Deep and accurate detection of m6A RNA modifications using miCLIP2 and m6Aboost machine learning

Abstract

N6-methyladenosine (m6A) is the most abundant internal RNA modification in eukaryotic mRNAs and influences many aspects of RNA processing. miCLIP (m6A individual-nucleotide resolution UV crosslinking and immunoprecipitation) is an antibody-based approach to map m6A sites with single-nucleotide resolution. However, due to broad antibody reactivity, reliable identification of m6A sites from miCLIP data remains challenging. Here, we present miCLIP2 in combination with machine learning to significantly improve m6A detection. The optimized miCLIP2 results in high-complexity libraries from less input material. Importantly, we established a robust computational pipeline to tackle the inherent issue of false positives in antibody-based m6A detection. The analyses were calibrated with Mettl3 knockout cells to learn the characteristics of m6A deposition, including m6A sites outside of DRACH motifs. To make our results universally applicable, we trained a machine learning model, m6Aboost, based on the experimental and RNA sequence features. Importantly, m6Aboost allows prediction of genuine m6A sites in miCLIP2 data without filtering for DRACH motifs or the need for Mettl3 depletion. Using m6Aboost, we identify thousands of high-confidence m6A sites in different murine and human cell lines, which provide a rich resource for future analysis. Collectively, our combined experimental and computational methodology greatly improves m6A identification.

Prediction of m6A and m5C at single-molecule resolution reveals a cooccurrence of RNA modifications across the transcriptome

“…The two most common non-DRACH motifs identified by CHEUI-solo were 5’-GGACG-3’ (203 unique genomic positions) and 5’-GGATT-3’ (121 unique genomic sites). These motifs coincided with the two most common non-DRACH motifs identified by miCLIP2 experiments in the same cell lines and occurred at 245 (5’-GGACG-3’) and 96 (5’-GGATT-3’) sites (Körtel et al 2021). Most of the sites with DRACH or non-DRACH motifs occurred in mRNAs.…”

Deep learning modeling m6A deposition reveals the importance of downstream cis-element sequences

“…To make sure that the iM6A model was accurate for all m 6 A sites independent of the experimental methods that precisely mapped them, we examined whether iM6A could identify m 6 A sites mapped by alternative experimental methods (Supplementary Fig. 1g ) including m6A-label-seq 36 , MAZTER-seq 34 , m6ACE-seq 35 , and miCLIP2 47 . The m6A-label-seq method detected m 6 A sites by chemically substituting the m 6 A with a 6 A ( N 6 -allyladenosine) at the m 6 A sites, MAZTER-seq identified a relatively small subset of m 6 A sites that were in the m 6 ACA motifs by a methyl-sensitive RNase, and m6ACE-seq detected m 6 A sites by its crosslinking to the m 6 A-antibody and followed with the exonuclease digest to achieve single-base resolution.…”

“…The m6A-label-seq method detected m 6 A sites by chemically substituting the m 6 A with a 6 A ( N 6 -allyladenosine) at the m 6 A sites, MAZTER-seq identified a relatively small subset of m 6 A sites that were in the m 6 ACA motifs by a methyl-sensitive RNase, and m6ACE-seq detected m 6 A sites by its crosslinking to the m 6 A-antibody and followed with the exonuclease digest to achieve single-base resolution. In addition, miCLIP2 was an optimized CLIP method that combined miCLIP with machine learning to improve m 6 A detection 47 . The precisely mapped m 6 A sites by all these alternative experimental methods were identified with high probability values by iM6A (Fig.…”

RNA modification mapping with JACUSA2

“…To demonstrate this transfer, we use a murine model system, mouse embryonic stem cells (mESC). Jenjaroenpun et al [ 7 ] has sequenced mESC mRNA on the Nanopore and Köertel et al [ 16 ] mapped m6A residues via a novel miCLIP approach (antibody-based reference set). Additional file 1 : Fig.…”