BindRNAgen: Protein-binding RNA Sequence Generation Using Latent Diffusion Models

RNA-binding proteins (RBPs) are pivotal regulators of gene expression, and their dysregulation is implicated in a wide range of human diseases. Designing synthetic RNA molecules to modulate RBP activity represents a promising therapeutic strategy, being constrained by the inefficiency of experimental screening and the limited generalization of existing computational models that require RBP-specific interaction data. Here, we present BindRNAgen, a hybrid RNA design framework that couples a variational autoencoder (VAE) with a conditional latent diffusion model (LDM), enabling the generation of binding RNA sequences given the RBP sequence as the input. Using RBP-binding RNAs derived from 168 eCLIP-seq datasets of diverse RBPs, we first pretrain the VAE on RBP-binding RNA sequences to construct a continuous latent representation for RBP binding sequence specificity. The LDM subsequently generates novel RBP-binding RNA sequences within the latent space, conditioned on protein-specific embeddings from the protein language model. For RBPs in the training set, BindRNAgen produces computationally predicted RBP-binding RNA sequences that are biophysically comparable to natural RBP-binding RNAs, outperforming existing benchmarks. Although the generalization for RBP targets outside the training set may be influenced by underlying homology to the training RBPs, BindRNAgen generates RNA sequences with high computationally predicted binding scores, as validated by in silico docking and molecular dynamics (MD) simulations.