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

<jats:title>Abstract</jats:title><jats:p>The <jats:italic>N</jats:italic><jats:sup>6</jats:sup>-methyladenosine (m<jats:sup>6</jats:sup>A) modification is deposited to nascent transcripts on chromatin, but its site-specificity mechanism is mostly unknown. Here we model the m<jats:sup>6</jats:sup>A deposition to pre-mRNA by iM6A (<jats:underline>i</jats:underline>ntelligent m<jats:sup>6</jats:sup>A), a deep learning method, demonstrating that the site-specific m<jats:sup>6</jats:sup>A methylation is primarily determined by the flanking nucleotide sequences. iM6A accurately models the m<jats:sup>6</jats:sup>A deposition (AUROC = 0.99) and uncovers surprisingly that the <jats:italic>cis</jats:italic>-elements regulating the m<jats:sup>6</jats:sup>A deposition preferentially reside within the 50 nt downstream of the m<jats:sup>6</jats:sup>A sites. The m<jats:sup>6</jats:sup>A enhancers mostly include part of the RRACH motif and the m<jats:sup>6</jats:sup>A silencers generally contain CG/GT/CT motifs. Our finding is supported by both independent experimental validations and evolutionary conservation. Moreover, our work provides evidences that mutations resulting in synonymous codons can affect the m<jats:sup>6</jats:sup>A deposition and the TGA stop codon favors m<jats:sup>6</jats:sup>A deposition nearby. Our iM6A deep learning modeling enables fast paced biological discovery which would be cost-prohibitive and unpractical with traditional experimental approaches, and uncovers a key <jats:italic>cis</jats:italic>-regulatory mechanism for m<jats:sup>6</jats:sup>A site-specific deposition.</jats:p>