Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

FusionQ: a novel approach for gene fusion detection and quantification from paired-end RNA-Seq.

BMC Bioinformatics · 2013
L1 65/100 3/4
Why this verdict

The main results reproduced, with only marginal, non-material deviations.

Reproduced on the brainbox compute brainarbeit.com
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴
✓ What held up
  • Nothing in this column.
What did not (or only partly)
  • 🟡Could not use the authors’ exact input data
  • 🔴Reported values were only indirectly comparable
  • 🔴A deviation arose in the data or preprocessing
  • 🔴A deviation was attributed to the published material
  • 🟡Reported values were not (fully) derivable from the shared data
  • 🔴The deviation was non-trivial in magnitude
  • 🔴The central claim did not (fully) hold under reproduction
  • 🔴Overall, the reproduction showed a material discrepancy
How its reproducibility compares
65/100
Reproducibility score
0.5 SD below mean
vs. all fields · 1173 studies
🎯 Scores higher than 27% of all assessed papers rank 843 of 1173 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

FusionQ (recovered via Wayback Machine, not from the assigned GitHub URL which is only the wgsim read-simulator) was successfully built and made to run completely cleanly end-to-end after 2 source-code fixes (prior segment) and 3 environment fixes this segment (broken bundled samtools binary, missing LD_LIBRARY_PATH for libboost_thread, and a samtools-CLI-version incompatibility requiring legacy samtools 0.1.19). On the shipped ground-truth smoke-test data (wgsim-simulated reads encoding a known APOD-KCTD7 fusion), the pipeline runs error-free and correctly flags the two expected candidate genes upstream, but calls zero fusions in any output file -- a clean algorithmic mismatch, not an environment artifact. wgsim's stated 50bp read length was confirmed exact; its fragment-length distribution (mean 164/SD 48) could not be verified from any recovered artifact and is graded partial. SRP003186's 7 SRA runs were confirmed fully present and integrity-verified but were not run through the pipeline (Table 3 real-data reproduction was an explicit stretch goal, not attempted, given scope/time already spent on environment repair). No claim in this report was fabricated beyond what shipped logs/data support; no self-invented deadlines caused any premature stop.

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-07-29
Rubric version
v1.0
Assessed by
🤖 AI curator · claude (ai-curator room) · v1.0 · run #1 2026-07-31
no human curator yet
Last updated
2026-07-31

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: opus
Founding hypothesis

Can gene fusions be detected from paired-end RNA-Seq data with both high sensitivity and specificity while also reconstructing chimerical transcript structures and estimating their abundances? The authors propose that extending short residual read segments by aggregating overlapping reads ("residual sequence extension"), applying a filter list, and quantifying whole chimerical transcripts by an EM algorithm with sparse optimization will outperform existing split-read/spanning-pair tools.

