Programmable design of functional RNAs through multimodal generative learning

Abstract

RNA engineering is bottlenecked by the tight coupling among sequence, structure, and function, leaving the field reliant on heuristics or screening. We introduce RNAverse, a multimodal generative model that learns a conditionable design grammar from RNA sequences, experimental structural probing profiles, and functional annotations. Beyond setting a new state of the art across 16 predictive modeling tasks, RNAverse inverts into generative design across four distinct RNA classes through three regimes, requiring experimental testing of no more than 20 candidates per-task. Zero-shot in-painting generates snoRNAs that promote premature termination codon (PTC) readthrough, and chimeric aptazymes for ligand-controlled gene regulation, exceeding rational design or large-scale screens. In both settings, RNAverse outperforms family-adapted sequence-only baselines. Fine-tuned in-painting extends to an unseen functional class, yielding IGFBP3-binding aptamers with picomolar affinity that outperform SELEX winners. Iterative optimization engineers 5' UTRs that outperform clinical benchmarks, enhancing protein production in cell-based assays for both the Zolgensma gene therapy and the BNT162b2 COVID-19 vaccine, and boosting antibody responses to BNT162b2 in mice. Together, these results establish multimodal generative learning as a route from predictive RNA modeling to programmable specification of functional RNAs.