Diverse database and machine learning model to narrow the generalization gap in RNA structure prediction

Understanding macromolecular structures of proteins and nucleic acids is critical for discerning their functions and biological roles. Advanced techniques-crystallography, nuclear magnetic resonance, and cryo-electron microscopy-have facilitated the determination of more than 180,000 protein structures, all cataloged in the Protein Data Bank. This comprehensive repository has been pivotal in developing deep learning algorithms for predicting protein structures directly from sequences. In contrast, RNA structure prediction has lagged and suffers from a scarcity of structural data. Here, we present the secondary structure models of 1098 primary microRNAs and 1456 human messenger RNA regions determined through chemical probing. We develop a deep learning architecture inspired from the Evoformer model of Alphafold and traditional architectures for secondary structure prediction. This model, eFold, was trained on our newly generated database and more than 300,000 secondary structures across multiple sources. We benchmark eFold on two challenging test sets of long and diverse RNA structures and show that our dataset and architecture contribute to increasing the prediction performance, compared to similar state-of-the-art methods. Together, our results reveal that merely expanding the database size is insufficient for generalization across families, whereas incorporating a greater diversity and complexity of RNA structures allows for enhanced model performance.