Abstract
Geometric graph neural networks (GNNs) that respect E(3) symmetries have achieved strong performance on small molecule modeling, but they face scalability and expressiveness challenges when applied to large biomolecules such as RNA and proteins. These systems require models that can simultaneously capture fine-grained atomic interactions, long-range dependencies across spatially distant components, and biologically relevant hierarchical structure, such as atoms forming residues, which in turn form higher-order domains. Existing geometric GNNs, which typically operate exclusively in either Euclidean or Spherical Harmonics space, are limited in their ability to capture both the fine-scale atomic details and the long-range, symmetry-aware dependencies required for modeling the multi-scale structure of large biomolecules. We introduce DualEquiNet, a Dual-Space Hierarchical Equivariant Network that constructs complementary representations in both Euclidean and Spherical Harmonics spaces to capture local geometry and global symmetry-aware features. DualEquiNet employs bidirectional cross-space message passing and a novel Cross-Space Interaction Pooling mechanism to hierarchically aggregate atomic features into biologically meaningful units, such as residues, enabling efficient and expressive multi-scale modeling for large biomolecular systems. DualEquiNet achieves state-of-the-art performance on multiple existing benchmarks for RNA property prediction and protein modeling, and outperforms prior methods on two newly introduced 3D structural benchmarks demonstrating its broad effectiveness across a range of large biomolecule modeling tasks.
Introduction
Modeling large biomolecules like RNA and proteins presents a unique challenge for deep learning. These systems are massive and hierarchical, requiring models to capture fine-grained atomic interactions, long-range dependencies across folded chains, and the biological grouping of atoms into residues and domains. Existing Geometric Graph Neural Networks (GNNs) often struggle with this multi-scale complexity, either focusing on local Euclidean (EU) geometry at the expense of long-range context or using computationally heavy Spherical Harmonics (SH) that ignore hierarchical structure. DualEquiNet addresses these limitations by bridging the gap between local atomic precision and global structural context.
Methodology
DualEquiNet introduces a hierarchical approach that simultaneously leverages EU and SH spaces to model biomolecular geometry while maintaining E(3) equivariance.
Dual-Space Representation and Neighborhoods
The model maintains parallel feature sets in Euclidean (EU) and Spherical Harmonics (SH) spaces. EU space captures local geometric interactions based on physical distance cutoffs. In contrast, SH space encodes rotation-equivariant features that identify "neighbors" based on structural orientation similarity, even if they are physically distant. This allows the network to bridge long-range dependencies essential for understanding molecular folding.
DualEqui Layer and Bidirectional Message Passing
The core building block is the DualEqui Layer, which facilitates bidirectional message passing. Information flows both within each space (inner-space) and between them (cross-space). This bidirectional exchange ensures that local atomic details inform global structural patterns and vice versa. To maintain efficiency, the model utilizes an inner-product-based scalarization trick, avoiding the high computational cost typically associated with high-order tensor products.
Cross-Space Interaction Pooling (CSIP)
To model the biological hierarchy, DualEquiNet introduces CSIP. This mechanism aggregates atom-level features into residue-level representations. Unlike simple pooling, CSIP explicitly exchanges information between EU and SH representations during aggregation, allowing the model to "zoom out" to higher-order structures without losing geometric precision.
Experiments and Results
The researchers evaluated DualEquiNet on a diverse suite of benchmarks, including RNA property prediction (CovidVaccine, Ribonanza), protein modeling (SASA), and new 3D structural datasets.
- RNA Prediction: DualEquiNet reduced error (RMSE) by 25.7% on CovidVaccine and 33.5% on Ribonanza compared to the best previous methods.
- Structural Accuracy: The model achieved a 28.8% improvement in predicting the Solvent-Accessible Surface Area (SASA) of proteins and a 2.5% improvement in RNA torsion angle prediction.
- Expressivity and Efficiency: In synthetic tasks like the "N-chain" task, DualEquiNet achieved 100% accuracy where most standard GNNs failed. Despite its complexity, training times remain comparable to simpler "scalar-only" models due to efficient scalarization.
Conclusion
DualEquiNet represents a significant advancement in biomolecular modeling by integrating dual-space geometric reasoning with hierarchical pooling. Its ability to capture multi-scale dependencies while remaining computationally efficient makes it a powerful tool for structural biology. Future work aims to integrate 3D structure prediction directly into the pipeline, allowing the model to work from raw sequences while leveraging its dual-space geometric reasoning.
Summary updated