UTR-Insight: integrating deep learning for efficient 5′ UTR discovery and design
Saichao Pan, Hanyu Wang, Hang Zhang, Zan Tang, Lianqiang Xu, Zhixiang Yan, Yong Hu
Abstract
The 5' UTR is critical for mRNA stability and translation efficiency in therapeutics. We developed UTR-Insight, a model integrating a pretrained language model with a CNN-Transformer architecture, explaining 89.1% of the mean ribosome load (MRL) variation in random 5' UTRs and 82.8% in endogenous 5' UTRs, surpassing existing models. Using UTR-Insight, we performed high-throughput in silico screening of hundreds of thousands of endogenous 5' UTRs from primates, mice, and viruses. The screened sequences increased protein expression by up to 319% compared to the human α-globin 5' UTR, and UTR-Insight-designed sequences achieved even greater expression levels than high-performing endogenous 5' UTRs.