and (right) performing genome-free de novo transcript assembly followed by identification of chimeric transcript alignments. b Given a well-defined truth set of fusions, true- and false-positive predictions are tallied according to minimum threshold for fusion-supporting reads. F1 accuracy values are computed at each minimum evidence threshold to determine the threshold that yields peak prediction accuracy for each method. Similarly, precision and recall values are computed at each minimum evidence threshold, plotted as a precision-recall curve, and the area under the curve (AUC) is computed as a measure of overall prediction accuracy] Implementations of the various prediction methods vary in the read alignment tools employed, the genome database and gene set resources used, and criteria for reporting candidate fusion transcripts and for filtering out likely false positives. Available fusion predictors vary in prediction accuracy, installation complexity, execution time, robustness, and hardware requirements. Depending on the fusion prediction tool chosen, processing one RNA-seq sample containing tens of millions of reads can take several days’ worth of computing and result in a list of hundreds to thousands of gene fusion candidates, including many likely false positives, with little evidence supporting these predictions. Thus, fast and accurate methods for fusion detection are urgently needed, particularly as RNA-seq is increasingly adopted in precision medicine and clinical diagnostics. Earlier evaluations of fusion prediction methods based on RNA-seq have highlighted the shortcomings of contemporary methods, but were mostly limited to small numbers of fusion candidates, compared few tools, and relied heavily on simulated test data for accuracy assessment [15, 16]. Here, we advance fusion transcript prediction benchmarking to include thousands of fusion transcripts at varied expression levels, devise a strategy to benchmark fusion accuracy leveraging real cancer transcriptome data without a priori defined gold standard truth sets, and test a large number of tools. Specifically, we evaluate 23 different fusion detection methods (from 19 different software packages) that can leverage solely RNA-seq as input for fusion transcript detection (Table 1, Fig. 1b). This includes two software packages that we developed, STAR-Fusion and TrinityFusion: STAR-Fusion leverages chimeric and discordant read alignments identified by the STAR aligner [18, 51, 52] to predict fusions and TrinityFusion leverages chimeric reads and Trinity de novo transcriptome assembly [48, 53] to reconstruct fusion transcripts and identify fusion candidates. We assess each method, evaluating sensitivity and specificity of fusion detection, and identify those methods best suited for accurate fusion detection from cancer RNA-seq. [Table: RNA-seq-based fusion transcript predictors evaluated] RESULTS A panel of methods for fusion transcript detection We assessed 23 methods for fusion transcript detection, including 18 methods primarily based on read-alignments (Table 1): Arriba [17], ChimeraScan [19], ChimPipe [21], deFuse [24], EricScript [25], FusionCatcher [28], FusionHunter [30], InFusion [31], JAFFA-Direct [32], MapSplice [34], nFuse [36], Pizzly [37], PRADA [39], SOAPfuse [40], STARChip [42], STAR-Fusion, STAR-SEQR [44], and TopHat-Fusion [45], and four methods primarily based on transcript assembly: JAFFA-Assembly [32] and three execution modes of TrinityFusion: TrinityFusion-C, TrinityFusion-D, and TrinityFusion-UC. An additional assessed method, JAFFA-Hybrid [32], leverages a combination of both read mapping and de novo assembly approaches. For each method, we used its own recommended alignment and analysis strategy and parameters, as implemented in its respective package (Table 1). We benchmarked each method using simulated data and real RNA-seq from cancer cell lines. In certain cases, we assessed methods in either alternative execution modes, or according to assigned fusion confidence levels (see the “Methods” section). For example, we assessed Arriba using either all predicted fusions or restricting to only those that Arriba labeled as high confidence predictions (Arriba_hc). We assessed TrinityFusion in each of its three alternative execution modes, involving assembly of all input reads (TrinityFusion-D), only chimeric reads (TrinityFusion-C), or both unmapped and chimeric reads (TrinityFusion-UC). We assessed accuracy using both strict and lenient scoring criteria: while strict scoring relied on the pair of gene symbols corresponding to the genes predicted to be fused, lenient scoring also allowed for likely paralogs to serve as acceptable proxies for fused target genes. We show lenient scoring here

Read the complete text