Efficient design of rna sequences with desired properties, structure, and motifs using a grammar variational autoencoder (2507.15912v1)

Abstract: Designing structurally stable RNA sequences with specific motifs and other desirable properties is an important challenge in bioinformatics. The potential design space increases exponentially with the length of the RNA to be engineered, which makes this a difficult combinatorial optimization problem. In this paper, we propose an RNA grammar variational autoencoder (RGVAE) that can efficiently generate novel RNA sequences with specific target properties. The proposed RGVAE builds on the recently proposed grammar VAE, where we incorporate the stochastic context-free grammar (SCFG) to design strutural RNAs with desired motifs and characteristics. Using the SCFG can ensure that the generated RNA sequence can form a thermodynamically stable secondary structure. Given a RNA sequence, the SCFT is used to find the parse tree, which is represented in a continuous low-dimensional latent space by the RGVAE encoder. We can optimize the RNA in the latent space, where the latent representation can be decoded by the RGVAE decoder to reconstruct the RNA sequence. Based on a number of practical uses cases, we demonstrate that RGVAE can be used to efficiently design structurally stable RNAs with specific target properties, which significantly outperform other alternatives such as randomized design and regular VAEs that do not utilize the SCFG. Code availability: the source code of RGVAE and the data used in this study are provided in https://github.com/nzarnaghinaghsh/RGVAE/tree/main, DOI 10.5281/zenodo.15569206.