rewirebio.iobenchmarks

Use caseDNA and genomesResearch and clinical research

Select a tumour RNA fusion-detection workflow

Which RNA-sequencing workflow detects and prioritises tumour gene fusions with useful sensitivity and a manageable false-positive review burden?

What the evidence shows

Four comparisons join the Seraseq reference and paediatric NCH cohort evidence; values cannot be pooled across studies.

Tamura et al. 2026 ran 12 short-read callers on haematologic cancer cell lines; one, Genomon, comes from a co-author's group. On 61 driver fusions in conventional RNA-seq of 170 cell lines, Arriba detected all 61 (TPR 1), FusionCatcher 0.98, STAR-Fusion 0.93, InFusion 0.43 and STARChip 0.41. On 24 driver fusions in targeted RNA-seq of 26 cell lines, Arriba again reached 1 and STAR-Fusion 0.92, InFusion 0.17 and STARChip 0.

A driver fusion entered these truth sets only if at least four callers, including the one scored, found it and it had literature or RT-PCR support, so fusions that most callers miss are under-represented and these values may overstate sensitivity. Against a consensus of the other callers, the number of single-caller calls, a measure of review burden, ranged on conventional RNA-seq from 327 for TrinityFusion-D and 1,061 for STAR-Fusion (PPV 0.39) to 9,567 for Arriba (PPV 0.08), 22,445 for FusionCatcher and 695,295 for Pizzly. On nanopore long reads from 51 paediatric B-ALL samples (Lin et al. 2026), scored on the dominant fusion per sample, JAFFAL reached sensitivity 0.76, LongGF 0.70 and FusionSeeker 0.63, with precision 0.79 to 0.95; the authors' own FUSILLI reached 0.81.

At about one eighth of the depth, sensitivity was 0.09 to 0.27. In an independent long-read benchmark (Dorney et al. 2026 preprint), no caller exceeded F1 0.35 on 400 simulated fusions at 100 Gb and 95% read identity: CTAT-LR-Fusion 0.35 (recall 0.28, precision 0.46), JAFFAL 0.29, Genion 0.20, FusionSeeker 0.14, GFSeeker 0.07 and LongGF 0.04. On all ONT PCR-cDNA reads of the MCF7 cell line, CTAT-LR-Fusion's calls matched 20 of 53 previously reported fusions, with 51 other calls; JAFFAL matched 19 with 38 other calls and FusionSeeker 11 with 359.

The other calls mix real fusions missing from the literature list and false calls. On the Huh7 cell line, pooling six long-read callers per long-read platform and three short-read callers for Illumina, PacBio Kinnex recovered 20 of 34 previously reported fusions, ONT PCR-cDNA 17 and Illumina 15. Against karyotyping, FISH and RT-PCR in 101 adult and paediatric acute leukaemia patients (Kim et al. 2026), an Arriba workflow with manual review detected 40 of 48 conventionally detected fusions (sensitivity 83.3%) and called fusions in 21.4% of 56 conventionally negative results; these count against specificity, although the authors describe several as well-known recurrent fusions and confirmed one by RT-PCR and Sanger sequencing.

These are cell lines, simulations and leukaemia samples: no linked comparison covers solid-tumour specimens, FFPE tissue or novel-fusion discovery, only Kim et al. score against orthogonal clinical assays, none reports interval estimates, and none establishes clinical performance.

You bring
  • Tumour RNA-seq reads: paired-end short reads (conventional or targeted capture) or long reads
  • The intended specimen type (frozen, FFPE or other) and acceptable false-positive review burden
You want
A sourced set of exact evaluated sensitivity/precision values on a synthetic reference standard and a real clinical cohort, with the ascertainment-bias caveat explicit; no patient-level fusion classification.
Evidence in this release
7 comparisons shown · 19 tools · 32 configurations
Kept up to date
Evidence last checked 11 October 2026 · re-checked monthly, next due 11 November 2026
Executive Takeaway

For "Select a tumour RNA fusion-detection workflow", on Validated driver fusions in haematologic cell lines (Tamura et al. 2026), Conventional RNA-seq (Recall), Arriba scored 100%, reported by Comparison of gene fusion detection algorithms reveals frequently overlooked driver fusions in hematologic malignancies.

