Capturing Natural Evolution in Function-guided RNA Design via Genomic Foundation Models

RNAs play critical roles in gene expression regulation, catalysis, genetic information transmission, and immune response. However, current RNA engineering approaches face limitations in designing RNA for efficient in vivo folding and functional execution. Here, we present a novel, zero-shot strategy that integrates genomic large language models (LLMs) with inverse folding models (IFMs) to generate RNA sequences that mirror natural evolutionary patterns while preserving the structural integrity of critical functional regions. Benchmarking 9 state-of-the-art genomic foundation models, we show that their unsupervised log-likelihood scores correlate strongly with RNA fitness, enabling zero-shot sequence design. Applied to fluorescent RNA aptamers, our approach yielded 20 Broccoli variants, over half of which achieved up to 155% in fluorescence and a twofold increase in binding affinity, with robust performance both in vitro and in vivo. For Pepper, two rounds of optimization produced 40 variants, 16 of which demonstrated up to 2.6 fold in fluorescence and a threefold boost in binding affinity, including five with superior in vivo brightness.This approach offers a scalable, efficient, and resource-light method for RNA design, with significant implications for RNA-based therapeutics and diagnostics. By leveraging general foundation models to facilitate RNA evolution in a zero-shot fashion, we believe our work represents a major advancement in the field of RNA engineering.