RNAformer: A Simple Yet Effective Deep Learning Model for RNA Secondary Structure Prediction

Predicting RNA secondary structure is essential for understanding RNA function and developing RNA-based therapeutics. Despite recent advances in deep learning for structural biology, its application to RNA secondary structure prediction remains contentious. A primary concern is the control of homology between training and test data. Moreover, deep learning approaches often incorporate complex multi-model systems, ensemble strategies, or require external data. Here, we present the RNAformer, a scalable axial-attention-based deep learning model designed to predict secondary structure directly from a single RNA sequence without additional requirements. We demonstrate the benefits of this lean architecture by learning an accurate biophysical RNA folding model using synthetic data. Trained on experimental data, our model overcomes previously reported caveats in deep learning approaches with a novel homology-aware data pipeline. The RNAformer achieves state-of-the-art performance on RNA secondary structure prediction, outperforming both traditional non-learning-based methods and existing deep learning approaches, while carefully considering sequence and structure similarities.