Diverse database and machine learning model to narrow the generalization gap in RNA structure prediction.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Macromolecular structures, particularly proteins and nucleic acids, are crucial for understanding biological functions.
• While protein structure prediction is advanced, RNA structure prediction faces challenges due to limited structural data.
• The Protein Data Bank holds over 180,000 protein structures, aiding deep learning advancements.

Purpose of the Study:

• To address the limitations in RNA secondary structure prediction.
• To present novel secondary structure models for microRNAs and messenger RNA regions.
• To develop and validate a new deep learning architecture for RNA structure prediction.

Main Methods:

• Generated secondary structure models for 1098 primary microRNAs and 1456 human messenger RNA regions using chemical probing.
• Developed eFold, a deep learning model inspired by AlphaFold's Evoformer and traditional architectures.
• Trained eFold on a newly generated database and over 300,000 existing secondary structures.

Main Results:

• eFold demonstrated improved prediction performance on challenging, diverse RNA structure test sets.
• The newly generated dataset and eFold architecture contributed to enhanced prediction accuracy.
• Results indicate that incorporating structural diversity and complexity is key for generalization.

Conclusions:

• Expanding RNA structure databases alone is insufficient for accurate prediction across different families.
• The eFold model and diverse dataset advance the field of RNA secondary structure prediction.
• Future efforts should focus on capturing the complexity and diversity of RNA structures for better models.