Abstract
Tertiary structure-based RNA design plays a crucial role in synthetic biology and therapeutics. While existing methods have explored structure-to-sequence mappings, they focus solely on RNA structures and overlook the role of complex-level information, which is crucial for effective RNA design. To address this limitation, we propose the Complex-Aware tertiary structure-based RNA Design model, CARD, that integrates complex-level information to enhance tertiary structure-based RNA sequence design. Specifically, our method incorporates protein features extracted by a protein language model (e.g., ESM-2), enabling the design model to generate more accurate and complex-relevant sequences. Considering the biological complexity of protein-RNA interactions, we introduce distance-aware filtering for local features from protein representation. Furthermore, we design a high-affinity design framework that combines our CARD with an affinity evaluation model. Extensive experiments demonstrate the effectiveness of our method with an improvement of 5.6% compared with the base model without our complex-aware feature integration.
Introduction
In synthetic biology, "inverse folding"—designing a nucleotide sequence that folds into a specific 3D shape—is a fundamental challenge. Most current deep learning models treat RNA as an isolated molecule, mapping a target structure directly to a sequence. However, in vivo, RNA function is often inextricably linked to its interactions with proteins and other ligands. Designing an RNA sequence based solely on its "naked" structure ignores the chemical and physical constraints imposed by its protein partners. To design effective therapeutics or biosensors, we must move toward complex-aware design, where the model incorporates the protein context during the RNA sequence generation process.
Methods
The CARD framework is built on three innovative pillars designed to capture the nuances of protein-RNA interactions:
- Complex-Aware Feature Fusion: The model utilizes a Geometric Vector Perceptron (GVP) to encode RNA 3D coordinates and a pre-trained Protein Language Model (ESM-2) to extract features from the partner protein. These features are fused using a novel CAFormer (Complex-Aware Transformer) architecture.
- Distance-Aware Filtering: To manage computational complexity and focus on relevant biological signals, CARD implements a filtering block. This identifies the $K$ amino acids closest to the RNA nucleotides, focusing the model’s attention specifically on the binding interface rather than the entire protein structure.
- High-Affinity Design Framework: Beyond structural matching, CARD employs an iterative screening pipeline. Candidate sequences are ranked by a dedicated affinity prediction model, and the top candidates are validated using structural tools like AlphaFold3.
Results
The model was evaluated using the PRI30K dataset (over 21,000 protein-RNA pairs) and the PRA201 blind test set. Key findings include:
- Superior Accuracy: CARD achieved an overall sequence recovery rate of 59.56% on the PRI30K test set, representing a 5.6% improvement over the base model (RhoDesign) which lacks complex-aware features.
- Robustness Across Lengths: CARD consistently outperformed sequence-based (LSTM/RNN) and other structure-based (GraphTrans, StructGNN) models across various RNA lengths.
- Enhanced Affinity: In a case study on the 2LBS complex, CARD designed sequences with higher predicted binding affinity than the native sequence while maintaining structural integrity (RMSD < 2.0Å).
Conclusion
CARD demonstrates that protein context is a foundational determinant in RNA design. By focusing on local interaction sites through distance-aware filtering, the model mimics biological reality more closely than global embedding approaches. Future work will focus on wet-lab validation of these designed sequences and extending the framework to multi-ligand complexes.
Summary updated