Enhanced Generalizability of RNA Secondary Structure Prediction via Convolutional Block Attention Network and...

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• Accurate RNA secondary structure (RSS) prediction is vital for understanding RNA function, therapeutic design, and synthetic biology.
• Traditional RSS determination methods (e.g., NMR) are laborious; existing computational methods, especially deep learning (DL), face generalizability issues due to overfitting and prediction inconsistencies.

Purpose of the Study:

• To develop a novel computational method, TrioFold, that improves the generalizability of RNA secondary structure prediction.
• To integrate base-pairing information from both thermodynamic and DL-based methods to overcome individual limitations.

Main Methods:

• Ensemble learning was employed to combine base-pairing clues from thermodynamic and DL models.
• A convolutional block attention mechanism was incorporated to enhance feature learning and integration.
• An online webserver was developed to provide accessible RSS prediction tools.

Main Results:

• TrioFold demonstrated higher accuracy in intra-family RSS predictions.
• The method exhibited enhanced generalizability in inter-family and cross-RNA-type predictions compared to existing approaches.
• The developed webserver offers a unified platform for RNA structure prediction and analysis.

Conclusions:

• TrioFold effectively improves the generalizability of RNA secondary structure prediction through ensemble learning of diverse base-pairing clues.
• The integration of thermodynamic and DL-based insights offers a promising direction for advancing RNA structure prediction.
• The accessible webserver facilitates broader adoption and application within the RNA research community.