The rational design of mRNA vaccine: From empirical method to artificial intelligence-based design

mRNA vaccines, recognized for their strong immunogenicity, low risk of gene integration, and cost-effective manufacturing, hold significant promise in preventing and treating infectious diseases and cancers. Indeed, mRNA vaccines were among the earliest vaccine platforms developed in response to the COVID-19 pandemic, demonstrating robust immunogenicity and playing a crucial role in protecting countless lives from COVID-19 infection. However, they also face practical challenges such as poor stability and suboptimal protein expression efficiency. To overcome these challenges, extensive research has focused on the design and optimization of mRNA vaccines. Traditionally, this design process has relied on iterative empirical refinements. With advances in artificial intelligence (AI) and bioinformatics, mRNA vaccine design is evolving toward more rational and efficient approaches. Thus, this review systematically introduces the characteristics of various vaccines and outlines the principles of mRNA vaccine design. We then focus on the optimization of mRNA vaccines, examining both empirical and AI-based design methods across five key structural domains of mRNA, including the 5' cap, 5' untranslated region (UTR), 3' UTR, coding sequence region, and poly-A tail. Traditional empirical methods involve designing and optimizing these regions to address some of the inherent deficiencies in mRNA vaccines. However, these methods often require repetitive experimental iterations, resulting in low development efficiency and high costs. Currently, the evolvement of vaccines is rapidly being revolutionized using advanced AI-based technologies. AI models can rapidly and efficiently optimize and generate highly druggable mRNA sequences by learning from publicly available or proprietary biological data. By comparing empirical and AI-based design approaches, we highlight the advantages of AI in mRNA vaccine design while also discussing its limitations and future potential. Centainly, AI models also face certain challenges when applied to mRNA design. Firstly, AI models require large-scale and high-quality experimental data for training, but currently, the available experimental data is limited in quantity, varies in quality, and is also scattered. Secondly, the interpretability of AI models is relatively poor, as they are often referred to as “black box” models, making it difficult to explain the decision-making processes of AI models. With the development of AI technology and the accumulation of biological data, the predictive capabilities of AI models will keep improving. It is even more foreseeable that base models based on large language models, such as AlphaFold3, targeting biomolecules like RNA and proteins, will play a significant role in drug development. AI models will be able to more comprehensively analyze the interaction patterns between RNA and other biomolecules like DNA and proteins, thereby further enhancing the effectiveness of mRNA design. In conclusion, we argue that personalized and precise mRNA design driven by AI could revolutionize the biomedical field, offering unprecedented therapeutic possibilities for patients, enhancing the vaccine development process, and providing new strategies to address future challenges.