Secondary-Structure-Informed RNA Inverse Design via Relational Graph Neural Networks.

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• RNA inverse design is vital for RNA therapeutics.
• Current machine learning models predict RNA sequences from structure but struggle with conformational switching.
• Modeling multiple RNA structures remains a challenge.

Purpose of the Study:

• To develop an improved RNA inverse design method.
• To explicitly incorporate alternative RNA structures into sequence prediction.
• To enhance the accuracy of designing RNA sequences for therapeutic applications.

Main Methods:

• A relational geometric graph neural network was proposed.
• RNA structures were converted into geometric graphs.
• Edge types distinguished primary, secondary, and spatial nucleotide information.

Main Results:

• The proposed model achieved higher native sequence recovery rates than existing methods (72% vs. 66% and 60% vs. 57%).
• Secondary-structure edge types significantly impacted sequence recovery.
• The model demonstrates improved performance in RNA sequence prediction.

Conclusions:

• Explicitly incorporating alternative RNA structures improves inverse design.
• More complex and specific RNA characterization is needed for successful inverse design.
• This approach advances the design of RNA regulators like riboswitches.