Gene Fusion Detection from RNA-Seq: Why Hundreds of Chimeric Transcripts Are Noise and a Handful Are the Diagnosis

Zetobit · Bioinformatics Insight Series

Gene Fusion Detection from RNA-Seq: Why Hundreds of Chimeric Transcripts Are Noise and a Handful Are the Diagnosis

A single tumor RNA-seq sample yields hundreds of chimeric transcripts. Almost all are biology or artifact, not cancer. The clinical task is not finding fusions — it is telling the one that matters apart from the crowd that doesn't.

By Kanna Nandakumar, PhD Zetobit LLC

An oncogenic gene fusion is one of the most actionable findings in cancer genomics. A BCR::ABL1 fusion defines chronic myeloid leukemia and is targeted by imatinib; fusions involving ALK, RET, ROS1, and the NTRK family each unlock a specific targeted therapy.¹ RNA-seq is uniquely suited to find them, because it reads the fused transcript directly regardless of where in the intron the underlying DNA break occurred, and it has meaningfully increased diagnostic yield over targeted assays.²

But RNA-seq detection has a signal-to-noise problem that sits at the center of every clinical fusion pipeline. Hundreds of chimeric transcripts can be detected per sample and predicted to reflect fusions.³ The overwhelming majority are not somatic cancer drivers. They are normal transcriptional byproducts, sequence-similarity artifacts, and alignment errors. The entire engineering problem is separating the one clinically decisive event from the hundreds of look-alikes — without discarding the real fusion in the process.

What the reads actually show

A fusion pipeline doesn't see "a fusion." It sees two kinds of read evidence at a candidate junction. Split reads (junction reads) cross the breakpoint directly: part of the read aligns to one gene, the rest to another, so the read itself will not map contiguously to the genome. Spanning fragments (encompassing pairs) bracket the junction: one mate lands in each gene, straddling a breakpoint that neither mate actually covers.⁴ Callers such as STAR-Fusion and Arriba build on a chimeric aligner, count these two evidence types, and score candidates by how many supporting fragments they carry.⁴

The trouble is that this evidence is generated by many processes other than a somatic rearrangement. A discordant read pair or a split read is a necessary signature of a fusion, but far from a sufficient one — the same pattern arises whenever two genes' sequences end up adjacent in a transcript, or merely appear to because their sequences resemble each other. That is why raw caller output is a candidate list, not a result.

Split and spanning reads look the same regardless of what produced them. The clinical challenge is that read-through transcription, normally-occurring chimeras, and homology artifacts all generate the identical signature as a true somatic fusion — so evidence counts alone cannot distinguish driver from background.

The false positives are structured, not random

The look-alikes fall into recognizable classes, which is what makes them filterable — and dangerous when filtered carelessly.

Read-through transcription. Two genes sitting adjacent on the same strand can be transcribed into a single RNA with no underlying genomic rearrangement at all. Because these mimic a small deletion, tools flag fusions whose partners are within intron-scale distances — Arriba, for instance, marks deletion-like events under roughly 400 kb as probable read-throughs.⁵ They are common, reproducible, and biologically real — just not cancer.

Normally-occurring chimeras. Some chimeric transcripts appear in healthy tissue. The clearest cautionary case is PAX3::FOXO1 — a bona fide driver of alveolar rhabdomyosarcoma when it arises from a genomic translocation, yet the same chimeric transcript is transiently expressed in normal muscle without any structural variant behind it.³ The transcript alone cannot tell you which situation you're in.

Sequence homology. Paralogous genes and repetitive sequence produce reads that mismap between similar loci and masquerade as chimeric. This is the RNA-level echo of the multi-mapping problem this series has covered on the DNA side, and fusion callers specifically exclude candidate pairs with high sequence similarity or promiscuous, recurrent partners as likely artifacts.⁴

Filtering cuts both ways. The stringent filters that suppress false positives also discard real driver fusions — a documented failure mode, and worse in degraded FFPE material, where more fragmented RNA yields more "controversial" reads that get filtered away.⁶ A pipeline tuned only for specificity can silently drop the very fusion the test exists to find. The tuning is a clinical decision, not a default.

