Deep Learning for RNA Secondary Structure Determination: Gauging Generalizability and Broadening the Scope of...

Area of Science:

• RNA biology and bioengineering
• Computational biology
• Bioinformatics

Background:

• RNA structure is fundamental to its diverse biological functions and stability.
• Computational secondary structure prediction is a rapid, low-cost method for analyzing RNA.
• Deep learning (DL) has emerged as a powerful tool, mirroring its success in protein structure prediction.

Purpose of the Study:

• To assess the generalizability of RNA structure prediction methods, particularly DL approaches.
• To highlight the challenges posed by the generalization gap in RNA structure prediction.
• To explore advancements in DL for predicting RNA structure probing data.

Main Methods:

• Curated a new benchmark dataset of structured RNAs from the Protein Data Bank.
• Evaluated method generalizability on this dataset.
• Analyzed DL methods for predicting structure probing data using a dedicated dataset.

Main Results:

• The generalization gap currently hinders accurate prediction of novel RNA structures.
• Specific challenges exist for deep learning methods predicting structure probing data.
• Deep learning advances enable optimization and integration with traditional structure prediction methods.

Conclusions:

• Addressing the generalization gap is critical for advancing RNA structure prediction.
• Future work should focus on improving DL model generalizability and exploring new datasets.
• Deep learning offers new avenues for optimizing and integrating traditional RNA structure prediction techniques.