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

Abstract

RNA plays a crucial role in cellular functions, and designing functional RNA sequences is essential for both scientific exploration and bioengineering applications. Conventional RNA design approaches typically assume a shared secondary structure among designed sequences. However, even closely related RNAs can adopt different secondary structures, particularly when artificial mutations are introduced. We present a novel deep generative model that integrates context-free grammar (CFG) with a variational autoencoder (VAE) to generate RNA sequences while explicitly considering their individual secondary structures. In our method, RNA sequences and their structures are represented as parse trees based on CFG, which are then transformed into binary matrices for VAE training. The optimal parse tree is reconstructed using dynamic programming, ensuring structure-aware sequence generation. When evaluated on natural RNAs from the Rfam database, our model successfully generates high-quality RNA sequences. Furthermore, when applied to RNA aptazyme mutants with distinct secondary structures, our method reveals a strong correlation between the latent space representation of the VAE and self-cleaving activity. This underscores the importance of incorporating RNA-specific structural information in generative models. https://github.com/gterai/RNAgg (archived at Zenodo: https://doi.org/10.5281/zenodo.15354990).