Hybrid BiLSTM machine learning with RNAfold-based thermodynamic modeling for RNA secondary structure prediction

Abstract

RNA secondary structure plays an important role in various biological fields, such as gene regulation, catalysis, and RNA–protein interactions, yet accurate prediction is challenging due to long-range base-pairing and structural complexity. This work presents a hybrid framework that integrates RNAfold's thermodynamic modeling with machine learning to improve RNA secondary structure prediction and interpretability. Using a balanced dataset of approximately 9,700 RNA sequences spanning multiple RNA families (≈ 4.09 million nucleotides), traditional thermodynamic prediction with RNAfold was evaluated alongside a bidirectional long short-term memory (BiLSTM) neural network trained for base-wise pairing prediction. While RNAfold achieved a baseline test accuracy of 0.77, the BiLSTM reached 0.91 accuracy. Building on these results, several hybrid approaches were developed that selectively combine RNAfold and neural predictions, including a base-wise selector, a sequence-level meta-learner, and a Monte Carlo (MC) dropout uncertainty method. The best-performing hybrid model, the MC dropout uncertainty method, achieved a test accuracy of 0.917, outperforming both standalone approaches. This framework has been deployed through an interactive web interface, enabling users to input RNA sequences and compare prediction methods in real-time. To enhance interpretability, predicted structures are converted from dot-bracket notation into annotated visual diagrams depicting the corresponding secondary structure motifs. This study demonstrates that hybrid modeling with uncertainty-aware selection can improve RNA secondary structure prediction while maintaining accessibility and interpretability for later biological analysis.