AdarEdit

Adenosine-to-inosine (A-to-I) RNA editing, catalyzed by ADAR enzymes, is the most common form of RNA editing in animals. Because inosine is read as guanosine by the cellular machinery, editing can recode proteins, alter splicing, and reshape regulation — and its dysregulation is implicated in disease and immunity. Predicting which adenosines are edited, and how strongly, is hard because editing depends not only on local sequence but on RNA secondary structure: ADAR acts on double-stranded regions, often formed by inverted Alu repeats that fold back on themselves.

AdarEdit, presented by Rosenwasser, Levitt, Levanon, and Oren (corresponding author Gal Oren, Stanford University; with Erez Levanon, a leading RNA-editing researcher, at Bar-Ilan University) in a February 2026 bioRxiv preprint, is a structure-aware graph learning framework for predicting RNA editability. It represents RNA segments as graphs in which nucleotides are nodes connected by both sequential edges and base-pairing edges, so the model sees sequence and secondary structure jointly rather than as flat sequence alone.

Trained on inverted Alu duplexes derived from 8,603 GTEx RNA-seq samples across 47 human tissues, AdarEdit achieves strong discrimination (AUROC/AUPRC ≈ 0.96, F1 ≈ 0.90 on combined-tissue data) and — notably — transfers zero-shot to evolutionarily distant organisms, suggesting it has captured conserved principles of ADAR substrate recognition. Pretrained weights are released under GPL-3.0.

#Key Features

• Graph representation of RNA: Nucleotides are nodes; edges encode sequential adjacency and base pairing (canonical, wobble), so sequence and secondary structure are modeled together.
• Two architecture variants: A baseline graph attention network with 8-dimensional node features and a "bio-aware" version with 22-dimensional features, typed edges, and a parallel 1D CNN.
• Cross-tissue training at scale: Trained on inverted Alu duplexes from 8,603 GTEx RNA-seq samples spanning 47 human tissues.
• Cross-species zero-shot transfer: Generalizes to evolutionarily distant organisms (e.g., sea urchin, acorn worm, octopus), pointing to conserved ADAR recognition rules.
• Released pretrained weights: Checkpoints for all 31 train→validation settings, plus per-sample predictions and ROC/PR data, are publicly available (GPL-3.0).

#Technical Details

AdarEdit is a graph-attention model in which each RNA segment becomes a graph: node features encode base identity, pairing status, relative position, and (in the bio-aware variant) biochemical properties, while edges represent sequential connections and base-pairing types. The bio-aware variant augments the 8-feature baseline GAT to 22 node features, adds typed edges, and runs a parallel 1D CNN branch. Training used inverted Alu pair coordinates and per-tissue editing levels from GTEx (v7), with F1-optimized thresholding for checkpoint selection. On combined-tissue data the bio-aware model reaches AUROC/AUPRC ≈ 0.96 and F1 ≈ 0.90, and its attention maps highlight influential structural motifs, giving a degree of interpretability. The repository provides Alu coordinates (BED), per-tissue editing levels, cross-species datasets, training code, and pretrained checkpoints under GPL-3.0.

#Applications

AdarEdit serves RNA biologists and computational researchers studying RNA editing, ADAR-dependent regulation, and editing-based therapeutics. It can prioritize candidate editing sites in new tissues or organisms, help interpret how sequence and structure drive editability, and — via its attention mechanism — surface the structural motifs that govern ADAR substrate choice. Its cross-species generalization makes it useful for annotating editing in non-model organisms where experimental editing maps are sparse, and the released weights let groups apply or fine-tune the model without retraining from scratch.

#Impact

By framing RNA editability prediction as structure-aware graph learning and demonstrating zero-shot transfer to distant species, AdarEdit provides evidence that ADAR recognition follows conserved, learnable structural principles. The combination of strong benchmark performance, interpretable attention over structural motifs, and openly released checkpoints and datasets makes it a practical tool and a reproducible reference. As a February 2026 preprint, its claims await peer review and independent validation, but the open release lowers the barrier for the RNA-editing community to test and build on it.