Best reported valueArribaComparison of gene fusion detection algorithms reveals…
Recall100%Validated driver fusions in haematologic cell lines (Tamura et al. 2026), Conventional RNA-seq
Baseline on this protocolNone scoredNo margin is computed
Ranked tools1211 from independent studies

12 tools have a value on this protocol; 11 of them come from studies whose authors did not develop the tool. No baseline method is scored on this protocol.

Ranking

Ranked on Validated driver fusions in haematologic cell lines (Tamura et al. 2026), Conventional RNA-seq (Recall, higher is better). Only results on this protocol are ranked; other tools are listed below the table.

10 tools in 12 rows, from 2 sources. 11 of 12 results (92%) come from an independent study; 1 is reported by the tool's developers.

Rank▲ModelRecall (higher is better)EvidenceCoverageReported byLicense
#1
Arriba[-] Baseline
100%Source-checked Independent study3/7Comparison of gene fusion detection algorithms…Not recorded
#298%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#3
STAR-Fusion[-] Baseline
93%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#487%Source-checked Independent study4/7Comparison of gene fusion detection algorithms…Not recorded
#584%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#682%Source-checked Developer-reported2/7Comparison of gene fusion detection algorithms…Not recorded
#780%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#867%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#962%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#1061%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#1143%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
#1241%Source-checked Independent study2/7Comparison of gene fusion detection algorithms…Not recorded
13 tools are evaluated elsewhere in this use case but not on this protocol, so not ranked

Baseline

STAR-Fusion and Arriba. Earlier benchmarks rank STAR-Fusion and Arriba as the most accurate short-read fusion callers, as Tamura et al. 2026 report. Both call fusions from chimeric STAR alignments with fixed filters, not a trained classifier.

Machine-learning methods are compared with the baseline on the same data in 2 tables: All fusions against a consensus truth set (Tamura et al. 2026), Validated driver fusions in haematologic cell lines (Tamura et al. 2026). Baseline rows are marked in each table.

Tools compared

Comparisons

Unified model comparison for this use case. Compare tools against protocol baselines in a single view.

Validated driver fusions in haematologic cell lines (Tamura et al. 2026)

Direct evidenceIndependent study: Reported in a study whose authors did not develop the evaluated tool.Developer-reported: Reported by the developers of the evaluated tool.

Recall (higher is better), with the other metrics below each value. The best value in each column is highlighted.

ToolConventional RNA-seqTargeted RNA-seq
ArribaBaselineconventional
100% (best in column)100% (best in column)
FusionCatcherconventional
98%96%
STAR-FusionBaselineconventional
93%92%
JAFFA-directconventional
87%88%
Pizzlyconventional
84%92%
Genomonconventional
82%88%
EricScriptmachine learning
80%83%
TrinityFusion-D3 sweepsconventional
67%83%
InFusionconventional
43%17%
STARChipconventional
41%0%

P precision.

Caveats and method

What each column measures:

