# R3J-AGNN: GNN-Based Prediction of Inter-Branch Angles in RNA Three-Way Junctions from Secondary Structure

## TL;DR

This paper introduces R3J-AGNN, a machine learning model that predicts the 3D geometry of RNA three-way junctions from their secondary structure, improving RNA structure modeling.

## Contribution

R3J-AGNN is a novel dual-resolution graph neural network that predicts inter-branch angles in RNA three-way junctions from secondary structure data.

## Key findings

- R3J-AGNN integrates nucleotide-level and global topology features to predict inter-branch angles.
- The model achieves robust performance across diverse RNA three-way junction architectures.
- It provides actionable geometric constraints for RNA tertiary structure modeling and refinement.

## Abstract

Despite the essential role of three-dimensional RNA structures in cellular functions, accurately modeling the spatial organization of multi-branch junctions—particularly three-way junctions (3WJs)—remains a substantial challenge, even when the underlying secondary structure is known. Here, we introduce R3J-AGNN, a dual-resolution hierarchical graph neural network that predicts inter-branch angles of RNA 3WJs directly from secondary structure information. The model integrates fine-grained nucleotide-level interactions with a coarse-grained representation of global junction topology, enabling the inference of three-dimensional geometry from sequence-derived features alone. Evaluations on independent test sets demonstrate that R3J-AGNN achieves robust and consistent predictive performance across diverse junction architectures. By providing accurate geometric constraints for 3WJs, R3J-AGNN offers a practical tool for improving RNA tertiary structure modeling and refinement.

Despite advances in computational RNA 3D structure prediction, accurately modeling multi-branch RNA motifs remains challenging, even with known secondary structures, because their global topology is highly sensitive to the relative orientations of the helical stems at the junction. Here, we propose R3J-AGNN, a dual-resolution hierarchical graph neural network that predicts inter-branch angles of RNA three-way junctions (3WJs) directly from secondary structure information. The model integrates fine-grained nucleotide-level interactions with a coarse-grained representation of the global topology of 3WJ-containing RNA structures, enabling accurate inference of inter-branch geometry and subsequent reconstruction of their 3D scaffolds. Evaluations on an independent test set demonstrate that R3J-AGNN achieves robust and consistent predictive performance. Thus, by inferring inter-branch angles of RNA three-way junctions from secondary structure information, R3J-AGNN provides actionable geometric constraints for RNA tertiary structure modeling and refinement.