Deep3MVPF: Multiview Deep Framework for the Prediction of Stability and m6A in mRNA 3'UTR.

Abstract

Accurate prediction of mRNA stability and identification of N6-methyladenosine (m6A) sites are central to understanding post-transcriptional regulation. Because the 3' untranslated region (3'UTR) contains both stability-associated cis-elements and many m6A sites, it provides a suitable context for modeling RNA regulatory effects. However, most existing methods rely primarily on linear sequence information and do not adequately capture higher-order topology or RNA structural context. Here, we present Deep3MVPF, a multiview deep learning framework for 3'UTR stability prediction and m6A site identification. Deep3MVPF integrates a multiscale convolutional neural network, a k-mer de Bruijn graph neural network, and a secondary-structure graph neural network to jointly model sequence, topological, and structural representations. For 3'UTR stability prediction, the model was trained and evaluated on a zebrafish (Danio rerio) mRNA degradation data set and achieved an MSE of 0.0049. For m6A site identification, it was evaluated on nine human cell line data sets and achieved an average AUC of 0.970. Attribution analysis further showed that Deep3MVPF recovered regulatory features consistent with known biology, including the destabilizing GCACUU motif and stabilizing G-rich/G-quadruplex-associated signals. These results demonstrate that integrating heterogeneous RNA representations can improve predictive modeling and facilitate interpretation of post-transcriptional regulatory grammar.