AdarEdit: A Graph Foundation Model for Interpretable A-to-I RNA Editing Prediction

Zohar Rosenwasser, Michael Levitt, Erez Levanon, Gal Oren

Keywords: A-to-I RNA editing, ADAR, graph neural networks, Graph Attention Networks, RNA foundation model, interpretability, dsRNA, cross-tissue generalization, cross-species transfer, guide RNA design

TL;DR: A graph foundation model (AdarEdit) unifies RNA sequence and secondary structure to predict A-to-I editing with high accuracy and built-in mechanistic interpretability.

Abstract:

Adenosine-to-inosine (A-to-I) RNA editing, catalyzed by ADAR enzymes, is a prevalent post-transcriptional modification with roles in transcript stability, splicing, and protein recoding. Accurate prediction of editing sites remains difficult due to the intricate interplay between local sequence context and RNA secondary structure. Existing approaches either rely on brittle, handcrafted features or adapt generic foundation models trained on broad RNA datasets, which fail to capture the specific biochemical requirement for double-stranded RNA and often lack interpretability.

We introduce AdarEdit, a domain-specialized graph foundation model for A-to-I editing site prediction. RNA segments are represented as graphs with nucleotides as nodes and both sequential and base-pairing edges, enabling the model to learn biologically aligned features such as stem–loop motifs. A Graph Attention Network architecture yields mechanistic interpretability by highlighting influential structural and sequence neighbors. Across 25 cross-tissue evaluations, AdarEdit consistently outperforms prior methods (F1 > ~0.85) and generalizes to evolutionarily distant species, demonstrating that biology-aware foundation models can deliver superior accuracy, scalability, and insight for complex RNA modification tasks.

The sources of this work are available at our repository: https://github.com/Scientific-Computing-Lab/AdarEdit

Submission Number: 6