MetaFold-RNA: Accurate prediction of RNA secondary structure using a meta-learning-guided deep network

Abstract

Accurately predicting the secondary structure of ribonucleic acid (RNA) is a critical step toward deciphering its biological roles and engineering novel RNA-based technologies. However, achieving high accuracy and generalization, especially for unseen RNA families, has long remained a central challenge in computational biology. Here we present MetaFold-RNA, a novel meta-learning framework that sets a new state of the art in RNA secondary structure prediction. Instead of relying on a single predictive paradigm, MetaFold-RNA synergizes insights from disparate models through a meta-learning approach and refines predictions via a co-evolutionary optimization process. On the challenging bpRNA-new benchmark, which specifically tests performance on unseen RNA families, MetaFold-RNA surpasses all SOTA methods, outperforming the best previous approach by 7.3%. This breakthrough performance is consistently maintained across other stringent assessments, including the PDB-new dataset and blind CASP predictions. Furthermore, our method demonstrates high robustness by accurately resolving the complex topologies of artificial RNA nanostructures. This work establishes a new performance benchmark, particularly in cross-family generalization, and provides a powerful computational tool for accelerating the study and design of RNA.