RESM: Capturing sequence and structure encoding of RNAs by mapped transfer learning from ESM (evolutionary scale modeling) protein language model

RNA sequences exhibit lower evolutionary conservation than proteins due to their informationally constrained four-letter alphabet, compared to the 20-letter code of proteins. More limited information makes unsupervised learning of structural and functional evolutionary patterns more challenging from single RNA sequences. We overcame this limitation by mapping RNA sequences to pseudo-protein sequences to allow effective transfer training from a protein language model (protein Evolution-Scale Model 2, protESM-2). The resulting RNA ESM (RESM) outperforms 12 existing RNA language models in zero-shot prediction, not only in sequence classification but also in RNA secondary structure and RNA-RNA interaction prediction. Further supervised fine-tuning demonstrates RESMs generalizability and superior performance over the existing models compared across multiple downstream tasks, including mRNA ribosome loading efficiency and gene expression prediction, despite RESM being trained exclusively on noncoding RNAs. Moreover, RESM can generalize to unseen sequences beyond its 1,024-nucleotide training limit, achieving 81.3% improvement over state-of-the-art methods in supervised secondary structure prediction for RNAs up to 4,000 nucleotides, limited only by the available GPU memory, while providing >1000-fold speedup compared to MSA-based approaches. RESM provides a robust foundation for deciphering RNA sequence-structure-function relationships, with broad implications for RNA biology.