How Do You Identify m6 A Methylation in Transcriptomes at High Resolution? A Comparison of Recent Datasets

Abstract

A flurry of methods has been developed in recent years to identify N6-methyladenosine (m⁶A) sites across transcriptomes at high resolution. This raises the need to understand both the common features and those that are unique to each method. Here, we complement the analyses presented in the original papers by reviewing their various technical aspects and comparing the overlap between m⁶A-methylated messenger RNAs (mRNAs) identified by each. Specifically, we examine eight different methods that identify m⁶A sites in human cells with high resolution: two antibody-based crosslinking and immunoprecipitation (CLIP) approaches, two using endoribonuclease MazF, one based on deamination, two using Nanopore direct RNA sequencing, and finally, one based on computational predictions. We contrast the respective datasets and discuss the challenges in interpreting the overlap between them, including a prominent expression bias in detected genes. This overview will help guide researchers in making informed choices about using the available data and assist with the design of future experiments to expand our understanding of m⁶A and its regulation.

Keywords: N6-methyladenosine; RNA; bioinformatics; epitranscriptomics; m6A.

FIGURE 1

High throughput methods to detect or predict m⁶A in transcriptomes. (A) Crosslinking and immunoprecipitation (CLIP) methods involve UV crosslinking of the m⁶A antibody to purified RNA. m⁶A-CLIP and miCLIP differ in the antibodies used, complementary DNA (cDNA) library preparation, and computational processing, among other differences. (B) MazF Escherichia coli endoribonuclease preferentially cuts at nonmethylated ACA sites. This forms the basis of MAZTER-seq and m⁶A-REF-seq. (C) DART-seq expresses an APOBEC1-YTH fusion protein. The YTH domain targets APOBEC1 to m⁶A sites, where it deaminates surrounding cytosines to uracil. (D) Direct RNA sequencing with Nanopore technologies facilitates detection of m⁶A due to differences in ionic current intensities between A- and m⁶A-containing sequences and dwell time in the pore. Methods differ by how these signals are deconvolved. m⁶A identification using nanopore sequencing (MINES) is a combination of four random forest models, pretrained using CLIP m⁶A sites as true positives. NanoCompore relies on a comparison in signal between two conditions, for example wild type (WT) and METTL3 knockdown, or in vivo RNA vs. nonmodified in vitro transcribed RNA. (E) In silico prediction of m⁶A sites is performed by WHISTLE, a support vector machine algorithm that uses miCLIP and m⁶A-CLIP sites as training data.

FIGURE 2

m⁶A-containing genes identified by eight methods. (A) Bar chart showing the number of m⁶A-containing transcripts identified by each method. Some methods have data from multiple cell lines or apply several possible thresholds, which are shown separately. The cell lines for each dataset are indicated along with the type of method. The hashed bars denote genes that are commonly expressed between all the cell lines considered here. For DART-Seq, MAZTER-Seq, and MINES, several thresholds were possible: “DART-Seq M3” refers to sites identified by comparison with METTL3 knockdown. “Low” and “high” refer to two stringency thresholds applied by the authors. “MAZTER-Seq” refers to all sites with a cleavage efficiency <50%, and “MAZTER-Seq cond” refers to FTO overexpression, WT ≥ 20%, and/or Alkbh5 overexpression, WT ≥ 20%. “MINES” refers to all sites identified by MINES, and “MINES 30×” refers to MINES sites with ≥ = 30× coverage. (B) Bar chart showing the numbers of overlapping target genes between the eight methods, considering all the reported genes.

FIGURE 3

Comparing the top-ranking target genes identified by eight methods. (A) Bar chart showing the numbers of top-ranking genes that overlap between the eight methods. (B) Heatmap showing overlap between the top targets. Dendrograms are produced by complete-linkage hierarchical clustering using the Jaccard index as the distance metric. Dark blue indicates presence of the gene among the top targets for a method, and gray indicates absence. Colored bars denote the category of the method. (C) Proportions of top targets that are unique to each method. (D) Number of methods detecting a target gene plotted against its mean expression decile across all studied cell lines. (E) Minimum expression deciles for the top ranked genes were plotted for each method.