RNARL: reinforcement learning-driven unified generative framework for multi-objective RNA codon design

Area of Science:

• Computational biology
• Bioinformatics
• Synthetic biology

Background:

• Current RNA codon design methods struggle with long sequences and lack generalizability.
• Existing approaches often separate sequence generation from optimization, limiting efficiency.

Purpose of the Study:

• To introduce RNARL, a reinforcement learning framework unifying RNA sequence generation and multi-objective optimization.
• To develop a generalizable and efficient method for designing high-performance RNA sequences.

Main Methods:

• Developed a reinforcement learning-driven framework (RNARL) integrating sequence generation and multi-objective optimization.
• Applied RNARL to optimize RNA sequences, handling lengths up to 3,900 nucleotides.

Main Results:

• RNARL directly learns to generate high-performance RNA sequences.
• Demonstrated superior performance and universality across six species and five RNA types.
• Successfully optimized long RNA sequences efficiently.

Conclusions:

• RNARL establishes an effective and generalizable framework for RNA codon design.
• The framework overcomes limitations of previous decoupled methods.
• A user-friendly web platform is available for RNA therapeutic design applications.