Structure-aware Graph Learning Predicts RNA Editability Across Tissues and Species

Abstract

Programmable A-to-I RNA editing using endogenous ADAR enzymes is emerging as a therapeutic strategy, but editability remains difficult to predict because ADAR recognition depends on double-stranded RNA geometry and stability rather than sequence alone. We present AdarEdit, a structure-explicit graph-attention framework that represents each dsRNA substrate as a nucleotide graph with backbone and base-pair edges. The framework includes a baseline model and a bio-aware model, with the latter augmenting this representation with typed interactions and a motif-sensitive sequence branch. We trained and evaluated both models on high-confidence inverted Alu duplexes (n = 884) with secondary structures predicted by RNAfold and editing levels measured across 8,603 GTEx RNA-seq samples spanning 47 tissues. Across five tissue contexts, the baseline and bio-aware models achieved strong held-out performance (test F1 = 0.814-0.869, AUROC = 0.869-0.933) and outperformed a matched structure-string baseline on the Liver split. The same graph representation retained predictive ability in evolutionarily distant non-Alu species (sea urchin, acorn worm, and octopus), suggesting conserved principles of ADAR substrate recognition. Finally, attention profiles and in silico mutagenesis recapitulated known biochemical constraints, including suppression by an upstream guanosine, and revealed longer-range asymmetric structural influences on editing. Because Alu duplexes are edited predominantly by ADAR1, AdarEdit is geared primarily to the ADAR1 regime. ADAR1 is particularly relevant to therapeutic editing given its broad tissue expression. The sources of this work are available at our repository: https://github.com/Scientific-Computing-Lab/AdarEdit