WHISTLE: a high-accuracy map of the human N6-methyladenosine (m6A) epitranscriptome predicted using a machine learning approach

Abstract

N 6 -methyladenosine (m 6 A) is the most prevalent post-transcriptional modification in eukaryotes, and plays a pivotal role in various biological processes, such as splicing, RNA degradation and RNA–protein interaction. We report here a prediction framework WHISTLE for transcriptome-wide m 6 A RNA-methylation site prediction. When tested on six independent datasets, our approach, which integrated 35 additional genomic features besides the conventional sequence features, achieved a major improvement in the accuracy of m 6 A site prediction (average AUC: 0.948 and 0.880 under the full transcript or mature messenger RNA models, respectively) compared to the state-of-the-art computational approaches MethyRNA (AUC: 0.790 and 0.732) and SRAMP (AUC: 0.761 and 0.706). It also out-performed the existing epitranscriptome databases MeT-DB (AUC: 0.798 and 0.744) and RMBase (AUC: 0.786 and 0.736), which were built upon hundreds of epitranscriptome high-throughput sequencing samples. To probe the putative biological processes impacted by changes in an individual m 6 A site, a network-based approach was implemented according to the ‘guilt-by-association’ principle by integrating RNA methylation profiles, gene expression profiles and protein–protein interaction data. Finally, the WHISTLE web server was built to facilitate the query of our high-accuracy map of the human m 6 A epitranscriptome, and the server is freely available at: www.xjtlu.edu.cn/biologicalsciences/whistle and http://whistle-epitranscriptome.com .

Expression Status And Prognostic Value Of M6A-associated Genes in Gastric Cancer

“…N6-methyladenosine (m6A) modification is the methylation of the adenosine base at the nitrogen-6 position of mRNA which was first discovered as an abundant nucleotide modification in eukaryotic messenger RNA in 1974 [7][8][9]. M6A modifications is regulated by three types of enzymes: "writers" (methyltransferases, including WTAP, KIAA1429, RBM15/15B, and METTL3/14/16), "readers" (YTH domain containing RNA binding proteins and heterogeneous nuclear ribonucleoprotein, including YTHDF1/2/3, YTHDC1, HNRNPC and HNRNPA2B1) and "erasers" (demethylases, including ALKBH5 and FTO) [10][11][12].…”

“…It plays a pivotal role in regulating precursor mRNA maturation, translation and degradation [14]. In addition, m6A modification could also affect tissue development [15], cell self-renewal and differentiation, control of heat shock response [16], DNA damage response [17], circadian clock controlling and the development of multiple forms of human diseases, including cancer [7]. Emerging evidence has demonstrated that m6A modification play a critical role in a great variety of human cancers [14], including breast cancer [18,19], lung cancer [20], acute myeloid leukemia (AML) [21,22], glioblastoma [23], hepatoblastoma [24], colorectal cancer [25] and so on [14].…”

Predicting sites of epitranscriptome modifications using unsupervised representation learning based on generative adversarial networks

“…This problem is naturally a supervised learning task, which aims to train predictive models for each type by using labeled positive and negative modification sites. There is a large collection of algorithms for predicting m 6 A sites from mRNA sequences [4][5][6][7][8][9][10], most notably SRAMP. However, such predictive algorithms for other modifications are still scarce because training robust models for these modification sites face several challenges [11,12].…”

SICM6A: Identifying m6A Site across Species by Transposed GRU Network

“…The RNA secondary structure is analyzed before SRAMP prediction, and the calculation amount is very large, so the feature extraction time is very long. WHISTLE uses support vector machines to predict N6-methyladenosine sites in human epithelial cells, but it has over fitting problems and only provides query service for m 6 A sites 19 . BERMP 20 introduced deep learning BGRU net to the prediction of m 6 A sites on SRAMP data sets for the first time, the calculation time can be reduced and the deep learning model shows a particular application prospect.…”