🧬 biology

RNA folding below the nanopore impairs direct RNA sequencing fidelity and modification calling

This study identifies that RNA secondary structures folding below the nanopore during direct RNA sequencing cause systematic ionic current distortions and false-positive modification calls, thereby establishing a critical need for structure-aware models to improve sequencing fidelity.

Original authors: Jana Wiedermannová, Marek Schwarz, Josef Pánek, Saran Natarajan, Veronika Kočárková

Published 2026-07-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jana Wiedermannová, Marek Schwarz, Josef Pánek, Saran Natarajan, Veronika Kočárková

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to read a secret message written on a long, slippery ribbon of RNA as it zips through a tiny, high-tech tunnel called a nanopore. This is how Oxford Nanopore Direct RNA Sequencing (DRS) works: it listens to the electrical "hum" of the ribbon passing through to figure out the letters (A, U, C, G) and even spot tiny chemical stickers (modifications) attached to them.

But here's the twist: the paper reveals that the ribbon isn't just a straight line. Once the ribbon pops out the other side of the tunnel (the "trans" compartment), it can curl up and tie itself into a knot, like a hairpin. The authors found that these knots are causing the machine to get confused, leading to a lot of false alarms.

The "Hairpin" Problem
Think of the nanopore like a turnstile at a subway station. The RNA ribbon is the passenger. Usually, the passenger walks through smoothly. But if, just as they exit the turnstile, they suddenly stop to tie their shoelaces into a tight knot (a hairpin structure), it pulls on the ribbon. This tug changes how the passenger moves and how the turnstile's sensors read them.

The authors discovered that when these RNA hairpins form right below the pore, they create a "zone of confusion" about 15 letters before the knot. In this zone, the machine starts making mistakes:

It thinks there are chemical stickers (modifications) where there are none.
It gets the letters wrong (sequencing errors).
It lowers its confidence score (Quality Score) for those letters.

The "Ghost" Modifications
The paper shows that this isn't just a fluke. The researchers tested this with "in vitro" (lab-made) RNA that was guaranteed to have zero real chemical stickers. Yet, the machine still reported finding them right before the hairpins. It's like a metal detector beeping loudly near a metal pole, even though there's no treasure buried there. The "treasure" (the modification) was a ghost created by the knot pulling on the ribbon.

What They Ruled Out
The authors were very careful to prove that the problem wasn't the RNA being near the end of the strand or just being a "messy" sequence.

They cut the RNA short or made it longer, and the problem stayed exactly where the knot was, not where the end of the ribbon was.
They broke the knots by changing a few letters so the RNA couldn't tie itself up. When the knots disappeared, the ghost modifications and errors vanished too.
They even tried different software models (Dorado versions 5.0.0, 5.2.0, 5.3.0, and 6.0.0). While newer models helped a little, they didn't fix the problem completely. The ghost signals are still there, suggesting the issue is physical (the knot pulling the ribbon) rather than just a software bug.

The Evidence
The team didn't just guess; they measured it.

They looked at the electrical current and found it wobbled strangely right before the hairpins.
They measured the "dwell time" (how long the RNA stayed in the pore) and found it changed, too.
They tested this in bacteria (Bacillus subtilis), human cells, and E. coli, and the same "knot-caused confusion" happened everywhere.
They found that in the lab-made RNA, about 7.5 times more fake modifications appeared near low-quality spots than you'd expect by chance.

Why It Matters
The paper suggests that if scientists are looking for real chemical stickers on RNA to understand diseases, they need to be careful. If they see a modification right before a strong RNA knot, it might just be a "ghost" caused by the knot pulling on the ribbon.

The authors propose three ways to fix this:

Flag the knots: Tell the computer to be extra skeptical of any "modification" found within 15 letters before a predicted knot.
Train the AI better: Teach the software models to recognize these knots so they don't get fooled.
Stop the knots: Add chemicals to the solution to keep the RNA straight so it can't tie itself into hairpins below the pore.

In short, the paper shows that RNA folding below the nanopore is a hidden culprit that makes the machine see things that aren't there. It's a physical tug-of-war that tricks the sensors, and until we account for these knots, our map of RNA modifications might have some fake landmarks.

