henhaobio
Sign in

bioRxivBioinformaticsPreprint

STAR-Fusion: Fast and Accurate Fusion Transcript Detection from RNA-Seq

Brian J. HaasiD, Alexander DobiniD, Nicolas StranskyiD, Bo LiiD, Xiao Feng YangiD, Timothy L. TickleiD, Asma Bankapur, Carrie L. GanoteiD, Thomas G. DoakiD, Nathalie L.M.M. PochetiD, Jing Yi SuniD, Catherine J. WuiD, T GingerasiD, Aviv RegeviD

PDFDOI
In-depth full-text digest available · Jump to digest ↓

Abstract

MotivationFusion genes created by genomic rearrangements can be potent drivers of tumorigenesis. However, accurate identification of functionally fusion genes from genomic sequencing requires whole genome sequencing, since exonic sequencing alone is often insufficient. Transcriptome sequencing provides a direct, highly effective alternative for capturing molecular evidence of expressed fusions in the precision medicine pipeline, but current methods tend to be inefficient or insufficiently accurate, lacking in sensitivity or predicting large numbers of false positives. Here, we describe STAR-Fusion, a method that is both fast and accurate in identifying fusion transcripts from RNA-Seq data.\n\nResultsWe benchmarked STAR-Fusions fusion detection accuracy using both simulated and genuine Illumina paired-end RNA-Seq data, and show that it has superior performance compared to popular alternative fusion detection methods.\n\nAvailability and implementationSTAR-Fusion is implemented in Perl, freely available as open source software at http://star-fusion.github.io, and supported on Linux.\n\nContactbhaas@broadinstitute.org

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (24 Mar 2017) · CC BY-ND 4.0

Filtering STAR's chimeric alignments against gene annotations outscored 15 other fusion callers on 60 cancer cell lines and ran fastest, averaging 2.3 single-core hours per sample.

Fusion genes drive many cancers and RNA-seq captures the expressed ones, but existing callers can need days per sample and return hundreds to thousands of mostly false candidates. STAR-Fusion builds on the split and discordant read alignments the widely used STAR aligner already reports, adding annotation mapping and artifact filters, and a 'kickstart' mode can reuse chimeric output from STAR runs in existing pipelines.

Key results

  1. ShownOn 60 cancer cell line transcriptomes, STAR-Fusion had the highest area under the precision-recall curve (AUC) of all callers compared, with ChimPipe second, and stayed first under every alternative truth-set definition the authors tried.

    Truth set: fusions called by at least four methods; calls unique to one method counted as false positives; calls shared by 2 or 3 methods left ambiguous. The ranking held when truth required agreement of 2 to 5 methods, when ambiguous calls were ignored or scored as false, and when paralog hits counted as true. The AUC values themselves are in figures and Supplementary Tables 9-10, not in this text. · STAR-Fusion accurately and sensitively predicts…; Fusion prediction accuracy computation

  2. ShownSTAR-Fusion was the fastest pipeline tested; most of its run time is the STAR alignment itself, and mapping candidates to annotations plus filtering adds only minutes.

    Each cell line sample (20 M paired-end reads) was run as a separate job on the Broad Institute compute grid. STAR-Fusion averaged 2.3 hours per sample on one core and about 30 minutes with 10 cores, while the other callers took from hours to days; the mapping and filtering steps averaged 9 minutes per sample. Per-tool times for the other callers are not given in this text. · STAR-Fusion prediction pipeline is fast

  3. SuggestedOn simulated data with known fusions most callers did well and the leaders differed only slightly, but STAR-Fusion had the best accuracy and stayed sensitive across the full range of fusion expression levels.

    Ten simulated samples, each with 500 fusions spread over a wide expression range and 30 M paired-end reads; half had 50-base and half 101-base reads. Most callers improved with 101-base reads, missed fusions were mainly the lowest expressed, and JAFFA-Assembly lost sensitivity even at the highest expression. Accuracy values are in Supplementary Tables 4-8, not in this text. · STAR-Fusion accurately and sensitively predicts…; Simulated fusion transcripts and RNA-Seq data

  4. SuggestedAccuracy hinges on the minimum read-support cutoff: STAR-Fusion peaked at 3 supporting fragments and did better across a range of cutoffs, whereas TopHat-Fusion peaked at 5 and fell steeply at 3.

