Best practices for genome-wide RNA structure analysis: combination of mutational profiles and drop-off information

Eva Maria Novoa, Jean-Denis Beaudoin, Antonio J. Giráldez, John S. Mattick, Manolis KellisbioRxiv 2017 · DOI

TL;DR

This paper reveals that two common methods for mapping RNA structure – reverse transcriptase (RT) drop-off and mutational profiling (MaP) – capture complementary information and are not interchangeable. By combining these methods, researchers can achieve more accurate RNA structure predictions, especially when considering the influence of different reverse transcriptase enzymes.

What Problem Does This Paper Address?

Understanding RNA structure is crucial for deciphering its function, but accurately mapping it across an entire genome (genome-wide) is challenging. Historically, two main approaches have been used:

  1. RT Drop-off: This method relies on identifying where the reverse transcriptase enzyme stops during RNA sequencing, indicating a chemical modification on the RNA that impedes its progress.
  2. Mutational Profiling (MaP): This method looks for specific "mismatch" signatures (errors) introduced by the reverse transcriptase when it encounters a modified nucleotide.

While these methods have been used, it was often assumed they produced similar results. This paper addresses the problem that using only one method might miss significant structural information because they capture different aspects of RNA modification. The authors aim to clarify the relationship between these methods, understand what influences their outcomes, and propose a best practice for more comprehensive RNA structure analysis.

Methods

The study employs and compares two primary methodologies for RNA structure analysis:

  • Reverse Transcription (RT) Drop-off (or Accessibility): This standard method involves probing RNA with chemical agents (like DMS) that modify accessible nucleotides. The extent to which reverse transcriptase (RT) stops at these modified sites is measured, providing an "accessibility" score.
  • Mutational Profiling (MaP): This method focuses on identifying specific "mismatch" signatures (errors) introduced by the RT enzyme when it encounters a modified nucleotide. These signatures are sequence-dependent and can reveal information about the modification.

The authors also investigate the impact of different reverse transcriptase enzymes (SuperScript-III and TGIRT) on the MaP signals, as well as the role of sequence context surrounding the modified nucleotides. They develop filtering strategies by combining DMS-Seq with RNA-Seq data to remove noise and improve signal-to-noise ratio. Finally, they use Support Vector Machines (SVMs) to integrate signals from different methods and predict RNA pairing status and secondary structure.

Datasets

The study utilizes several datasets for analysis and evaluation:

  • In vitro probed RNA spike-ins: Tetrahymena ribozyme and a DsRed mRNA containing a spinach tRNA cassette. These are well-characterized RNA structures used to validate the methods.
  • In vivo DMS-Seq datasets: From 64-cell stage zebrafish embryos.
  • RNA-Seq datasets: From 64-cell stage zebrafish embryos, used for filtering out non-DMS-induced mismatches.
  • Structure-Seq datasets: From 64-cell stage zebrafish embryos, generated using different reverse transcriptase enzymes (SS3 and TGIRT).
  • DMS-probed yeast rRNA datasets: Previously published datasets (GSE45803 and GSE84537) with experimentally determined structures, used for training and evaluating prediction models.

The paper does not introduce new benchmark datasets but rather leverages existing ones to compare and combine methodologies.

Key Insights and Ideas

  • Complementarity is Key: RT drop-off and MaP are not interchangeable; they capture distinct, complementary information about RNA modifications and structure.
  • RT Enzyme Matters: The choice of reverse transcriptase enzyme significantly impacts both the frequency of mismatches and the specific "mutational signatures" observed, meaning signatures are not solely an intrinsic property of the RNA modification.
  • Nucleotide Identity Influences Outcome: The specific modified nucleotide (e.g., m¹A vs. m³C) influences whether RT is more likely to drop off or introduce a mismatch.
  • Sequence Context Plays a Role: While not definitively disentangled from library preparation biases, sequence context around the modified nucleotide can influence RT behavior.
  • Filtering is Crucial: Combining DMS-Seq with RNA-Seq is effective for filtering out false positives and increasing the signal-to-noise ratio in MaP analysis.
  • Combined Methods Improve Predictions: Integrating signals from both RT drop-off and MaP, especially when considering RT enzyme choice, leads to more accurate predictions of RNA pairing status and secondary structure.

Key Results

  • Low Correlation: The correlation between positions identified by RT drop-off and MaP is poor, indicating they capture different sets of modified nucleotides.
  • MaP Captures 5'/3' Ends: MaP is better at identifying modifications at the 5' and 3' ends of RNA molecules, which are often missed by RT drop-off due to library preparation steps.
  • Nucleotide Bias: RT drop-off preferentially identifies adenosine (A) modifications, while MaP preferentially identifies cytosine (C) modifications when using the SS3 enzyme.
  • RT Enzyme Impact: TGIRT enzymes, being more processive, generally increase mismatch frequency over drop-off compared to SS3. This effect is not uniform across all RNA modifications.
  • Improved Prediction Accuracy: Combining RT-stop and MaP signals improved RNA secondary structure prediction accuracy by 8% compared to RT-stop alone and 11% compared to MaP alone.
  • Enhanced Pairing Status Prediction: Combining mismatch frequencies and mutational signatures significantly improved the accuracy of predicting C mismatches.

Limitations and Future Work

  • Sequence Context vs. Library Bias: The study acknowledges that observed differences in sequence context might be partially due to library preparation biases, making it difficult to fully isolate the effect of sequence context on RT outcomes.
  • RT Enzyme Specificity of Signatures: The authors propose that mismatch signatures are specific to the RNA modification-reverse transcriptase combination, not an intrinsic property of the modification itself. This implies that signatures derived from one RT enzyme cannot be directly applied to datasets generated with a different enzyme.
  • Future Work: The paper suggests that further investigation into how additional variables like RT temperature or salt concentration might affect mismatch signatures is warranted. The authors also propose that combining RT truncation and mutational profiling methodologies can facilitate more complete identification of genome-wide RNA modifications.

Summary updated

Abstract

ABSTRACT Genome-wide RNA structure maps have recently become available through the coupling of in vivo chemical probing reagents with next-generation sequencing. Initial analyses relied on the identification of truncated reverse transcription reads to identify the chemically modified nucleotides, but recent studies have shown that mutational signatures can also be used. While these two methods have been employed interchangeably, here we show that they actually provide complementary information. Consequently, analyses using exclusively one of the two methodologies may disregard a significant portion of the structural information. We also show that the identity and sequence environment of the modified nucleotide greatly affect the odds of introducing a mismatch or causing reverse transcriptase drop-off. Finally, we identify specific mismatch signatures generated by dimethyl sulfate probing that can be exploited to remove false positives typically produced in RNA structurome analyses, and how these signatures vary depending on the reverse transcription enzyme used.
BibTeX
@misc{novoa2017best,
  author = {Eva Maria Novoa and Jean-Denis Beaudoin and Antonio J. Giráldez and John S. Mattick and Manolis Kellis},
  title = {{Best practices for genome-wide RNA structure analysis: combination of mutational profiles and drop-off information}},
  year = {2017},
  doi = {10.1101/176883},
  howpublished = {bioRxiv},
}

Loading related content...

Figure