Technical Summary: RNA Folding Below the Nanopore Impairs Direct RNA Sequencing Fidelity and Modification Calling

Problem Statement
Oxford Nanopore Technologies (ONT) Direct RNA Sequencing (DRS) offers the unique capability to sequence full-length native RNA molecules and infer chemical modifications directly from ionic current signatures. However, DRS currently suffers from lower per-read accuracy compared to short-read cDNA sequencing, with a high prevalence of indels and false-positive modification calls. While previous studies have attributed these errors to low signal-to-noise ratios, chemical similarities between nucleotides, and translocation variability in homopolymeric or specific sequence contexts, a significant source of systematic distortion remained unidentified. Specifically, the phenomenon of RNA secondary structure refolding in the trans compartment (below the nanopore) during sequencing had not been systematically characterized as a driver of basecalling errors and spurious modification signals.

Methodology
The authors employed a combination of in vivo transcriptomics and controlled in vitro experiments to isolate the effects of RNA folding on DRS signals.

Organisms and Samples: The study utilized Bacillus subtilis (both wild-type total RNA and modification-free in vitro transcribed (IVT) controls), E. coli, and human cell line (HEK293T) datasets.
Experimental Design: To establish causality, the researchers engineered specific IVT transcripts based on B. subtilis genes (cspD, yĬH, iolW, yxkC, kdgT). They generated variants where intrinsic transcription terminator hairpins or internal secondary structures were:
- Disrupted: By truncating the distal (3') arm of the hairpin ("half") or introducing point mutations to prevent base pairing ("mut").
- Relocated: By extending the 3' end ("long") to shift the hairpin position relative to the read end.
- Repositioned: By fusing hairpin sequences to high-quality regions ("del" and "del mut").
Sequencing and Analysis: Samples were sequenced on PromethION devices using R10.4.1 pores. Data were basecalled using Dorado (versions 5.0.0 through 6.0.0) with modified-base inference enabled. The study utilized tools such as modkit, m6Anet, SWARM, and Uncalled4 for signal alignment and modification calling.
Signal Analysis: The authors analyzed raw ionic current levels, dwell times (event lengths), and base quality scores (QS) relative to predicted secondary structures (minimum free energy, $\Delta G$ ).

Key Results

Hairpin-Associated Artifacts: In both in vivo and modification-free IVT samples, high-confidence modification calls (HCMs) and basecalling errors (mismatches, indels) were consistently enriched in a reproducible window immediately upstream (~12 nt) of stable RNA hairpins, including intrinsic transcription terminators and strong internal structures.
Causal Link via Disruption: Disrupting the formation of these hairpins through truncation or point mutations abolished the accumulation of false-positive modification calls and restored base quality scores to background levels. This confirmed that the downstream folding, not the sequence context alone, caused the artifacts.
Signal Perturbation: Signal-level analysis revealed that the presence of a downstream hairpin caused localized deviations in ionic current and increased variance in dwell time upstream of the structure. These perturbations are consistent with altered translocation dynamics or steric interference at the nanopore exit.
Generality: The phenomenon was observed across different organisms (B. subtilis, E. coli, human) and different modification callers (Dorado, SWARM), indicating a general physical mechanism rather than a software-specific bug.
UC-Rich Contexts: While UC-rich sequences are known to reduce basecalling accuracy, the study found that downstream hairpins act as an independent driver of artifacts that is amplified in UC-rich contexts. However, strong hairpins cause signal disruption even in the absence of UC-rich motifs.
Model Limitations: Analysis of multiple Dorado model versions (v5.0.0 to v6.0.0) showed that while some updates reduced specific artifacts, the underlying structure-linked distortion persists, with newer models still generating substantial false-positive signals in modification-free RNA.

Significance and Claims
The paper identifies RNA refolding in the trans compartment as a previously unrecognized, general source of signal distortion in DRS. The authors claim that this mechanism systematically compromises sequencing fidelity and generates false-positive modification calls, particularly upstream of stable secondary structures.

The study emphasizes that candidate modification sites located upstream of strong hairpins must be interpreted with caution unless supported by orthogonal validation (e.g., mass spectrometry) or rigorous negative controls. The authors propose three mitigation strategies:

Implementing structure-aware quality control in modification-calling algorithms to flag or downweight bases within ~15 nt upstream of predicted strong hairpins.
Training future basecalling models on synthetic RNA constructs containing defined hairpins or structure-aware data from non-modified IVT transcriptomes.
Physically suppressing hairpin formation in the flow cell trans compartment (e.g., via chemical additives or denaturing agents), though this requires careful optimization.

Ultimately, the findings provide a mechanistic explanation for context-dependent errors in DRS and establish a necessary foundation for structure-aware data interpretation to improve the reliability of epitranscriptomic research and clinical applications.

More like this