Generative machine learning of ADAR substrates for precise and efficient RNA editing

Abstract

Adenosine Deaminase Acting on RNA (ADAR) converts adenosine to inosine within certain double-stranded RNA structures. However, ADARs promiscuous editing and poorly understood specificity hinder therapeutic applications. We present an integrated approach combining high-throughput screening (HTS) with generative deep learning to rapidly engineer efficient and specific guide RNAs (gRNAs) to direct ADARs activity to any target. Our HTS quantified ADAR-mediated editing across millions of unique gRNA sequences and structures, identifying key determinants of editing outcomes. We leveraged these data to develop DeepREAD (Deep learning for RNA Editing by ADAR Design), a diffusion-based model that elucidates complex design rules to generate novel gRNAs outperforming existing design heuristics. DeepREADs gRNAs achieve highly efficient and specific editing, including challenging multi-site edits. We demonstrate DeepREADs therapeutic potential by designing gRNAs targeting the MECP2R168X mutation associated with Rett syndrome, achieving both allelic specificity and species cross-reactivity. This approach significantly accelerates the development of ADAR-based RNA therapeutics for diverse genetic diseases.