Why callers disagree with each other

Because each tool draws the sensitivity/specificity line differently — different aligners, annotations, evidence thresholds, and exclusion lists — the same sample run through different callers yields different fusion lists, with high false-positive rates and poor overlap between tools a repeatedly documented finding.⁷ This is why serious clinical fusion pipelines don't rely on a single caller. The established practice is to run several and combine them: a pediatric-cancer pipeline overlapping seven callers and requiring at least three to agree recovered all expected fusions in a reference control while controlling false positives, and cross-referenced a curated knowledge base of known events.⁸

That ensemble design carries an implication worth stating plainly: a fusion call's credibility depends on how many independent methods saw it, whether it is a known recurrent driver, and whether orthogonal evidence supports it — not on its mere presence in one tool's output. A single-caller hit on a novel gene pair and a multi-caller hit on a known driver are not the same kind of result, even if both show comparable read counts.

What a real fusion pipeline has to do

Turning chimeric reads into a reportable fusion is a curation problem layered on a detection problem:

• Detect with a chimeric-aware aligner and multiple callers. Use split-read and spanning-fragment evidence, and take the union or a voting consensus across tools rather than trusting one.⁷
• Classify the background explicitly. Flag read-throughs by partner distance and orientation, and screen against panels of normal chimeras seen in healthy tissue, so recurrent non-somatic events are labeled rather than reported.⁵
• Suppress homology artifacts by excluding sequence-similar and promiscuous partner pairs — while keeping a curated rescue list so known low-support drivers aren't lost to the same filter.⁴
• Confirm orthogonally. Corroborate clinically reportable fusions against DNA structural-variant evidence (WGS/WGS-informed) or an orthogonal assay such as FISH, RT-PCR, or targeted NGS before acting on them.⁹
• Weight by prior knowledge. Prioritize known recurrent driver fusions and reading-frame-preserving events, and treat novel pairs as hypotheses requiring more evidence, not conclusions.⁸

As with the other blind spots in this series, the honest output declares its own limits: which fusions were interrogated, how many callers supported each, and whether orthogonal confirmation was obtained. "Candidate fusion, single caller, unconfirmed — reflex to orthogonal assay" is a truthful result. A clean-looking fusion call with no provenance behind it is not.

The takeaway

Gene fusion detection inverts the usual scarcity problem: the pipeline finds too much, not too little. Hundreds of chimeric transcripts per sample are read-throughs, normal chimeras, and alignment artifacts; a small number are somatic drivers that change how a patient is treated. Getting it right is less about the aligner and more about the filtering, the ensemble, the knowledge base, and the orthogonal confirmation that together separate a decisive finding from expensive noise — and about reporting the difference, rather than presenting a candidate as a conclusion.

References

1. Experimentally deduced criteria for detection of clinically relevant fusion 3′ oncogenes from FFPE bulk RNA-seq (RTK fusions: ALK, NTRK1-3, RET, ROS1, FGFR1-4, ERBB2, ABL1; BCR-ABL1/imatinib). Biomedicines. 2022. 10.3390/biomedicines10081866
2. Systematic discovery of gene fusions in pediatric cancer by integrating RNA-seq and WGS (RNA-seq diagnostic-yield gain; Fusion-sq). BMC Cancer / bioRxiv. 2023. PMC10318758
3. Hundreds of chimeric transcripts per sample; read-through and trans-splicing; PAX3::FOXO1 in normal muscle. bioRxiv. 2021. biorxiv.org
4. Haas BJ, et al. STAR-Fusion: fast and accurate fusion transcript detection from RNA-Seq (split/junction and spanning fragments; similarity and promiscuity filters). bioRxiv. 2017. biorxiv.org
5. Uhrig S, et al. Arriba — accurate and efficient detection of gene fusions from RNA-seq (read-through flagging <~400 kb; recovers events other tools miss). Genome Res. 2021. genome.cshlp.org
6. Stringent filters discard driver fusions, worse in degraded FFPE. Biomedicines. 2022. 10.3390/biomedicines10081866
7. INTEGRATE and cross-tool comparisons — high false-positive rates and poor overlap between fusion callers; two-tool union/intersection practice. Genome Res / PMC. PMC4691743
8. Discovery of clinically relevant fusions in pediatric cancer (ensemble of 7 callers; ≥3-caller overlap + knowledge base recovered all reference fusions). bioRxiv. 2021. biorxiv.org
9. SeekFusion — a clinically validated fusion transcript detection pipeline, orthogonally benchmarked, NYS-approved LDT. Front Genet. 2021. PMC8569241

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com