DENA: training an authentic neural network model using Nanopore sequencing data of Arabidopsis transcripts for detection and quantification of N6-methyladenosine on RNA

Longxian Chen,Liang Ou,Jian Gao,Jia-Wei Wang,Pei Hao,Hang Qin,Xuan Li

doi:10.1186/s13059-021-02598-3

Abstract

Models developed using Nanopore direct RNA sequencing data from in vitro synthetic RNA with all adenosine replaced by N6-methyladenosine (m6A) are likely distorted due to superimposed signals from saturated m6A residues. Here, we develop a neural network, DENA, for m6A quantification using the sequencing data of in vivo transcripts from Arabidopsis. DENA identifies 90% of miCLIP-detected m6A sites in Arabidopsis and obtains modification rates in human consistent to those found by SCARLET, demonstrating its robustness across species. We sequence the transcriptome of two additional m6A-deficient Arabidopsis, mtb and fip37-4, using Nanopore and evaluate their single-nucleotide m6A profiles using DENA.

Highlights

N6-methyladenosine (m6A) is the most abundant modification found in messenger RNA [1, 2]
We found the existence of clustered multiple mismatches within or around “RRACH” five-mers in m6A-modified reads of synthetic data (Fig. 1a)
In two “AAACC” five-mers from in vitro synthetic sequences, we observed distinguishing signals and consecutive mismatches within the “AAACC” pattern and its surrounding “A” residues (Fig. 1b). This situation was infrequent in direct RNA-Seq data of in vivo transcribed RNAs from A. thaliana (Fig. 1d), and only sporadic distinction of signals and mismatches were observed around the m6A sites

Summary

METHOD

DENA: training an authentic neural network model using Nanopore sequencing data of Arabidopsis transcripts for detection and quantification of N6-methyladenosine on RNA. Hang Qin1,2†, Liang Ou2,3†, Jian Gao, Longxian Chen, Jia-Wei Wang2,4*, Pei Hao2,3* and Xuan Li1,2*

Background

Results and discussion

Methods