Core claims
  • FusionQ is a novel tool that detects gene fusions, constructs chimerical transcript structures, and estimates their abundances from paired-end RNA-Seq data. resource
  • The "residual sequence extension" approach, which lengthens short read segments by merging overlapping (reference) reads, confirms alignment on both sides of a fusion point and largely decreases the false-positive rate. method
  • Chimerical transcript abundance can be estimated with the RSEM framework using an EM algorithm with an L1-norm (group sparsity) constraint, and this quantification further improves fusion identification accuracy because near-zero-abundance fusions can be discarded as misalignments. method
  • FusionQ exceeded deFuse and TopHat-Fusion in both sensitivity and specificity on simulated data, especially for low-coverage fusion detection. finding
  • A filter list (read number, sequence similarity, read position distribution) controls the false-positive rate and gives results with high specificity. method
  • Constructing the reference transcriptome directly from the RNA-Seq data (Cufflinks with guiding reference annotation) lets FusionQ report fusions involving novel genes/isoforms, not only annotated genes. method
  • On breast cancer cell line RNA-Seq data, FusionQ detected both previously reported and new fusions, reported their structures and expressions, and showed better specificity and sensitivity than the other two tools; some highly expressed detected fusion genes are important breast cancer biomarkers. finding
  • Read-count-based fusion quantification used by existing tools only reflects expression of the small region around the fusion point, so quantification should be based on the whole chimerical transcript. mechanism
Experimental setups
Assay System Perturbation Readout Platform
Simulated paired-end RNA-Seq benchmark (fusion detection: sensitivity/specificity comparison of FusionQ vs deFuse vs TopHat-Fusion) 50 simulated datasets generated with "Eric the Simulator"; each dataset = 2 million "background" reads from published RNA-Seq of untreated human pulmonary microvascular endothelial cells plus "broken exon" (BE) fusion reads in silico spike-in of 50 simulated gene fusions per dataset at coverage 1-50 detection sensitivity and specificity of simulated fusions Eric the Simulator (read simulation); FusionQ, deFuse, TopHat-Fusion (detection)
Paired-end RNA-Seq fusion detection and quantification on cancer cell lines Four breast cancer cell lines: BT474, SKBR3, KPL4, MCF7 (published dataset NCBI SRA SRP003186), known to contain 27 gene fusions none (cancer cell lines as-is) detected fusion genes (reported vs novel), chimerical transcript structures, and fusion expression levels; specificity/sensitivity vs deFuse and TopHat-Fusion NCBI SRA: SRP003186
Paired-end RNA-Seq control comparison RNA-Seq data from normal tissue (control group within the SRP003186 dataset) none (normal tissue control) fusion calls in normal tissue used as control group NCBI SRA: SRP003186
Read-mappability test (in silico) supporting the >50bp read-length restriction 10,000 randomly chosen 25bp reads mappable to the constructed transcriptome none fraction of 25bp reads mapping uniquely vs to more than three genes Bowtie / Bowtie-build index of Cufflinks-constructed transcriptome
Read-position distribution filter calibration (Kolmogorov-Smirnov test) A large number of randomly selected "real" transcripts with both read ends mapped to the transcriptome, pooled as "background positions" none K-S test p-value distribution for supporting-read positions; fusions discarded if either end's p-value falls below the significance level Kolmogorov-Smirnov test
Paralog/homology similarity filter Every pair of reported partner isoforms from detected fusion junctions none pairwise sequence similarity (percent match) between fusion partner isoforms EMBOSS Water
Transcriptome assembly for the fusion-detection reference RNA-Seq reads of the analyzed sample, assembled with reference annotation guiding assembly none constructed transcript library containing reference transcripts plus novel isoforms Cufflinks; Bowtie-build
Chimerical transcript abundance quantification Detected chimerical transcripts and all paired-end reads of the analyzed dataset none estimated transcript abundances (theta) maximizing the read sampling likelihood RSEM framework with EM algorithm plus L1-norm sparsity constraint (Lagrange multiplier M-step)
Key results
  • On simulated data FusionQ exceeded both deFuse and TopHat-Fusion in sensitivity and specificity, with the advantage most pronounced for low-coverage fusions.
  • On breast cancer cell line data FusionQ showed better specificity and sensitivity than deFuse and TopHat-Fusion.
  • FusionQ detected both previously reported and new fusions in the four breast cancer cell lines and reported their structures and expression levels.
  • Some highly expressed fusion genes detected by FusionQ are important biomarkers in breast cancer.
  • Almost 75% of randomly chosen 25bp reads mapped uniquely to the constructed transcriptome, and fewer than 5% mapped to more than three genes, justifying seed-based splitting for reads longer than 50bp. ~75% unique; <5% mapped to >3 genes
  • Residual sequence extension makes short residual sequences longer, raising the probability of unique mapping and increasing detection specificity.
  • EM-based abundance estimation maps reads to optimized positions, so fusions with near-zero estimated expression can be identified as misalignment artifacts and discarded, further improving identification accuracy.
  • Fusion candidates whose partner isoforms share high sequence homology are removed by the similarity filter, and fusions with atypical supporting-read position distributions are removed by the K-S distribution filter.
Key statistics
  • count 50 simulated paired-end RNA-Seq datasets, each containing 50 simulated gene fusions with coverage from 1 to 50 (Simulation design using "Eric the Simulator")
  • count two million "background" reads per simulated dataset (Background reads drawn from published RNA-Seq of untreated human pulmonary microvascular endothelial cells)
  • count 27 gene fusions known to be contained in the four breast cancer cell lines (BT474, SKBR3, KPL4, MCF7 (SRA: SRP003186))
  • other almost 75% aligned uniquely (Randomly chosen 25bp reads aligned to the constructed transcriptome)
  • other less than 5% of reads mapped to more than three genes (Same 25bp read mappability test)
  • other similarity more than 50% → fusion discarded (EMBOSS Water similarity filter between reported partner isoforms)
  • pvalue significance level 0.025 (K-S test read-position distribution filter; fusion discarded if either end's p-value is significantly low)
  • other read lengths restricted to more than 50bp; seeds allowed no more than one mismatch; 20bp segment with the fusion point in the middle used as contig; residual sequences shorter than 10bp handled by the 20bp containing segment (FusionQ parameter constraints for splitting reads and residual sequence extension)