    On the cell line data, requiring at least 3 supporting fragments per 20 M reads, equivalent to 0.15 fusion fragments per million total fragments (FFPM), appeared to give near-optimal accuracy for the well-behaved callers. TopHat-Fusion's drop at 3 came from an influx of false positives with few added true positives. · STAR-Fusion accurately and sensitively predicts…

  5. ShownDoing well on simulated data did not guarantee doing well on real data: callers varied far more on cell lines, mostly in their false positives, and in the simulations most false positives were paralogs of the true partner genes.

    At each caller's peak F1 on the cell lines, calls made by only one method ranged from tens to over a thousand; nFuse, a top performer on simulated data, lacked predictive power on cell lines. In simulations, reporting a paralog of a true partner (A'-B alongside or instead of A-B) explained most false positives; scoring paralogs as correct raised every method's accuracy, with STAR-Fusion still on top. · STAR-Fusion accurately and sensitively predicts…; STAR-Fusion accurately and sensitively predicts…

Methods, data and reproducibility

STAR is run with fusion-tuned chimeric alignment settings; split reads and discordant pairs are mapped to reference exons with interval trees, then candidates are filtered on read support, breakpoint alignment length, isoform dominance, BLAST similarity between partners and promiscuous partners. Accuracy and speed were compared with 15 published callers on simulated and cancer cell line RNA-Seq.

  • Simulated paired-end Illumina RNA-Seq with random fusion transcripts (Fusion Simulator Toolkit, RSEM read simulation) · human · n = 10 samples, each with 500 fusions and 30 million PE reads
  • Illumina Human Body Map 2.0 RNA-Seq, used as the template for the 50-base simulated reads · human · E-MTAB-513
  • Cancer Cell Line Encyclopedia (CCLE) RNA-Seq, used as the template for the 101-base simulated reads · human
  • Public paired-end RNA-Seq of cancer cell lines: 52 from CCLE plus BT474, KPL4, MCF7, SKBR3, VCaP, LC2/ad, H2228 and K562, each subsampled to 20 million paired-end reads · human · n = 60 cell lines
Methods & tools
STAR-Fusion, STAR, BLAST+ (blastn), RepeatMasker, RSEM (rsem-simulate-reads), Fusion Simulator Toolkit, Set::IntervalTree (Perl CPAN), Berkeley DB, Gencode Release 19, Trinity Cancer Transcriptome Analysis Toolkit Galaxy Portal, Compared callers: ChimeraScan, ChimPipe, deFuse, EricScript, FusionCatcher, FusionHunter, InFusion, JAFFA (Assembly, Hybrid, Direct), MapSplice, nFuse, PRADA, SOAPfuse, TopHat-Fusion
Reagents & environment
Perl, supported on Linux; uses Perl CPAN module Set::IntervalTree and a Berkeley DB of all-vs-all BLAST hits; the Genome Resource Lib used here was built from Gencode Release 19 (libs provided for human and mouse); STAR, RepeatMasker and blastn command lines printed; software versions not stated.
Assay conditions & compute
Broad Institute compute grid: on average 2.3 hours per sample of 20 M paired-end reads on one core, about 30 minutes with 10 cores; annotation mapping plus filtering 9 minutes per sample. Genome indexing used 4 threads and --limitGenomeGenerateRAM 40419136213.
Reproducibility gaps
Code, documentation, benchmark data and scripts, simulated data and sampled cell-line reads are said to be available, but every link was lost in extraction; the Galaxy portal has no address here. Tool parameters and data sources are in supplementary tables not included.
Limitations (5)
  • No perfect truth set exists for the cell lines, so accuracy was scored against agreement among callers; other truth definitions could change the rankings, although STAR-Fusion stayed first in the variants tested. · stated by the authors
  • Prediction accuracy depends strongly on the minimum supporting-evidence threshold chosen for split reads and spanning fragments. · stated by the authors
  • On simulated data, most fusions the callers missed had the lowest expression levels and little or no read evidence, and paralogs of true partners produced most false positives. · stated by the authors
  • The benchmark was designed and run by the STAR-Fusion authors, and a consensus truth set can favour callers whose behaviour resembles most of the compared methods. · noted in this summary
  • Every simulated fusion joins exons at consensus splice sites, so the simulation does not test breakpoints that fall inside exons or introns. · noted in this summary

Read this if you call gene fusions from tumour RNA-seq and want a fast STAR-based caller, a benchmark against 15 alternatives, and guidance on read-support cutoffs.

Citations
284

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMohebbi '21Tan '20Dorney '26Melsted '17Davidson '15Nicorici '14
75% semantic matchBioinformatics
Hamidreza Mohebbi, Nurit Haspel·bioRxiv·Nov 17, 2021· 1 citations