RBPtool: A Deep Language Model Framework for Multi-Resolution RBP-RNA Binding Prediction and RNA Molecule Design

Jiyue Jiang, Yitao Xu, Zikang Wang, Yihan Ye, Yanruisheng Shao, Yuheng Shan, Jiuming Wang, Xiaodan Fan, Jiao Yuan, Yu LiEMNLP 2025 · OpenReview

TL;DR

This paper introduces RBPtool, a novel deep learning framework that integrates RNA sequences with their 3D structures to predict RNA-binding protein (RBP) interactions with unprecedented accuracy. Beyond prediction, RBPtool can also design novel RNA molecules that specifically bind to target RBPs, opening new avenues for RNA-based therapeutics and biological research.

What Problem Does This Paper Address?

RNA-binding proteins (RBPs) are crucial for regulating gene expression by interacting with specific RNA molecules. Understanding these interactions is key to deciphering cellular processes and designing new RNA-based drugs. However, existing methods often rely on simplified sequence or structural features, failing to capture the complex, multi-modal nature of RBP-RNA binding. This limitation hinders both accurate prediction and the design of functional RNA molecules. The authors aim to bridge this gap by developing a framework that leverages the power of large language models (LLMs) and geometric deep learning to integrate diverse biological data for a more comprehensive understanding of RBP-RNA interactions.

Methods

RBPtool is a multi-task, multi-resolution framework built upon a dual-channel architecture.

  • Sequence Encoding: It utilizes pre-trained LLMs like RNA-FM to capture rich contextual information from RNA sequences.
  • Structure Encoding: A Geometric Vector Perceptron Graph Neural Network (GVP-GNN) module is employed to extract SE(3)-equivariant structural features from RNA backbones. This allows the model to understand the 3D geometry of RNA molecules.
  • Multi-Modal Integration: The framework fuses sequence and structural embeddings. For tasks involving protein information, it incorporates protein sequence embeddings using an attention mechanism.
  • Hierarchical Prediction: RBPtool predicts RBP-RNA binding at three levels: sequence (overall binding), residue (specific nucleotides involved), and atom (individual atom interactions). This multi-resolution approach provides a detailed understanding of binding.
  • Generative Design: A generative module, conditioned on protein targets, cell types, and species, is incorporated to design novel RNA sequences with specific binding properties.

Datasets

The authors utilized several datasets for training and evaluation:

  • CLIP: A dataset containing RNA sequences with experimentally verified binding sites.
  • RNAcompete: A dataset derived from in vitro RNAcompete assays, providing RNA sequences and binding scores.
  • RPI15223: A newly curated, multi-resolution, structure-level RBP-RNA dataset. This dataset was constructed by retrieving RBP-RNA complex structures from PDB and EMDB, filtering them based on resolution and size, and identifying binding pairs. A refined version of RPI15223 was also created to support residue-level predictions.

Key Insights and Ideas

  • Synergy of Sequence and Structure: Combining RNA sequence information with its 3D structure significantly boosts RBP-RNA binding prediction accuracy, highlighting the importance of both modalities.
  • LLMs as a Foundation: Pre-trained LLMs like RNA-FM provide a powerful foundation for understanding RNA sequences, significantly improving performance even without structural information.
  • Hierarchical Prediction Matters: Predicting binding at different resolutions (sequence, residue, atom) offers a more nuanced understanding of RBP-RNA interactions and can reveal different aspects of the binding process.
  • Generative Design for Function: The ability to design RNA sequences conditioned on specific biological contexts (protein, cell type, species) is a significant step towards creating functional RNA molecules for therapeutic or research purposes.

Key Results

RBPtool achieved state-of-the-art (SOTA) performance across various benchmarks:

  • Binding Prediction: It outperformed existing methods on CLIP and RNAcompete datasets for sequence-level binding prediction, achieving higher ACC, AUPR, and AUROC scores. For instance, on the CLIP dataset, RBPtool achieved an ACC of 0.773 ± 0.067, significantly higher than baselines.
  • Fine-grained Prediction: RBPtool demonstrated strong performance in residue-level and atom-level binding prediction on the RPI15223 dataset, with AUPR scores of 0.697 and 0.687, respectively.
  • RNA Design: The RNA sequences designed by RBPtool showed significantly higher Weighted Success Rates (WSR) compared to random or genetic algorithm-based designs, indicating their ability to specifically bind target RBPs.

Limitations and Future Work

  • Experimental Validation: The authors acknowledge that model scoring for RNA design cannot fully replace wet-lab experimental validation. Sequences predicted to bind highly may not always achieve their expected functions in vitro or in vivo.
  • Generalizability: The current experiments focus on typical RBPs, species, and cell types. Further validation is needed for less common or structurally complex RBPs, viral RNAs, or rare species.
  • Future Directions: The authors suggest that incorporating richer biological data and more advanced generative algorithms could further enhance cross-species and pathological predictions. They also highlight the potential for RBPtool to advance fundamental RBP biology research and RNA-based drug design.

Summary updated

Abstract

RNA-binding proteins (RBPs) play essential roles in post-transcriptional gene regulation via recognizing specific RNA molecules as well as modulating several key physiological processes in cellulo, represented by alternative splicing and RNA degradation. Despite extensive research, most existing approaches still rely on superficial sequence features or coarse structural representations, limiting their ability to capture the intricate nature of RBP-RNA interactions. The recent surge in large language models (LLMs), combined with advances in geometric deep learning for extracting three-dimensional representations, enables the integration of multi-modal, multi-scale biological data for precise modeling and biologically informed de novo RNA design. In this work, we curate and extend RPI15223 into a multi-resolution, structure-level RBP-RNA dataset, and introduce RBPtool, a multi-task, multi-resolution framework that combines a geometric vector perception (GVP) module together with a deep language model encoder to fuse sequence and structural information. Our tool achieves state-of-the-art performance on public benchmarks and the RPI15223 dataset, while also supporting fine-grained level predictions and enabling de novo RNA design through a generative module conditioned on protein, cell-type, and specified species. RBPtool provides a fast and versatile platform for both fundamental RBP-RNA research and practical RNA drug design, delivering enhanced predictive accuracy and fine-grained structural insights.
BibTeX
@inproceedings{jiang2025rbptool,
  author = {Jiyue Jiang and Yitao Xu and Zikang Wang and Yihan Ye and Yanruisheng Shao and Yuheng Shan and Jiuming Wang and Xiaodan Fan and Jiao Yuan and Yu Li},
  title = {{RBPtool: A Deep Language Model Framework for Multi-Resolution RBP-RNA Binding Prediction and RNA Molecule Design}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=k4ExQJcN8s},
}
Figure