Monte Carlo Inverse RNA Folding.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• The inverse RNA folding problem is crucial for designing functional RNA molecules.
• Existing methods often struggle with the complexity of predicting RNA structures from sequences.
• Monte Carlo search algorithms offer a framework for optimizing sequence design.

Purpose of the Study:

• To develop an AI-driven approach for solving the inverse RNA folding problem.
• To leverage a transformer neural network for generating effective priors in RNA sequence design.
• To evaluate the performance of the Generalized Nested Rollout Policy Adaptation (GNRPA) algorithm with a transformer-generated prior.

Main Methods:

• Training a transformer neural network on the Rfam database for inverse RNA folding.
• Utilizing the trained transformer to generate priors for Eterna100 puzzles.
• Applying the Generalized Nested Rollout Policy Adaptation (GNRPA) algorithm with the transformer prior to solve RNA folding instances.
• Comparing results against handcrafted heuristics.

Main Results:

• The transformer neural network successfully generated priors for RNA folding problems.
• GNRPA, guided by the transformer prior, effectively solved instances from the Eterna100 dataset.
• The transformer-generated prior demonstrated superior performance compared to handcrafted heuristics.

Conclusions:

• AI-driven priors significantly enhance the performance of search algorithms like GNRPA for inverse RNA folding.
• Transformer neural networks offer a powerful tool for advancing RNA sequence design.
• This approach shows promise for designing novel RNA molecules with desired structures and functions.