RiboDiffusion: tertiary structure-based RNA inverse folding with generative diffusion models.

Area of Science:

• Computational Biology
• Synthetic Biology
• Biotechnology

Background:

• RNA's critical roles in biology drive applications in synthetic biology and therapeutics.
• The inverse RNA folding problem, designing functional sequences for specific structures, remains a challenge, especially from 3D conformations.
• Existing computational methods primarily focus on secondary structures, with direct 3D structure-based design facing data scarcity and conformational flexibility issues.

Purpose of the Study:

• To develop a novel generative model, RiboDiffusion, for RNA inverse folding using 3D structural information.
• To enable the design of RNA sequences that conform to specified 3D backbone structures.
• To explore the trade-off between sequence recovery and diversity in RNA design.

Main Methods:

• RiboDiffusion employs a generative diffusion model integrating a graph neural network for structure and a Transformer for sequence.
• The model iteratively refines random sequences into target sequences conditioned on 3D RNA backbone structures.
• Test sets were stratified by RNA clustering (sequence/structure similarity) to evaluate model performance.

Main Results:

• RiboDiffusion significantly outperforms baseline methods in sequence recovery, showing 11-16% relative improvement on similarity-based splits.
• The model demonstrates consistent performance across diverse RNA lengths and types.
• In silico folding validation confirmed that generated sequences can adopt the target 3D RNA backbones.

Conclusions:

• RiboDiffusion presents a powerful new approach for RNA inverse folding from 3D structures.
• The method effectively navigates the sequence space to find novel RNA candidates meeting structural constraints.
• This tool has significant potential for advancing RNA design in synthetic biology and therapeutic development.