Accurate detection of m6A RNA modifications in native RNA sequences

Huanle Liu, Oguzhan Begik, Morghan C Lucas, Christopher E Mason, Schraga Schwartz, John S Mattick, Martin A Smith, Eva Maria Novoa

Preprint posted on January 21, 2019

Hidden in plain sight: A machine learning approach uses sequencing errors to identify native RNA modifications in nanopore sequencing.

Selected by Christian Bates


The RNA content of the cell encodes a vast amount of information. Aside from simply encoding the amino-acid sequence required to build a protein, portions of RNA are capable of regulating cellular processes such as splicing, mRNA translation and chromosome inactivation. In recent years, it has emerged that RNA can be chemically modified in numerous different ways. All of these RNA modifications in the cell have been termed the epitranscriptome (‘above transcriptome’) as they do not alter the nucleotide sequence of the biomolecule, but they can impact its ability to function in many of the processes outlined above, providing an additional layer of regulatory information [1].

Whilst modifications to species of RNA such as rRNA and tRNA have been long known – as early as the 1960s [2,3] – their impact on mRNA has recently been appreciated. One such modification is the methylation of a nitrogen atom in adenine, generating m6A. This modification is widespread in higher eukaryotes and can have a profound impact on mRNA stability and translation. However, despite the ubiquity of this modification, studying m6A modifications at the transcriptome-wide level has been particularly challenging because m6A modifications do not impact Watson-Crick base pairing. As such, m6A modifications cannot be identified via reverse-transcription methods, which are routinely used to assess other RNA modifications. Therefore, current epitranscriptomic studies interrogating m6A infer its presence through the use of antibodies. These effectively enrich RNA with a specific modification, which are then identified via high-throughput sequencing platforms.

This preprint from Liu et al aims to identify m6A modifications directly using an emerging sequencing platform provided by Oxford Nanopore Technologies (ONT). This platform directly sequences nucleic acids, unlike existing next-gen technologies, such as those provided by Illumina, which sequence DNA and RNA by synthesis. Specifically, ONT consists of thousands of individual polymer membranes, each with a single nanopore embedded within them. Nucleic acid is captured by these nanopores and ratcheted through the membrane. Due to the electrochemical potential of each base, this ratcheting perturbs the current between the two sides of the membrane. As each base is chemically different, they generate an idiosyncratic perturbation in the current, which can be deconvoluted by a recurrent neural network, converting this signal into a sequence of bases (Figure 1).

Figure 1: Modified version of Figure 1A from Liu et al (2019) pre-print


Key Findings

RNA modifications exert their effects within the cell through their ability to provide unique physico-chemical properties to the modified nucleotide. This unique ‘signature’ can then be read by specific enzymes, subsequently dictating the fate of the modified RNA [4]. With this in mind, Liu, et al. begin with the hypothesis that, due to their unique chemical signature, modified nucleotides will also generate unique current intensity change as the RNA is ratcheted through the pore. As a consequence, the authors suggest that if a base is modified, it may be less likely to be assigned correctly, particularly if the program used to decipher the nucleotide sequence has not been trained to look for modifications.

To test this hypothesis, Liu, et al. sequenced two versions of the same oligo: one which contained unmodified adenine, and another in which all instances of adenine were substituted for m6A. This showed that, as they had predicted, m6A-modified reads contained more errors, and that these errors were primarily found at adenine nucleotides. Importantly, these errors were reproduced across several repeats, suggesting that they are not random errors, but rather that they are the result of some underlying feature of the RNA at that position.

Next, they went on to prove that these errors are sufficient to determine whether a nucleotide was modified or not, by testing whether modified RRACH (R = G/A; H = A/C/U) oligos are sufficiently different to non-modified RRACH oligos. This is important as RRACH is the most common m6A motif; in some cell types, more than 85% of m6A sites were found to occur at this motif [5]. Not only were modified RRACH motifs different from unmodified motifs, but by using a combination of error features of the modified adenine, such as the confidence of the base call and the frequency of errors at that base, a machine learning approach was capable of predicting whether a motif was modified with 91% accuracy.

Together, these experiments demonstrate a proof-of-principle that ONT sequencing platforms could be used to identify nucleotide modifications on RNA directly. The capacity to do so would greatly enhance the accuracy and resolution of existing technologies.


Why I chose this pre-print

The epitranscriptome represents an important facet of the regulation of gene expression. Yet, whilst it has been possible in the past to study m6A modifications at the genome-wide level, previous approaches rely upon the use of antibodies, making them expensive and limited in resolution. This pre-print demonstrates the capability of calling m6A modifications directly, with single nucleotide resolution using an emerging sequencing platform, provided by ONT.

I particularly like the fact that the authors have made their data and code available on public repositories. This allows quick dissemination of their work and also enables other groups to test whether alternative machine learning approaches may call m6A sites with higher accuracy.


Future Directions and Questions

This work highlights the ability to predict modified nucleotides on a synthetic RNA sequence, with either all or no adenines possessing modifications. As such, it would be interesting to see whether this approach could call differences on non-uniformly modified RNA extracted directly from cells. I also think it would be interesting to test whether the characteristic differences in base quality and error frequency could be used to specifically differentiate between m6A modifications, and other adenine modifications such as m1A, or whether these errors can only be used to infer that some modification has occurred at a specific base.



1         Roundtree IA, Evans ME, Pan T, He C. (2017) Dynamic RNA Modifications in Gene Expression Regulation. Cell; 169: 1187–1200. doi:10.1016/j.cell.2017.05.045.

2         Cohn WE. (1960) Pseudouridine, a carbon-carbon linked ribonucleoside in ribonucleic acids: isolation, structure, and chemical characteristics. J Biol Chem; 235: 1488–1498.

3         Holley RW, Everett GA, Madison JT, Zamir A. (1965) Nucleotide Sequences in the Yeast Alanine Transfer Ribonucleic Acid. J Biol Chem ; 240: 2122–2128.

4         Sloan KE, Warda AS, Sharma S, Entian KD, Lafontaine DLJ, Bohnsack MT. (2017) Tuning the ribosome: The influence of rRNA modification on eukaryotic ribosome biogenesis and function. RNA Biol; 14: 1138–1152. doi:10.1080/15476286.2016.1259781.

5         Chen T, Hao Y-J, Zhang Y, Li M-M, Wang M, Han W et al. (2015) m6A RNA Methylation Is Regulated by MicroRNAs and Promotes Reprogramming to Pluripotency. Cell Stem Cell; 16: 289–301. doi:10.1016/j.stem.2015.01.016.

Tags: gene expression, machine learning, nanopore, rna, rna modifications, sequencing

Posted on: 2nd April 2019

Read preprint (No Ratings Yet)

  • Author's response

    Eva Maria Novoa shared

    Could the platform be used to differentiate between m6A and m1A modifications?

    We are indeed testing this with other modifications, to see how well we can predict each of them using this strategy.

    How well do you think this platform will work for non-uniformly modified RNAs?

    We are currently testing this this and it seems to also work, but some adjustments to the code must be made to account for/predict stoichiometry. We expect to release these results and corresponding code soon in a newer version of EpiNano.  

    Have your say

    Your email address will not be published. Required fields are marked *

    This site uses Akismet to reduce spam. Learn how your comment data is processed.

    Sign up to customise the site to your preferences and to receive alerts

    Register here