Solving the RNA design problem with reinforcement learning.

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• RNA secondary structure prediction is crucial for understanding RNA function.
• Designing RNA sequences to fold into desired structures remains a significant challenge in computational biology.
• Existing algorithms often struggle with complex or arbitrary target structures.

Purpose of the Study:

• To develop a novel reinforcement learning agent for *in silico* RNA design.
• To create a versatile model capable of handling arbitrary RNA target structures of any length.
• To evaluate the agent's performance against established benchmarks and analyze its learned strategies.

Main Methods:

• Utilized reinforcement learning to train a computational agent.
• Employed a novel graph convolutional neural network architecture.
• Trained the agent on randomly generated target RNA structures.
• Tested the agent on the Eterna100 benchmark dataset.

Main Results:

• The reinforcement learning agent significantly outperformed all previous RNA design algorithms on the Eterna100 benchmark.
• The agent successfully learned and applied advanced design strategies observed in human players of the Eterna game.
• The model demonstrated the ability to solve highly complex RNA secondary structures.

Conclusions:

• The developed agent represents a significant advancement in computational RNA design.
• The graph convolutional architecture enables generalization to diverse and complex target structures.
• Further improvements in training protocols may enhance the agent's ability to learn a wider range of design strategies.