  • Conventional RNA-seq: TP, FP, FN, TPR, PPV and F1 for driver fusions on conventional RNA-seq of haematologic cell lines for 12 algorithms
  • Targeted RNA-seq: TP, FP, FN, TPR, PPV and F1 for driver fusions on targeted RNA-seq of haematologic cell lines for 12 algorithms
  • Haematologic cancer cell lines only (170 for conventional, 26 for targeted RNA-seq); no solid tumours, FFPE tissue or clinical specimens. Per-caller results on 165 clinical samples are only in a figure.
  • Circular truth selection: a driver pair entered the truth set only if four or more of the 12 algorithms detected it, counting the algorithm being scored, and it was reported in the literature or confirmed by RT-PCR and Sanger sequencing; 2 pairs found by three algorithms were added from the literature (Methods 'Comparison of detection algorithms'). Driver fusions that most callers miss are therefore under-represented, so TPR may overstate sensitivity, most for callers that agree with the majority.
  • False positives are listed driver fusions detected by one algorithm only; listed driver calls by two or three algorithms that are not in the truth set count as neither TP nor FP, so the driver PPV understates review burden.
  • Truth: 61 driver fusion-cell line pairs (20 unique fusions) from a curated list of 202 haematologic driver fusions.
  • Genomon rows are author_reported: a co-author is first author of the Genomon fusion paper; the other 11 algorithms were developed elsewhere.
  • Default or recommended parameters only; the authors show that parameter changes recover some missed driver fusions (Results).
  • Circular truth selection: a driver pair entered the truth set only if four or more of the 12 algorithms detected it, counting the algorithm being scored, and it was reported in the literature; 2 pairs from the conventional analysis were added (Methods 'Comparison of detection algorithms'). Driver fusions that most callers miss are therefore under-represented, so TPR may overstate sensitivity.
  • Truth: 24 driver fusion-cell line pairs (17 unique fusions); with 24 pairs, one fusion moves TPR by about 0.04.
  • Targeted assay is the authors' own hybridisation-capture panel (44 fusion-target genes and 134 known fusion junctions).

Protocols: Driver fusions, conventional RNA-seq of haematologic cell lines, validated truth (Tamura et al. 2026 Supplementary Table 5), Driver fusions, targeted RNA-seq of haematologic cell lines, validated truth (Tamura et al. 2026 Supplementary Table 5)

Primary benchmark evaluation shown (Validated driver fusions in haematologic cell lines (Tamura et al. 2026)). 6 additional study protocols (NCH clinically relevant fusion ascertainment, EnFusion undiluted Seraseq duplicate benchmark, Long-read fusion callers in paediatric B-ALL (Lin et al. 2026) and others) are cited in individual tool drawer rows and under Question, sources and review.

Not covered yet

  • Table 2 caption (v3) vs the optimisation narrative (v2) is an unresolved source version conflict.
  • STAR-Fusion precision is printed as both 43.6% (Table 2) and 43.8% (paragraph) in the source; both are preserved, neither is silently reconciled.
  • No confidence interval or dispersion is printed for the Seraseq Table 2 rows.
  • No linked comparison of several callers on adult solid tumours or FFPE tissue with printed per-caller values; Br J Cancer 2026 (1,233 FFPE solid tumours) reports Arriba and DRAGEN against panel assays only in figures.
  • Driver-fusion truth sets in Tamura et al. 2026 were selected mostly from fusions called by at least four algorithms, including the algorithm scored, so sensitivity for fusions most callers miss is not measured.
  • Per-caller results on 165 clinical haematologic samples in Tamura et al. 2026 are shown only in a figure.
  • Lin et al. 2026 scores only the dominant fusion among 74 listed B-ALL fusions, so the review burden of long-read callers' full output is not measured.
  • Short- and long-read evidence on the same RNA comes from one cell line (Huh7), with calls pooled over callers per platform (Dorney et al. 2026 preprint); no per-caller comparison across platforms on the same samples is linked.
  • The DREAM SMC-RNA fusion challenge results (Cell Systems 2021) were not retrieved, and Haas et al. 2019 prints per-method accuracy only in figures.
  • No linked study reports interval estimates for per-caller sensitivity or precision.
  • Only one workflow (Arriba with manual review) is scored against orthogonal clinical assays (Kim et al. 2026, acute leukaemia); its specificity counts RNA-seq-only fusions as errors, although some are well-known recurrent fusions and one was confirmed by RT-PCR and Sanger sequencing.
  • Long-read caller precision and recall come from simulated reads only; on cancer cell lines only counts of known and other calls are available.
  • The only machine-learning method compared with STAR-Fusion or Arriba on the same data in the linked evidence is EricScript (Tamura et al. 2026), an alignment-based caller that ranks its candidates with a trained AdaBoost score; no caller that learns fusion detection itself is compared.

