Nature Machine Intelligence ( IF 23.8 ) Pub Date : 2020-08-03 , DOI: 10.1038/s42256-020-0211-4 Fei Tan , Tian Tian , Xiurui Hou , Xiang Yu , Lei Gu , Fernanda Mafra , Brian D. Gregory , Zhi Wei , Hakon Hakonarson
Research on DNA methylation on N6-adenine (6mA) in eukaryotes has received much recent attention. Recent studies have generated a large amount of 6mA genomic data, yet the role of DNA 6mA in eukaryotes remains elusive, or even controversial. We argue that the sparsity of DNA 6mA in eukaryotes, the limitations of current biotechnologies for 6mA detection and the sophistication of the 6mA regulatory mechanism together pose great challenges for elucidation of DNA 6mA. To exploit existing 6mA genomic data and address this challenge, here we develop a deep-learning-based algorithm for predicting potential DNA 6mA sites de novo from sequence at single-nucleotide resolution, with application to three representative model organisms, Arabidopsis thaliana, Drosophila melanogaster and Escherichia coli. Extensive experiments demonstrate the accuracy of our algorithm and its superior performance compared with conventional k-mer-based approaches. Furthermore, our saliency maps-based context analysis protocol reveals interesting cis-regulatory patterns around the 6mA sites that are missed by conventional motif analysis. Our proposed analytical tools and findings will help to elucidate the regulatory mechanisms of 6mA and benefit the in-depth exploration of their functional effects. Finally, we offer a complete catalogue of potential 6mA sites based on in silico whole-genome prediction.
中文翻译:
通过深度学习阐明N 6-腺嘌呤上的DNA甲基化
真核生物中N 6-腺嘌呤(6mA)的DNA甲基化研究受到了广泛关注。最近的研究产生了大量的6mA基因组数据,但DNA 6mA在真核生物中的作用仍然难以捉摸,甚至有争议。我们认为,真核生物中DNA 6mA的稀疏性,当前6mA检测生物技术的局限性以及6mA调控机制的复杂性,对DNA 6mA的阐明提出了巨大的挑战。为了利用现有的6mA基因组数据并应对这一挑战,我们在此开发了一种基于深度学习的算法,可从单核苷酸分辨率的序列中预测从头开始的潜在DNA 6mA位点,并将其应用于三种代表性的模式生物拟南芥(Arabidopsis thaliana)果蝇和大肠杆菌。与传统的基于k -mer的方法相比,大量实验证明了我们算法的准确性及其优越的性能。此外,我们基于显着性图的情境分析协议揭示了6mA位点周围有趣的顺式调控模式,而传统的基序分析则错过了这种模式。我们提出的分析工具和发现将有助于阐明6mA的调节机制,并有助于深入探索其功能效果。最后,我们根据计算机模拟全基因组预测提供了完整的6mA潜在位点目录。