Modular Deep Learning for Direct RNA Sequence Design via Self-Contained RNA Units

Area of Science:

• Synthetic biology
• Computational biology
• Structural biology

Background:

• Deep learning for RNA sequence design is limited by scarce high-resolution 3D structure data.
• Existing methods like NA-MPNN and RiboDiffusion are computationally intensive due to limited training data.

Purpose of the Study:

• To address the data scarcity and granularity issues in RNA sequence design.
• To develop a scalable and physically grounded framework for generating accurate RNA sequences.

Main Methods:

• Introduction of SCRU-DB, a database of over 61,000 Self-contained RNA Units (SCRUs) derived from tertiary contact clustering.
• Development of SCRU-Seq (a direct GNN) and SCRU-Diff (an iterative diffusion model) leveraging the SCRU-DB.
• Validation of structural fidelity using 3D backbone superposition and C4' RMSD.

Main Results:

• SCRU-DB contains over 8,200 unique structural clusters of SCRUs.
• SCRU-Seq achieved 63.7% native sequence recovery (NSR) on the set112 benchmark.
• SCRU-Diff achieved a superior Best NSR of 79.2% with high structural fidelity (1.5Å C4' RMSD).

Conclusions:

• The SCRU-DB framework provides a scalable solution for RNA sequence design by utilizing modular RNA units.
• SCRU-Seq and SCRU-Diff demonstrate improved accuracy and efficiency in generating structurally sound RNA sequences.
• The modular approach ensures the generation of diverse and physically realistic RNA sequences.