Deep generative models design mRNA sequences with enhanced translational capacity and stability.

Despite the success of messenger RNA (mRNA) COVID-19 vaccines, extending this modality to more diseases necessitates substantial enhancements. We present GEMORNA, a generative RNA model that uses transformer architectures tailored for mRNA coding sequences (CDSs) and untranslated regions (UTRs) to design mRNAs with enhanced expression and stability. GEMORNA-designed full-length mRNAs exhibited up to a 41-fold increase in firefly luciferase expression compared with an optimized benchmark in vitro. GEMORNA-generated therapeutic mRNAs achieved up to a 15-fold enhancement in human erythropoietin (EPO) expression and substantially elicited antibody titers of COVID vaccine in mice. Additionally, GEMORNA's versatility extends to circular RNA, substantially enhancing circular EPO expression and boosting antitumor cytotoxicity in chimeric antigen receptor T cells. These advancements highlight the vast potential of deep generative artificial intelligence for mRNA therapeutics. Editor's summary: Generative models have advanced small-molecule and protein design, but their application to messenger RNA (mRNA) remains challenging because of the complexity of translational regulation and sequence-structure-function relationships. Zhang et al. present GEMORNA, a generative AI framework that learns intrinsic biological patterns from natural mRNA sequences and enables de novo sequence design with strong contextual coherence and diversity through autoregressive decoders. GEMORNA has demonstrated potential as a generalized mRNA design platform through case studies showing enhanced and prolonged protein expression in cell-based assays and animal models. These findings are relevant to current mRNA therapeutic applications, including mRNA vaccines, protein replacement therapies, and in vivo chimeric antigen receptor T cell treatments. —Di Jiang INTRODUCTION: mRNA therapeutics represent a transformative modality, but their broader application is limited by several challenges—in particular, the need to design sequences that achieve enhanced and durable protein expression for optimal therapeutic outcomes. Traditional design methods typically optimize one or two objectives, whereas existing deep-learning approaches often lack generalizability or focus on property prediction rather than generation, both failing to outperform commercial benchmarks. We introduce GEMORNA (generative models for RNA), a deep generative model capable of designing mRNA sequences with unprecedented translational capacity and durability. Extensive in vitro and in vivo evaluations demonstrate GEMORNA's ability to substantially enhance mRNA performance, highlighting its enormous potential for future therapeutic applications. RATIONALE: Previous RNA language models, typically based on BERT (bidirectional encoder representations from transformers)–style architectures, are better suited for property prediction than de novo generation and often yield suboptimal designs when combined with optimization methods such as genetic algorithms. GEMORNA addresses this limitation through a generative modeling approach, specifically adapted for RNA sequence design. For coding sequence (CDS) design, GEMORNA uses a transformer encoder-decoder architecture, framing the task as a language translation problem in which the encoder captures the semantic representation of the protein sequence, and the decoder generates a codon-constrained CDS that preserves the input protein sequence. By contrast, GEMORNA's untranslated region (UTR) design uses a decoder-only architecture, enabling greater flexibility to explore the UTR design space on the basis of the intrinsic patterns learned from natural sequences. These zero-shot high-performing CDS and UTR elements generated by GEMORNA are further integrated into full-length mRNA designs, offering a practical solution tailored to different protein targets and therapeutic applications. RESULTS: GEMORNA-generated firefly luciferase CDSs showed markedly higher expression levels as compared with sequences from other sources, and its UTR designs exhibited low sequence identity to natural UTRs yet consistently outperformed a broad range of benchmarks, including commercial products. Full-length mRNAs designed by GEMORNA demonstrated up to 41-fold improvement in the expression level of firefly luciferase. GEMORNA's full-length designs for therapeutic mRNAs demonstrated substantially increased expression levels, prolonged durability, and improved potency in mouse studies. GEMORNA COVID-19 mRNA vaccines doubled antibody titers relative to that with BNT162b2, and GEMORNA erythropoietin (EPO) constructs achieved 15- to 150-fold higher expression as compared with a leading benchmark. Furthermore, GEMORNA extends to circular RNA (circRNA) design, producing circRNAs with improved expression, increased stability, and superior antitumor activity in RNA-transduced chimeric antigen receptor T cells (CAR T cells) versus conventional codon optimization and top patented designs. CONCLUSION: GEMORNA is a generative artificial intelligence (AI) model that uses a tailored deep-learning architecture specifically optimized for therapeutic mRNA design, capable of producing effective sequences through zero-shot generation. Extensive experimental validation demonstrates that GEMORNA-derived sequences consistently outperform commercially available mRNA products and leading published designs in terms of expression level, durability, and potency, achieving a very high success rate. GEMORNA shows robust performance for both chemically modified mRNA sequences and circular RNA constructs. GEMORNA represents a promising and versatile platform to advance next-generation mRNA therapeutics and vaccines. GEMORNA designs superior mRNA sequences by leveraging deep generative AI models.: GEMORNA is a deep generative AI platform tailored for mRNA design, capable of generating sequences with enhanced expression and stability. Its designs have been experimentally validated across a range of therapeutic applications, including vaccine, gene therapy, and in vivo CAR T therapy. [Figure created with BioRender.com] [ABSTRACT FROM AUTHOR]