Detection of m6A from direct RNA sequencing using a multiple instance learning framework.

Hendra C, Pratanwanich PN, Wan YK, Goh WSS, Thiery A, Göke J. · 2022

Read Detection of m6A from direct RNA sequencing using a multiple instance learning framework. by Hendra C, Pratanwanich PN, Wan YK, Goh WSS, Thiery A, Göke J. (2022) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc9718678.txt.

Introduction

Hendra and co-authors present m6Anet, a method that finds m6A marks from direct RNA sequencing reads. The m6A mark adds a layer of control to RNA messages. Nanopore reads carry the mark in their raw current signal. Training data, though, labels whole sites and not single molecules.

That gap is where most learning tools stumble at the start. m6Anet uses a multiple instance learning frame to bridge it. The frame learns from site labels while scoring single reads. Tests showed it beat rival computing methods on the task. Its calls matched wet lab method accuracy in the checks. It also worked on new cell lines and species with no retraining. Read level scores gave a sense of mark rates per site. One run can thus both find and size marks genome wide.

Catalog record

CREATORHendra C, Pratanwanich PN, Wan YK, Goh WSS, Thiery A, Göke J.
YEAR2022
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC9718678

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Detection of m6A from direct RNA sequencing using a multiple instance learning framework By Christopher Hendra, Ploy N Pratanwanich, Yuk Kei Wan, W S Sho Goh, Alexandre Thiery, Jonathan Göke Published in: Nature Methods DOI: 10.1038/s41592-022-01666-1 ABSTRACT Abstract RNA modifications such as m6A methylation form an additional layer of complexity in the transcriptome. Nanopore direct RNA sequencing can capture this information in the raw current signal for each RNA molecule, enabling the detection of RNA modifications using supervised machine learning. However, experimental approaches provide only site-level training data, whereas the modification status for each single RNA molecule is missing. Here we present m6Anet, a neural-network-based method that leverages the multiple instance learning framework to specifically handle missing read-level modification labels in site-level training data. m6Anet outperforms existing computational methods, shows similar accuracy as experimental approaches, and generalizes with high accuracy to different cell lines and species without retraining model parameters. In addition, we demonstrate that m6Anet captures the underlying read-level stoichiometry, which can be used to approximate differences in modification rates. Overall, m6Anet offers a tool to capture the transcriptome-wide identification and quantification of m6A from a single run of direct RNA sequencing. Subject terms: RNA, Software, Transcriptomics, Genome informatics MAIN Modifications in RNA nucleotides were first discovered in the 1950s1,2, and today, more than 150 different modifications have been described3,4. One of the most common RNA modifications is m6A, the main internal methylation on mammalian mRNA5,6. This modification presents mostly at the consensus motif DRACH (D–A, G, or U, R–A or G while H is A, C or U) and has been shown to impact RNA structure7, stability8,9, splicing10, and translation11. Disruption of m6A homeostasis in animal models affects regulation of stem cells12,13, fertility and the developmental process14, while in humans, this modification plays an important role in cancer15,16, cell-fate transition and determination17,18 and transition, development19, and diseases20,21. Experimental identification of trasncriptome-wide RNA modifications can be achieved with three main approaches. Immunoprecipitation methods such as MeRIP-Seq22, m6A-Seq23, PA-m6A-Seq24, m6A-CLIP/IP25, miCLIP26, m6A-LAIC-Seq27, m6ACE-Seq28, and m6A-Seq229 use antibodies that specifically bind to the modified ribonucleotide. Chemical-based detection methods such as Pseudo-Seq30, AlkAniline-Seq31, utilize chemical compounds that selectively react with the modified ribonucleotide. Approaches such as Mazter-Seq32, m6A-REF-Seq33, or DART-Seq34 use specific enzymes to selectively distinguish modified and unmodified bases. These approaches are similar in that they isolate the RNA after inducing changes to the surrounding nucleotides, followed by reverse transcription and short-read cDNA sequencing to detect these changes. While these approaches provide a transcriptome-wide map of RNA-modification sites, they are limited by the availability of antibodies or compounds for specific modifications35. Also they lack single-nucleotide resolution22,23 and are incapable of identifying modifications for single RNA molecules. The ability to sequence native RNA using Nanopore direct RNA sequencing (RNA-Seq) can potentially overcome these limitations36. Direct RNA-Seq infers the RNA sequence using the current intensity when RNA molecules