Statistical methods review

Model: sonnet

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

This is a bioinformatics software/methods paper rather than a classical hypothesis-testing biology study: the FusionQ gene-fusion detection and quantification tool is evaluated on 50 simulated paired-end RNA-Seq datasets and on a public breast-cancer cell-line RNA-Seq dataset (with a normal-tissue comparator), benchmarking detection performance (sensitivity and specificity) against two other tools, deFuse and TopHat-Fusion. Within the FusionQ algorithm itself, a Kolmogorov-Smirnov (K-S) test is used as one of several heuristic filters, comparing the distribution of supporting-read positions for a candidate fusion transcript to an empirically derived background distribution, and a candidate fusion is discarded if its K-S p-value falls below a fixed significance level (0.025) modeled from the empirical p-value distribution. Results in the provided text are reported mainly as detection-performance outcomes (sensitivity/specificity, expression estimates via an EM algorithm) rather than through classical inferential statistics with dispersion measures, exact p-values, or corrected multiple-comparison thresholds.

Replicationunclear Sample size50 simulated paired-end RNA-Seq datasets, each containing 50 simulated gene fusions with coverage ranging from 1x to 50x; no formal power/sample-size calculation described GroupsFusionQ vs. deFuse vs. TopHat-Fusion, evaluated on simulated fusion data and on real breast-cancer cell-line RNA-Seq data (BT474, SKBR3, KPL4, MCF7) with a normal-tissue control group Pairingunclear Randomization/blindingnot stated Dispersionnone Exact p-valuesno Multiplicity correctionnone stated; a fixed significance threshold (0.025) is applied to each fusion's K-S p-value based on an empirically modeled p-value distribution, rather than a formal multiple-testing correction
Statistical tests used
Test Applied to n Assumptions
Kolmogorov-Smirnov (K-S) test Distribution filter comparing read-position distribution of a candidate fusion transcript to a 'background positions' distribution, per detected fusion not stated
Approaches that could also have been used
  • Candidate fusions are filtered using a K-S test p-value against a fixed empirical significance threshold (0.025), applied separately to each candidate fusion, without a stated multiple-testing correction.
    Could also: A formal false-discovery-rate procedure (e.g., Benjamini-Hochberg) or a permutation-based empirical FDR could also be applied across the full set of candidate fusions tested. — Since many candidate fusions are being tested simultaneously, an FDR-controlling approach would explicitly bound the expected proportion of false positives across the whole candidate set, which is a common practice when many similar statistical tests are performed in parallel on genomic features.
  • FusionQ's sensitivity and specificity on the 50 simulated datasets are compared against deFuse and TopHat-Fusion as point estimates/performance curves, without an inferential test reported in the excerpted text.
    Could also: Because all three tools are run on the same 50 simulated datasets, a paired comparison (e.g., McNemar's test on matched detection outcomes, or a paired t-test/Wilcoxon signed-rank test on per-dataset sensitivity/specificity) could also be used. — A paired test exploits the matched-dataset structure and would let readers judge whether the observed performance differences between tools exceed what could arise from dataset-to-dataset variability alone.
  • Sensitivity and specificity across the 50 simulated datasets appear to be summarized without an accompanying measure of variability (e.g., across the 50 replicate datasets or across the coverage range 1-50x).
    Could also: Reporting mean ± SD, or a 95% confidence interval, of sensitivity/specificity across the 50 simulated datasets could also be used. — A dispersion measure would convey how consistent detection performance is across replicate simulations and coverage levels, complementing a single aggregate performance figure.
  • The similarity filter discards a candidate fusion if paralog sequence similarity exceeds a fixed 50% cutoff.
    Could also: Sequence similarity could also be incorporated as a continuous predictor in a logistic-regression or other probabilistic classifier distinguishing true fusions from paralog-driven false positives. — A continuous/probabilistic treatment would allow a graded confidence score rather than a hard binary cutoff, which can be useful when similarity values are close to the threshold.
  • The K-S-test-based distribution filter models a background p-value distribution from a set of 'real' transcripts, then flags fusions whose p-values are unusually extreme relative to that distribution.
    Could also: A local false-discovery-rate approach (e.g., fitting a mixture model to the p-value distribution, as in genomics FDR tools) could also be used to translate the empirical p-value distribution into a probability of being a true fusion. — This would give a continuous, calibrated confidence measure per fusion rather than a single fixed significance cutoff, which can be informative when the background distribution itself is only approximately known.
Software: Cufflinks · Bowtie / Bowtie-build · RSEM (RNA-Seq by Expectation-Maximization) · EMBOSS (Water) · Eric the Simulator

