Author:
Li Yupeng,Wang Fan,Yang Jiaqi,Han Zirong,Chen Linfeng,Jiang Wenbing,Zhou Hao,Li Tong,Tang Zehua,Deng Jianxiang,He Xin,Zha Gaofeng,Hu Jiekai,Hu Yong,Wu Linping,Zhan Changyou,Sun Caijun,He Yao,Xie Zhi
Abstract
ABSTRACTMessenger RNA (mRNA) therapeutics show immense promise, but their efficacy is limited by suboptimal protein expression. Here, we present RiboCode, a deep learning framework that generates mRNA codon sequences for enhanced protein production. RiboCode introduces several advances, including direct learning from large-scale ribosome profiling data, context-aware mRNA optimization and generative exploration of a large sequence space.In silicoanalysis demonstrate RiboCode’s robust predictive accuracy for unseen genes and cellular environments.In vitroexperiments show substantial improvements in protein expression, with up to a 72-fold increase, significantly outperforming past methods. In addition, RiboCode achieves cell-type specific expression and demonstrates robust performance across different mRNA formats, including m1Ψ-modified and circular mRNAs, an important feature for mRNA therapeutics.In vivomouse studies show that optimized influenza hemagglutinin mRNAs induce ten times stronger neutralizing antibody responses against influenza virus compared to the unoptimized sequence. In an optic nerve crush model, optimized nerve growth factor mRNAs achieve equivalent neuroprotection of retinal ganglion cells at one-fifth the dose of the unoptimized sequence. Collectively, RiboCode represents a paradigm shift from rule-based to data-driven, context-sensitive approach for mRNA therapeutic applications, enabling the development of more potent and dose-efficient treatments.
Publisher
Cold Spring Harbor Laboratory