pass through the pores. Modified nucleotides will emit a different signal intensity compared to unmodified nucleotides, allowing the computational identification of modified sites for each individual RNA molecule using either supervised or comparative approaches37. Comparative approaches do not require training data for known RNA modifications but instead use control or reference samples to detect meaningful shifts in signal-based features that correlate to the presence of modifications. Comparative methods such as Tombo38, DRUMMER39, nanoDOC40, Nanocompore41, ELIGOS42, xPore43, and Yanocomp44 detect m6A sites by comparing against samples without m6A modifications. While these methods are accurate, their success relies on the availability of m6A-free control samples. This involves silencing of specific writer genes, which can be a limiting factor. Supervised detection of m6A modifications involves training a classifier using labels that can be obtained from synthetically modified RNA samples or experimental protocols such as miCLIP26, MeRIP-Seq22 or m6ACE-Seq28. Methods such as EpiNano45,46, MINES47, and nanom6A48, use training data to identify m6A using the sequencing error profile or shifts in the current signal intensity. Supervised methods can be applied on a single sample, overcoming the main limitation of comparative methods for detection of specific RNA modifications. Furthermore, given the availability of training data, one can retrain supervised classifiers to detect other modifications such as pseudouridine, which is detected by NanoRMS49. However, existing approaches are limited to a specific nucleotide content45–48, and they are currently less accurate than comparative approaches using an m6A-free control41,43,44. One of the main challenges for supervised approaches is that training labels are provided for a set of reads at the site level, but not for each individual read, which is known as a multiple instance learning (MIL) problem50,51. Existing methods address this problem by averaging read-based features45–47. However, at any given site, we are likely to have a mixture of modified and unmodified reads and as such, not all reads provide useful features to detect m6A sites. Therefore, current approaches, which do not consider the MIL structure in the training data might fail to detect m6A modifications from sites with low stoichiometry as it tends to obscure signals from the lowly expressed modified RNAs, and it limits the ability to integrate variation in read-level features into a predictive model. To address these limitations, we developed m6Anet, a MIL-based neural network model that takes in signal intensity and sequence features to identify potential m6A sites from direct RNA-Seq data. Our model takes into account the mixture of modified and unmodified RNAs and outputs the m6A-modification probability at any given site for all DRACH fivemers represented in the training data. Unlike existing approaches, m6Anet learns a high-dimensional representation of individual reads from each candidate site before aggregating them together to produce a more accurate prediction of m6A sites. By applying m6Anet to human, Arabidopsis, and synthetic direct RNA-Seq data we demonstrate that it detects previously unlabeled m6A sites and generalizes across different cell lines and species without a reduction in performance. The approach is general enough that the network can be retrained to classify any natural or artificial RNA modifications given a set of labels. RESULTS m6ANet identifies methylated positions with a MIL approach Here we present m6Anet, a neural-network-based MIL model that combines learning the representation of each individual read with classifying m6A-modified sites. m6Anet comprises two separate modules that are optimized jointly—a read-level encoder and a pooling layer. The read-level encoder uses signal and sequence features from each read, and transforms them into a high-dimensional representation before predicting the probability of each read being modified (Fig. 1a). The read-level probability is then pooled to give a probability estimate that a site is modified (Fig. 1a). By combining features that represent signal and sequence properties, m6Anet can learn a model that can be applied for all fivemers that are represented in the training data. Furthermore, the end-to-end training of our model implicitly learns a representation of the data that is optimized towards predicting the probability that a site is modified on the basis of the assumption encoded within the pooling layer. In our case, the pooling layer represents the probability that a particular site contains at least one modified position, but in practice one can choose a pooling layer that best captures the labeling process associated with the data collection step. While we apply m6Anet to the task of m6A RNA-modification detection, the framework can also

Read the complete text