henhaobio
Sign in

bioRxivBioinformaticsPreprint

Fusion detection and quantification by pseudoalignment

Páll MelstediD, Shannon L. HateleyiD, Isaac Joseph, Harold PimenteliD, Nicolas BrayiD, Lior S. PachteriD

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

Abstract

RNA sequencing in cancer cells is a powerful technique to detect chromosomal rearrangements, allowing for de novo discovery of actively expressed fusion genes. Here we focus on the problem of detecting gene fusions from raw sequencing data, assembling the reads to define fusion transcripts and their associated breakpoints, and quantifying their abundances. Building on the pseudoalignment idea that simplifies and accelerates transcript quantification, we introduce a novel approach to fusion detection based on inspecting paired reads that cannot be pseudoaligned due to conflicting matches. The method and software, called pizzly, filters false positives, assembles new transcripts from the fusion reads, and reports candidate fusions. With pizzly, fusion detection from raw RNA-Seq reads can be performed in a matter of minutes, making the program suitable for the analysis of large cancer gene expression databases and for clinical use. pizzly is available at https://github.com/pmelsted/pizzly

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (20 Jul 2017) · CC BY 4.0

Read pairs that fail kallisto pseudoalignment because of conflicting matches are filtered by pizzly to call, assemble and quantify gene fusions from raw RNA-seq in minutes rather than hours.

Fusion callers for RNA-seq typically map reads, regroup them by realignment and apply ad hoc heuristic filters; the paper describes them as slow, hard to install, uneven in sensitivity and specificity, and liable to miss breakpoints or skip quantification. Here a fusion mode of the kallisto pseudoaligner supplies the candidate reads, and pizzly applies annotation-aware filters, assembles fusion transcripts at breakpoint resolution and quantifies their abundance.

Key results

  1. Shownpizzly ran far faster than every comparison fusion caller. It finished the larger spike-in dataset in minutes, while the other tools needed hours.

    On the spike-in dataset SRR1659964 (93,867,189 100nt paired-end reads), pizzly took 8.5 minutes. The other tools took from 1 hour 18 minutes (STAR-Fusion) to over 26 hours (MapSplice, which ran on a 16-CPU cloud instance). Tools used four cores where possible. · Results; Real Data with Validation; Tables

  2. Suggestedpizzly had high sensitivity on every benchmark and high specificity on the simulated ones. The text notes specific weaknesses for PRADA, EricScript and JAFFA, and only fair sensitivity and specificity for STAR-Fusion.

    Four benchmarks (Tables 2-5): simulated 75nt and 100nt positives with 50 fusions, a 30-million-read simulated negative, and a spike-in with nine synthetic fusions. pizzly detected the most fusions in the 75nt control. PRADA failed on all 100nt sets, EricScript kept many false positives, JAFFA degraded on real data and STAR-Fusion was fair. Per-tool counts appear only in the tables. · Results; Tables; Positive Control; Negative Control

  3. SuggestedA kallisto-pizzly-kallisto pipeline re-quantifies reads against an index augmented with the assembled fusion transcripts. It gives fusion abundances in transcripts per million (TPM), and the spike-in total fell inside the expected range.

    On SRR1659964 the nine synthetic fusions summed to 31.6 TPM (Table 8). The expected total is 11-60 TPM, derived from the -6.17 log10 pmol spike-in into 1 ug RNA and assuming mRNA is 1-5% of total RNA. Only the sum is checked. The other callers report read counts only. · Discussion; Real Data with Validation; Tables

  4. SuggestedA second kallisto pass after fusion assembly resolved the two false positives pizzly reported in the 75nt positive control. Both came from genes that overlap in the genome.

    The false calls AC069368.3-KIF4A and BSG-AL021546.6 appeared alongside the true PLEKHO2-KIF4A and BSG-COX6A1. Initial kallisto estimates had two of the eight BSG transcripts and COX6A1 expressed, but not AL021546.6 (Table 6). After the sixteen candidate fusion sequences were indexed and the reads re-quantified, reads were assigned correctly at transcript level (Table 7). · Discussion; Tables

Methods, data and reproducibility

kallisto's new fusion mode saves read pairs whose k-mer equivalence classes are non-empty on each side of a junction but empty when combined. pizzly uses a GTF annotation to drop multi-locus and same-gene artifacts, removes similar-gene artifacts by approximate k-mer matching to the partner end's candidate transcripts, aligns reads across junctions and filters fusions by read support and exon-boundary distance.

  • Simulated positive control: 75nt paired-end reads from 50 fusion transcripts, 4,300 of them spanning junctions (made by the FusionMap developers) · n = 57,209 paired-end reads
  • Simulated positive control: 100nt paired-end reads from the same 50 fusion transcripts, generated with the RSEM read simulator · n = 200,000 paired-end reads
  • Simulated negative control: 100nt paired-end reads simulated with RSEM, using expression from H1 human embryonic stem cells as background · human (H1 embryonic stem cell background sample) · n = 30 million paired-end reads · SRR066679
  • Real mRNA-seq spike-in dataset (Tembe et al.): nine synthetic fusion constructs pooled equimolar into total RNA, Illumina TruSeq stranded library, HiSeq 2500 · n = 93,867,189 100nt paired-end reads; 9 fusion constructs · SRR1659964
Methods & tools
kallisto, pizzly, RSEM, MapSplice, PRADA, FusionCatcher, EricScript, JAFFA, STAR-Fusion, Illumina TruSeq stranded mRNA, Illumina HiSeq 2500
Assay conditions & compute
Tools ran on four cores where possible. MapSplice ran on an Amazon EC2 c4.4xl Ubuntu instance (16 CPUs, 30 GiB memory). pizzly took 8.5 minutes on the spike-in dataset, and the authors say the analysis can be done on a laptop.
Reproducibility gaps
The paper does not state software versions, the reference transcriptome or annotation release, or pizzly's parameter settings. No benchmark scripts are linked. The simulated datasets are not stated to be deposited; only the run accessions SRR066679 and SRR1659964 are given.
Limitations (4)
  • pizzly reported two false positives in the 75nt positive control (AC069368.3-KIF4A, BSG-AL021546.6), both from genes that overlap in the genome. Removing them needed an extra kallisto re-quantification step after fusion assembly. · stated by the authors
  • The authors present specificity as open to improvement, for example by iteratively refining filters using pizzly output and biologically relevant heuristics. · stated by the authors
  • Accuracy rests on four author-run datasets: two simulated positives sharing the same 50 fusions, one simulated negative and one spike-in with nine synthetic fusions. No tumor samples with orthogonal validation are tested, and high specificity is claimed only for the simulated data. · noted in this summary
  • Split-read breakpoints must lie within 10 bp of annotated exon boundaries on both sides. Fusions that break inside exons, introns or unannotated exons would likely be filtered out, and this trade-off is not evaluated. · noted in this summary

Read this if you call gene fusions from RNA-seq and want a pseudoalignment-based pipeline that runs in minutes and reports breakpoint sequences and fusion abundances, not only read counts.

Citations
45

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
THISTan '20Haas '17Mohebbi '21Nicorici '14Dehghannasi '19Davidson '15
74% semantic matchBioinformatics
Yuxiang Tan·bioRxiv·Mar 15, 2020· 1 citations