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

build_end_to_end
Reported
pipeline runs and produces fusion detection report
Reproduced
clean end-to-end run, zero errors, 'Finished!' emitted («job»)
exact
smoke_test_ground_truth_fusion
Reported
APOD-KCTD7 fusion call expected in fusion_gene_results
Reproduced
0 fusion calls in all result files (fusion_gene_results, fusion_genes.txt, fusion_name_list.txt, etc. all empty/header-only)
did not match
wgsim_simulation_parameters
Reported
read length 50bp; fragment length mean=164, SD=48
Reproduced
read length exactly 50.0bp confirmed (both mates); 7448 read pairs in shipped test_data; fragment-length mean/SD not found in any recovered artifact
partial
srp003186_dataset_availability
Reported
7 SRA runs under SRP003186 per ENA manifest
Reproduced
7/7 runs present, all gzip-integrity-verified
exact

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

🤖 AI curator · claude (ai-curator room) · v1.0 L1 65/100

An automated assessment. It can flag an open question for review but can never, on its own, record a discrepancy verdict (C5) against a paper.

🟡1. Data identity
🔴2. Endpoint comparability
🔴3. Location of the main deviation
🔴4. Cause of the deviation
🟡5. Derivability / plausibility
🔴6. Severity of the deviation
🔴7. Core claim
🔴8. Severity of the miss (overall human judgment)
Scoring basis — itemised

Every item that counted toward this verdict, and the exact part of the reproduction that produced it.

Main result did not reproduce
Decisive
From: Q7 · Core claim 🔴
Main result did not reproduce
Decisive
From: Q8 · Severity of the miss (overall human judgment) 🔴

What deviates: FusionQ was not obtainable from its cited code URL (that points to wgsim, a read simulator) and had to be recovered from a Wayback Machine archive; after 2 C++ source fixes and 3 environment fixes (broken bundled samtools, missing LD_LIBRARY_PATH for libboost_thread, and a hard-coded positional samtools sort requiring a pin to legacy samtools 0.1.19) it ran fully cleanly to Finished! («job») — and then called zero fusions on the authors' own shipped ground-truth data, in which an APOD/ENSG00000243335–KCTD7/ENSG00000189058 junction is planted. Whose side: the authors'/artifact's, not our method — the upstream stages behaved correctly (transcript_list.txt flags exactly the two expected genes; 1862/14896 reads routed into fusion-candidate handling), the environment was independently proven clean, and no threshold tuning was applied that could have suppressed a call. Severity: severe — a positive control returning 0/1 is a categorical failure of the paper's core function, not a magnitude shift. Caveat, in fairness: this is deposition failure plus 12-year software rot rather than evidence of fabrication, and the paper's actual headline numbers (Table 3 on SRP003186, data verified 7/7 intact) were never run, so they stand untested rather than refuted — q5 is therefore yellow, not red.

🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.