REDfold

REDfold predicts RNA secondary structure from sequence using a convolutional encoder-decoder deep learning model to improve accuracy and scalability, including prediction of pseudoknots for studies of RNA stability and function.

Key Features:

• Encoder-decoder CNN: Uses a convolutional neural network-based encoder-decoder architecture with symmetric skip connections to capture short- and long-range dependencies in RNA sequences.
• Pseudoknot handling: Applies a constrained optimization post-processing step to produce valid secondary structures that include pseudoknots.
• Improved scalability: Addresses limitations of thermodynamic models combined with dynamic programming by providing improved performance and scalability for complex structures.
• Benchmarking: Experimental evaluations on the ncRNA database demonstrate higher efficiency and accuracy compared with contemporary state-of-the-art methods.

Scientific Applications:

• RNA structure prediction for functional studies: Infers secondary structures to inform analyses of RNA stability and function.
• Large-scale RNA analyses: Enables high-throughput or large-scale secondary structure prediction that is impractical with dynamic programming approaches.
• Pseudoknot analysis: Predicts and analyzes RNAs containing pseudoknots for studies requiring non-nested base-pairing information.

Methodology:

Convolutional encoder-decoder CNN with symmetric skip connections to model sequence dependencies, followed by constrained optimization post-processing; evaluated on the ncRNA database.