Abstract
HyperRNA is a novel generative framework designed to address the 3D RNA inverse folding problem. By representing RNA molecular structures as hypergraphs, the model captures higher-order, multi-atomic interactions that traditional graph neural networks often overlook. HyperRNA utilizes an encoder-decoder architecture featuring attention embeddings and a hypergraph neural network to achieve state-of-the-art performance in sequence recovery and structural accuracy. Evaluations on PDBBind and RNAsolo datasets demonstrate its superior ability to design diverse and novel RNA sequences that fold into target 3D structures.
Introduction
RNA inverse folding—the task of designing a nucleotide sequence that folds into a predefined 3D structure—is a critical challenge in bioinformatics. While protein design has seen significant progress, RNA remains difficult due to its high flexibility and reliance on complex interactions like base pairing and stacking. Traditional Geometric Deep Learning approaches typically use Graph Neural Networks (GNNs) to model these interactions as pairwise relationships. However, RNA biology involves intricate multi-node dependencies. HyperRNA addresses this by employing hypergraphs, where hyperedges can connect multiple nodes, effectively modeling the group-wise forces that stabilize RNA structures.
Methodology
The HyperRNA architecture consists of three primary stages: preprocessing, encoding, and decoding.
- Preprocessing: To manage the complexity of RNA structures, the model adopts a 3-bead coarse-grained representation, tracking three principal atoms per nucleotide (Phosphorus, C4', and the base nitrogen). A k-Nearest Neighbors (kNN) approach is then used to construct the initial graph.
- Encoding: This stage features an Attention Embedding module that uses self-attention to capture dependencies between scalar and vector features. The core innovation is the Hypergraph Neural Network (HGNN), which utilizes hyperedges to represent complex biomolecular neighborhoods more effectively than standard edges.
- Decoding: An autoregressive decoder, equipped with Geometric Vector Perceptrons (GVP), generates the RNA sequence from the 5' to the 3' end. This ensures that each predicted nucleotide satisfies the structural constraints imposed by the preceding sequence.
Experimental Setup
The model was benchmarked against two primary datasets: PDBBind (1,192 RNA-protein complexes) and RNAsolo (4,223 diverse RNA structures). To prevent data leakage and test generalization, the researchers employed a sequence-similarity split, clustering sequences with >80% identity. This ensures the model is not simply memorizing training examples.
Results
HyperRNA consistently outperformed baseline models like gRNAde across multiple metrics. On the PDBBind sequence-similar split, it achieved a recovery rate of 0.31 and an lDDT score of 0.56, indicating high structural fidelity. On the RNAsolo dataset, the model demonstrated superior diversity (0.47) and novelty, suggesting it can generate a wide range of valid structures rather than repeating a few "safe" patterns.
Ablation studies further confirmed that the hypergraph component is "plug-and-play." Integrating the hypergraph representation into the existing gRNAde model led to immediate performance gains, highlighting the general superiority of hypergraphs for molecular geometry.
Conclusion
HyperRNA demonstrates that hypergraph representations are inherently better suited for capturing the complex geometry of RNA. By modeling multi-atomic interactions, the framework achieves significant improvements in inverse folding accuracy. While limited by the availability of high-quality 3D RNA data, future work will focus on optimizing hypergraph construction and extending the model to tasks such as RNA-ligand binding and functional site identification.
Summary updated