Protein-RNA interaction prediction with deep learning: structure matters

Junkang Wei, Siyuan Chen, Licheng Zong, Xin Gao, Yu Li

TL;DR

This paper provides a comprehensive survey of how deep learning is transforming our ability to predict protein-RNA interactions. The authors argue that while sequence-based models have dominated the field, the "AlphaFold era" is ushering in a new generation of structure-aware models that offer far greater accuracy and biological insight.

What Problem Does This Paper Address?

Protein-RNA interactions (PRIs) are the "gears" of the cell, driving everything from gene expression to protein synthesis. When these interactions fail, it leads to serious diseases, including cancer and neurodegenerative disorders.

Historically, researchers have faced a major bottleneck: experimental methods like CLIP-seq are expensive and slow, while computational methods have mostly relied on sequence data (the "letters" of DNA/RNA). However, molecules are 3D objects. A protein might have the "right" sequence but the "wrong" shape to bind an RNA molecule. The challenge the authors address is how to move the field from simple sequence matching to complex 3D structural modeling, especially now that tools like AlphaFold have made high-quality protein structures widely available.

Methods

The paper categorizes computational approaches into two main tracks:

  1. Binding Site Prediction (Protein-centric): Predicting where on a protein's 3D surface an RNA molecule will attach.
  2. Binding Preference Prediction (RNA-centric): Predicting which RNA sequences or structures a specific protein prefers to bind.

The authors trace the evolution of these methods from:

  • Traditional Machine Learning: Using hand-crafted features (like hydrophobicity or charge) with Support Vector Machines (SVMs) or Random Forests.
  • Deep Learning (Sequence-based): Using CNNs and LSTMs to automatically extract motifs from raw sequences (e.g., DeepBind).
  • Deep Learning (Structure-based): The cutting edge. These models use 3D CNNs (treating proteins like 3D images/voxels), Graph Neural Networks (GNNs) (treating atoms as nodes in a network), and Point Clouds to capture the actual geometry of the interaction.

Datasets

The paper provides a roadmap of essential data sources:

  • Sequence Datasets: doRiNA, iCount, and AURA 2 (derived from CLIP-seq experiments).
  • Structural Datasets: The Protein Data Bank (PDB) remains the gold standard for experimental 3D structures, while the AlphaFold DB has recently added millions of predicted structures.
  • RNA Structure: bpRNA and RASP provide data on RNA secondary structures (loops, stems, and pseudoknots), which are critical for binding.

Key Insights and Ideas

  • The "AlphaFold" Catalyst: We are entering a "post-AlphaFold era." Previously, structure-based models were limited by a lack of 3D data. Now, we can predict the structure of almost any protein, allowing structure-based PRI models to scale to the entire proteome.
  • Local vs. Global Context: The paper highlights that binding is a local phenomenon. Successful models (like NucleicNet) focus on the physicochemical environment of specific surface "patches" rather than the whole protein.
  • Structure vs. Sequence: While sequence tells you what a molecule is, structure tells you what it can do. The authors emphasize that structural features (like solvent accessibility and electrostatic patches) are often more conserved and informative than the raw sequence.
  • The Merged Approach: The most promising future direction is the "Merged Approach," which considers both protein and RNA information (sequence + structure) simultaneously, rather than looking at one side of the interaction in a vacuum.

Key Results

The survey summarizes several performance trends:

  • Structure Wins: In binding site prediction, methods utilizing 3D structural information consistently outperform sequence-only methods because they can "see" the binding pockets.
  • Hybrid Models: For RNA preference, hybrid architectures (CNN + RNN) like DeepRAM generally perform better than single-type networks by capturing both local motifs and long-range dependencies.
  • In Vivo Matters: Models like PrismNet show that incorporating in vivo RNA structure data (how RNA folds inside a living cell) significantly improves prediction accuracy over models that only use "test tube" (in vitro) data.

Limitations and Future Work

The authors identify several hurdles that still need to be cleared:

  • Resolution: Most current models predict binding at the "residue" level. We need to move toward atomic-level resolution to truly understand the chemistry of the interaction.
  • Dynamics: Molecules are not rigid statues; they "breathe" and change shape. Future models need to incorporate molecular dynamics to account for this flexibility.
  • Ab Initio Prediction: The ultimate goal is to predict interactions directly from sequences by folding both the protein and RNA and docking them in one end-to-end computational step.
  • Interpretability: As models become more complex (GNNs, 3D CNNs), it becomes harder to understand why they make a prediction. Improving "Explainable AI" in this field is crucial for biological discovery.

Summary updated

Abstract

Protein–RNA interactions are of vital importance to a variety of cellular activities. Both experimental and computational techniques have been developed to study the interactions. Because of the limitation of the previous database, especially the lack of protein structure data, most of the existing computational methods rely heavily on the sequence data, with only a small portion of the methods utilizing the structural information. Recently, AlphaFold has revolutionized the entire protein and biology field. Foreseeably, the protein–RNA interaction prediction will also be promoted significantly in the upcoming years. In this work, we give a thorough review of this field, surveying both the binding site and binding preference prediction problems and covering the commonly used datasets, features and models. We also point out the potential challenges and opportunities in this field. This survey summarizes the development of the RNA-binding protein–RNA interaction field in the past and foresees its future development in the post-AlphaFold era.
BibTeX
@article{wei2022protein,
  author = {Junkang Wei and Siyuan Chen and Licheng Zong and Xin Gao and Yu Li},
  title = {{Protein-RNA interaction prediction with deep learning: structure matters}},
  journal = {Briefings in Bioinformatics},
  year = {2022},
  doi = {10.1093/bib/bbab540},
}

Loading related content...

Figure