Physics-supervised learning of the sequence--chemistry relationship for therapeutic RNA editing

RNA editing is a promising therapeutic approach, offering safe and durable treatment across a broad range of diseases. A viable editing guide must be optimized jointly over its nucleotide sequence and its chemical modification pattern, yet no current machine-learning design method addresses guide chemistry for RNA editing, leaving chemistry to costly empirical screening. Here, we present a model that uses joint representations of target and guide sequence and positional guide chemistry for RNA editing. We further show how features derived from all-atom molecular dynamics simulations of the guide--target duplex help improve the model's ability to generalize to unseen targets. We evaluate on 11 unseen target editing sites, ranking chemical designs within each site. Adding molecular dynamics supervision nearly doubles the mean within-site Spearman correlation, from $0.27$ to $0.53$ ($0.20$-$0.90$ across sites). To our knowledge this is the first evidence that molecular dynamics paired with deep learning informs experimental guide ranking, enabling candidate prioritization before wet-lab screening.