Contribute evidence or propose a correction

How to cite

Select a tumour RNA fusion-detection workflow. Rewire biological model benchmark database, release 2026-11-21-adc15377fab2. https://benchmarks.rewirebio.io/use-cases/tumour-rna-fusion-detection/

Release 2026-11-21-adc15377fab2. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.

The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence

Question, sources and review

Who this is for, setting and exclusions

Inspect per-caller sensitivity for validated driver fusions, the single-caller call counts that drive review burden, and the Seraseq and NCH evidence, matched to the intended specimen type and sequencing (short or long reads, conventional or targeted capture), before selecting a workflow and review burden.

Who this is for
  • Clinical researchers scoping the available diagnostic-genomics evidence
  • Computational researchers comparing exact evaluated configurations
Biological setting
A 14-fusion synthetic Seraseq reference standard and a 229-sample paediatric cancer and haematologic cohort (67 ensemble-ascertained clinically relevant fusions) for Arriba, STAR-Fusion and the EnFusion ensemble; 12 short-read callers on conventional and targeted RNA-seq of haematologic cancer cell lines, scored against literature-supported driver fusions and against a consensus of the other callers (Tamura et al. 2026); and four long-read callers on nanopore whole-transcriptome sequencing of paediatric B-ALL at two depths, scored by dominant-fusion classification (Lin et al. 2026).
Clinical research scope
Clinical applicability is bounded: the NCH clinical-cohort sensitivity figures are ascertained against the optimised EnFusion ensemble's own calls, not an independent exhaustive truth set, so a false-negative rate for fusions the ensemble itself missed cannot be derived from this evidence.

Outside this use case

  • Novel-fusion discovery performance (the optimised known-fusion rescue measures recovery of 14 known synthetic fusions, not novel-fusion discovery)
  • Adult solid-tumour cohorts outside the transcribed NCH cohort
  • Non-Seraseq background fusions, all counted false positive in this intake though some may be real endogenous fusions
Evidence collection plan
Evidence collection planCollecting evidence

Already on record: 7 comparisons shown · 19 tools, in the evaluated evidence above. The status above describes only the specific comparison in this plan, which remains open; it does not mean no evidence has been collected. The plan defines a comparison to investigate; it does not establish model performance or suitability.

Comparison question

Which RNA-sequencing workflow detects and prioritises tumour gene fusions with useful sensitivity and a manageable false-positive review burden?

Baselines, outcomes and validation requirements

Baselines to include

  • The conventional/author-introduced workflow measured in the linked primary source(s).

Outcomes to measure

  • The declared endpoint in this use case's active mapping(s); see evidence_gaps for what remains open.

Validation requirements

  • Independent held-out population matched to the intended clinical setting.
  • Qualified human scientific review before any clinical-validation claim.

Next collection task

A dedicated rewire-benchmarks protocol/run task with an independent fusion truth set not ascertained by the same ensemble being evaluated.

Sources, review and downloads

Automated source review · 2026-10-07 · Claude Sonnet AMP-integration worker, bounded transcription of Codex-checked primary values; independently reviewed by Codex (workbench/amp-supervision/primary-review.md, integration-review-corrections.md)

Bounded primary-source transcription, independently Codex-checked. No new model execution, independent experimental reproduction, qualified human scientific review or clinical validation.

Release 2026-11-21-adc15377fab2 · question use-case-tumour-rna-fusion-detection

Use-case input digest bfbb7ab9303becad1e3cb395568c26f41e72ddb67cd1a89a06d9eaa0f76bd95f

Download questions and review metadata (JSON) (gzip) · Verify release checksums (gzip)