GALS-Fold: Geometry-Aware Long-Short RNA Inverse Folding with Linear Scaling
TL;DR
GALS-Fold is proposed, a Geometry-Aware Long-Short framework that reconciles local geometric fidelity with efficient global context, and achieves state-of-the-art performance on long sequences, reaching an average recovery rate of 52.2% on long sequences.
Abstract
RNA inverse folding aims to design nucleotide sequences that adopt a prescribed 3D backbone, enabling programmable non-coding RNAs (ncRNAs) for therapeutics and synthetic biology. However, scaling 3D structure-guided design to long RNAs remains challenging due to a key trade-off we term the ''Long RNA Dilemma'': graph neural networks (GNNs) are efficient and capture local motifs but struggle to propagate long-range tertiary constraints, while Transformer-style global attention can model distant couplings but incurs prohibitive O(N2) time and memory and may dilute local geometric signals. We propose GALS-Fold, a Geometry-Aware Long-Short framework that reconciles local geometric fidelity with efficient global context. GALS-Fold combines an SE(3)-equivariant GNN based short-range encoder with a linear-time long-range module based on anchor attention. To ensure the long branch focuses on non-local information, we introduce a local exclusion regularization that discourages adjacent nucleotides from sharing the same anchors. We finally fuse long and short features with a length-aware gating strategy. Extensive evaluations on a rigorous 5-fold structural split demonstrate that GALS-Fold achieves state-of-the-art performance, reaching an average recovery rate of 52.2% (SOTA baseline 48.0%). On long sequences (>200 nt), GALS-Fold attains 62.4% recovery, outperforming strong baselines such as gRNAde (57.7%) and Graphormer (45.6%). Our code is available at: https://github.com/wenxy59/GALS-Fold.