5UTRnet: A deep learning model for designing high-efficiency 5′ UTRs in mRNA therapeutics

Optimizing the 5′ untranslated region (5′ UTR) sequence to increase protein expression is crucial for mRNA medicines to reach their full therapeutic potential. Here, we developed 5UTRnet, a deep learning framework integrating both predictive and reconstructive capabilities for the tailored design of 5′ UTRs. We designed a Two-Stage training method for accurate prediction of Mean Ribosome Load (MRL). Additionally, we introduced an innovative Gradient-Generation optimization algorithm to engineer 5′ UTRs of diverse lengths with high MRLs. Experimental validation across different systems demonstrated that 5′ UTRs generated by 5UTRnet significantly enhanced protein expression compared to artificially designed benchmark sequence. The top-performing 5′ UTR sequence, UTRnet2, subsequently demonstrated its potential by eliciting superior specific antibody titers in a preventive vaccine model and inducing robust antitumor efficacy in a therapeutic vaccine setting, underscoring the strong translational potential of UTRnet2. Collectively, this study established 5UTRnet as a versatile and high-performance tool for rational 5′ UTR engineering, thereby accelerating the development of novel mRNA therapeutics.