Deep generative models generate mRNA sequences with enhanced translation capacity and stability

Abstract

Despite the tremendous success of messenger RNA (mRNA) COVID-19 vaccines, the extension of this modality to a broader spectrum of diseases necessitates substantial enhancements, particularly in the design of mRNAs with elevated expression levels and extended durability. Here we present GEMORNA, a deep generative model designed to generate novel mRNA coding sequences (CDSs) and untranslated regions (UTRs) with superior translation capacity, comparable to the sophisticated task of language translation and free-form poetry composition with accurate grammar and semantics. Our AI model was trained on an extensive collection of RNA sequences from diverse families, further enhanced with labeled data to refine its performance. Remarkably, we demonstrate that our AI-generated mRNAs exhibited 8.2-fold and 15.9-fold increases in firefly luciferase expression compared to benchmark mRNAs in two different cell types. Additionally, Our AI- designed COVID-19 mRNA vaccine elicited a 4-fold increase in anti-COVID antibody titer in mice relative to BNT162b2. Furthermore, GEMORNAs versatility extends to circular mRNA design, which we facilitated a 27-fold increase in human erythropoietin protein expression in vivo than a systematically optimized benchmark sequence. We also created circular mRNAs with substantial improvements in expression levels, durability and anti-tumor cell cytotoxicity in mRNA-transduced CAR-T cells compared with an experimentally validated benchmark. In summary, GEMORNA generates novel mRNA sequences with significant performance improvements and has the potential to enable a wide range of therapeutic and vaccine applications.