Deep generative model of RNAs based on variational autoencoder with context-free grammar.

Area of Science:

• Computational Biology
• Bioengineering
• Molecular Biology

Background:

• RNA molecules are vital for cellular processes and bioengineering.
• Traditional RNA design methods often assume a conserved secondary structure.
• RNA sequences can exhibit diverse secondary structures, especially with mutations.

Purpose of the Study:

• To develop a novel deep generative model for RNA sequence design.
• To explicitly incorporate individual RNA secondary structures into the generation process.
• To explore structure-activity relationships in RNA sequences.

Main Methods:

• Integration of context-free grammar (CFG) with a variational autoencoder (VAE).
• Representation of RNA sequences and structures as CFG parse trees, converted to binary matrices for VAE training.
• Dynamic programming for optimal parse tree reconstruction to ensure structure-aware generation.

Main Results:

• The model successfully generates high-quality RNA sequences, validated against the Rfam database.
• A strong correlation was observed between the VAE's latent space and the self-cleaving activity of RNA aptazyme mutants.
• Demonstrated the significance of integrating RNA-specific structural information into generative models.

Conclusions:

• The developed model enables structure-aware RNA sequence generation.
• The findings highlight the importance of secondary structure in determining RNA function.
• This approach advances RNA design for both research and bioengineering applications.