Helix: a structure-aware deep learning model for accurate prediction of A-to-I RNA editing by endogenous ADARs

Abstract

Adenosine deaminase acting on RNA (ADAR) converts adenosine to inosine within double-stranded RNA (dsRNA) and can be co-opted for therapeutic RNA editing by introducing dsRNA substrates in trans using programmable guide RNAs (gRNAs). However, ADARs natural promiscuity necessitates sophisticated gRNA designs to achieve efficient and specific editing. Here we present Helix, a predictive model that achieves highly accurate, zero-shot per-adenosine editing predictions for any target sequence. Helixs performance arises from two architectural choices: a transformer framework that scales effectively with increasing and imbalanced training data; and a structure-aware attention mechanism that incorporates predicted RNA secondary structure, a key determinant of ADAR activity. Helixs predictive accuracy enables seamless integration with DeepREAD, our previously reported generative model, in a noisy-student distillation framework termed DeepHelix. This workflow supports both zero-shot gRNA design and the generation of complex, constraint-based designs. We demonstrate DeepHelixs utility by designing gRNAs that efficiently edit a therapeutically relevant adenosine and by leveraging its flexible design space to engineer species cross-reactive gRNAs to accelerate pre-clinical development.