Deciphering the comprehensive relationship between 5' UTR and 3' UTR sequences with deep learning.

Abstract

Recent advances in mRNA therapeutics have driven further research on the untranslated regions (UTRs) of mRNA. However, prior studies have mainly focused on either the 5' or 3' UTR individually. Increasing evidence suggests potential cooperative effects between these two regions, which remain largely unexplored in computational studies. We present a deep learning-based approach to predicting relationships between 5' and 3' UTRs by leveraging latent representations from a pre-trained RNA language model and contrastive learning. Our method effectively identifies highly related UTRs, uncovering sequence and expression characteristics that suggest functional interplay. Our analysis revealed that Highly Related UTRs (HRUs) are significantly enriched in genes associated with neural development, exhibit distinctive UTR length and secondary structure characteristics, and are involved in cell type-specific regulation of translation efficiency. These findings provide new insights into UTR co-optimization for mRNA therapeutics. The source code is available for free at https://github.com/hmdlab/utr\_pairpred.git. The data and intermediate files used in our analysis are available at https://waseda.box.com/v/utr-pairpred-data.