From sequence to structure: A comprehensive review of deep learning models for RNA structure prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Biology

Background:

• RNA structure prediction is vital for gene regulation, drug design, and synthetic biology.
• Deep learning has advanced protein structure prediction but faces challenges in RNA due to limited data, noncanonical interactions, and flexibility.
• Traditional methods include Direct Coupling Analysis and physics-based simulations.

Purpose of the Study:

• To review the evolution of RNA structure prediction methods.
• To systematically examine deep learning approaches for RNA secondary and tertiary structure prediction.
• To identify future research directions for improving RNA structure prediction.

Main Methods:

• Review of traditional physics-based methods (e.g., Direct Coupling Analysis, simulations).
• Systematic review of three deep learning paradigms: language models, end-to-end predictors, and geometry-distance predictors.
• Identification of future research needs, including advanced tokenization and explainable AI.

Main Results:

• Deep learning methods show significant progress in RNA structure prediction.
• Challenges remain, including data scarcity, noncanonical interactions, and conformational flexibility.
• Key future directions involve advanced tokenization for data scarcity and explainable AI for interpretability.

Conclusions:

• Continued methodological innovation tailored to RNA's unique characteristics is essential.
• Expansion of high-quality structural datasets is critical for transformative performance.
• Integrating advanced AI techniques is necessary to overcome current limitations in RNA structure prediction.