Abstract
RNA-protein complexes underlie numerous cellular processes, including basic translation and gene regulation. The high-resolution structure determination of the RNA-protein complexes is essential for elucidating their functions. Therefore, computational methods capable of identifying the native-like RNA-protein structures are needed. To address this challenge, we thus develop DRPScore, a deep-learning-based approach for identifying native-like RNA-protein structures. DRPScore is tested on representative sets of RNA-protein complexes with various degrees of binding-induced conformation change ranging from fully rigid docking (bound-bound) to fully flexible docking (unbound-unbound). Out of the top 20 predictions, DRPScore selects native-like structures with a success rate of 91.67% on the testing set of bound RNA-protein complexes and 56.14% on the unbound complexes. DRPScore consistently outperforms existing methods with a roughly 10.53-15.79% improvement, even for the most difficult unbound cases. Furthermore, DRPScore significantly improves the accuracy of the native interface interaction predictions. DRPScore should be broadly useful for modeling and designing RNA-protein complexes.
Introduction: The Scoring Problem in RNA-Protein Docking
RNA-protein complexes are fundamental to cellular machinery, driving protein synthesis and gene regulation. Understanding their function often requires high-resolution 3D structures. However, experimental determination via X-ray crystallography or NMR is frequently hindered by the inherent flexibility of RNA. Computational "docking" serves as a vital alternative, generating thousands of potential structural "decoys." The primary challenge lies in the "scoring problem": identifying the single native-like structure from this vast pool of candidates.
Traditional scoring functions rely on physics-based statistical potentials, which often struggle with "unbound" docking—scenarios where the RNA or protein undergoes significant conformational changes upon binding. To overcome these limitations, researchers have developed DRPScore, a deep-learning-based scoring function designed to capture the complex spatial and sequential nuances of RNA-protein interfaces.
Methodology: The 4D-CNN Architecture
DRPScore distinguishes itself from previous 3D convolutional neural networks (3DCNNs) by introducing a fourth dimension: sequence order. While 3DCNNs treat atoms as static points in a volume, DRPScore's 4D architecture stacks convolutional layers along the sequence dimension. This allows the model to integrate 3D spatial geometry with 1D sequential context, enabling it to recognize how a residue interacts not just with its immediate spatial neighbors, but with residues further along the molecular chain.
Feature Extraction and Input
The model focuses on the interface, defined by a 6Å cutoff between the RNA and protein. It constructs a 32Å grid around each interface residue, incorporating 85 atom types for RNA and 225 for proteins. By including atomic properties like mass and charge, DRPScore captures high-fidelity chemical and physical interactions at the binding site.
Performance Evaluation
DRPScore was rigorously tested against industry-standard benchmarks (ITScore-PR, DARS-RNP, and 3dRPC) using two primary datasets:
- Bound-Bound Set: 36 complexes where structures are already in their bound shapes (rigid docking).
- Unbound-Unbound Set: 57 complexes where structures are in their free, unbound shapes. This represents the "gold standard" of difficulty due to binding-induced conformational shifts.
In the bound-bound set, DRPScore achieved a remarkable 91.67% success rate within the top 20 predictions. More importantly, in the difficult unbound-unbound set, it maintained a 56.14% success rate, consistently outperforming existing methods by a margin of 10.53% to 15.79%.
Comparative Analysis and Structural Accuracy
The superiority of DRPScore is evident not just in aggregate success rates but in the precision of the identified models. By analyzing the Root Mean Square Deviation (RMSD) from the native structure, DRPScore consistently identifies models with lower error than its competitors.
As shown in the case study of the Bacillus subtilis YxiN protein (Figure 5c), DRPScore identified a model with an RMSD of only 1.98 Å. In contrast, benchmark methods produced models with errors exceeding 8 Å, failing to capture the true binding mode. This suggests that DRPScore has learned a more accurate representation of the physical interface, including critical hydrogen bond networks.
Conclusion and Future Directions
DRPScore represents a significant advancement in the computational modeling of RNA-protein complexes. By combining spatial and sequential data through a 4D-CNN, it provides a robust solution to the scoring problem, particularly for flexible unbound docking.
Despite these gains, challenges remain. A 56% success rate in unbound cases indicates that nearly half of the most difficult complexes are still not correctly identified. Furthermore, a scoring function is limited by the quality of the decoys provided; if a docking simulation fails to generate any native-like structures, no scorer can succeed. Future work may involve integrating molecular dynamics (MD) simulations to improve the sampling of flexible RNA movements, providing DRPScore with higher-quality candidates for evaluation.
Summary updated