Unlocking Your Programmable and Creative RNA Sequence Designer with RDiffusion

Abstract

RNA, a pillar of the central dogma, has shaped three billion years of evolution. Despite cataloging tens of millions of non-coding RNAs and annotating millions, we have barely scratched the surface of the vast, unexplored RNA sequence space. Here, we introduce RDiffusion, a discrete-diffusion-based generative transformer model for RNA sequence design. Conditioned on diverse biological features, such as desired function, family type, secondary and tertiary structure, or binding protein partners, it can guide the generation of novel RNA sequences tailored to specific design specifications. We evaluate RDiffusion across a broad spectrum of RNA design tasks and find that it not only surpasses all baseline methods in design success rate and sequence diversity but also achieves advanced performance on downstream tasks. To demonstrate its translational potential, we applied RDiffusion with a customized seed-screening pipeline to de novo design therapeutic MicroRNA(miRNA) mimics in osteoarthritis (OA). This strategy identified mimic 0-4 as a lead candidate targeting SDC4 to preserve ECM homeostasis and cartilage integrity, with robust efficacy validated in clinical OA samples. Altogether, RDiffusion holds strong potential to serve as a cornerstone for RNA generation and representation learning, establishing a generalizable AI-driven paradigm for RNA therapeutic discovery.