CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction

Rong Han, Xiaohong Liu, Tong Pan, Jing Xu, Xiaoyu Wang, Wuyang Lan, Zhenyu Li, Zixuan Wang, Jiangning Song, Guangyu Wang, Ting ChenAAAI 2025 · DOI

TL;DR

This paper introduces CoPRA, a novel framework that synergistically combines protein and RNA language models with structural information to predict protein-RNA binding affinity. By developing a Co-Former architecture and a bi-scope pre-training strategy, CoPRA achieves state-of-the-art performance, offering a more comprehensive understanding of these crucial biological interactions.

What Problem Does This Paper Address?

Protein-RNA interactions are fundamental to numerous biological processes, from gene regulation to cellular function. Accurately predicting the strength of these interactions (binding affinity) is vital for understanding disease mechanisms and designing targeted therapies. However, existing computational methods often rely on either sequence or structure features in isolation, failing to capture the complex interplay that governs binding. While recent advances in pre-trained language models (PLMs and RLMs) have shown promise for individual tasks, their collaborative use for complex-level predictions, especially when integrating structural information, remains largely unexplored. This gap limits our ability to fully model and predict protein-RNA binding mechanisms.

Methods

CoPRA tackles this challenge by proposing a novel architecture and training strategy:

  • Co-Former Architecture: This core component fuses information from pre-trained protein and RNA language models (PLMs and RLMs) with structural data. It takes sequence embeddings from the interface regions of both molecules and combines them with structural pair embeddings. A "Structure-Sequence Fusion module" within the Co-Former integrates these modalities.
  • Bi-Scope Pre-training: To enhance the Co-Former's understanding of protein-RNA interactions, a two-pronged pre-training strategy is employed:
  • Contrastive Protein-RNA Interaction (CPRI): This task learns to distinguish interacting protein-RNA pairs from non-interacting ones, akin to image-text matching.
  • Mask Interface Distance Modeling (MIDM): This task focuses on predicting the precise distances between atoms at the binding interface, providing fine-grained interaction details. This is achieved by masking portions of the interface structure information and requiring the model to reconstruct the distances.
  • Interface Representation: CoPRA specifically focuses on the binding interface. It extracts sequence embeddings from the relevant regions of protein and RNA sequences and generates pair-wise structural embeddings based on backbone atom positions, ensuring invariance to side-chain changes. Special nodes are also introduced to aggregate complex, protein, and RNA information.

Datasets

  • PRI30k: A newly curated, unsupervised pre-training dataset consisting of 30,000 protein-RNA complexes, derived from BioLiP2. This dataset is designed to capture diverse binding poses.
  • PRA310: The largest curated protein-RNA binding affinity dataset to date, containing 310 unique complexes. It was assembled by consolidating and carefully calibrating data from three public datasets: PDBbind, PRBABv2, and ProNAB.
  • PRA201: A subset of PRA310, specifically designed for stricter evaluation, containing only single-chain protein-RNA complexes.
  • mCSM blind test set: Used for evaluating the model's ability to predict mutation effects on binding affinity.

Key Insights and Ideas

  • Cross-modal collaboration is key: The paper demonstrates for the first time that combining pre-trained language models from different biological domains (protein and RNA) can significantly improve binding affinity prediction.
  • Structure matters, but so does sequence context: Integrating structural information alongside sequence embeddings from advanced language models provides a more holistic view of binding.
  • Interface focus is crucial: By concentrating on the binding interface, CoPRA effectively captures the most relevant information for predicting affinity.
  • Bi-scope pre-training enhances understanding: The CPRI and MIDM tasks provide complementary information, enabling the model to grasp both the presence of interaction and the fine-grained details of how it occurs. MIDM, particularly the focus on interface distance, appears to be more impactful for affinity prediction.
  • Larger models yield better results: Similar to unimodal models, increasing the size of the pre-trained language models used within CoPRA leads to improved performance, highlighting the importance of representation capacity.

Key Results

  • CoPRA achieves state-of-the-art performance on the PRA310 and PRA201 datasets for protein-RNA binding affinity prediction.
  • On PRA310, CoPRA achieved an RMSE of 1.391, MAE of 1.129, PCC of 0.580, and SCC of 0.589.
  • The model also demonstrates strong performance in predicting the binding affinity change caused by protein mutations, outperforming other models on the mCSM blind test set, especially after fine-tuning.
  • Ablation studies confirm the importance of each module within CoPRA, with the pre-training strategy and the Co-Former architecture being critical for its success.

Limitations and Future Work

  • Dataset limitations: While PRA310 is the largest curated dataset, the field still faces challenges with limited labeled complex structures and potential inconsistencies across datasets.
  • Future Directions: The authors suggest extending CoPRA to other biological domains, such as protein-DNA binding. Additionally, future work could focus on predicting the effects of multi-point mutations and RNA mutations, further expanding the model's predictive capabilities.

Summary updated

BibTeX
@inproceedings{han2025copra,
  author = {Rong Han and Xiaohong Liu and Tong Pan and Jing Xu and Xiaoyu Wang and Wuyang Lan and Zhenyu Li and Zixuan Wang and Jiangning Song and Guangyu Wang and Ting Chen},
  title = {{CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i1.32001},
}
Figure