From Sequence to Decision: Large Language Models as End-to-End Engines for Therapeutic RNA Design

Abstract

RNA-based therapeutics have achieved remarkable clinical validation, yet rational sequence design remains a fundamental bottleneck. Large language models (LLMs) trained on tens of millions of RNA sequences are beginning to reframe this challenge, offering a unified computational framework spanning structure prediction, immunogenicity optimization, and de novo sequence generation. Here, we survey the expanding landscape of RNA-focused foundation models — from early BERT-based encoders such as RNA-FM and ERNIE-RNA to large generative systems including GRAPE-LM — and map their end-to-end utility across six therapeutic modalities: messenger RNA, siRNA, antisense oligonucleotides, aptamers, circular RNA, and CRISPR guide RNAs. We identify three critical translation barriers: the RNA structural data famine, the absence of chemical-modification-aware architectures, and the lack of in vivo benchmarking datasets. We propose a roadmap — anchored by community data-generation initiatives and pre-competitive regulatory standards — for translating LLM-driven RNA design into clinical practice.