Abstract
Protein-RNA interactions (PRIs) are vital to a variety of cellular activities. While experimental and computational techniques have been developed to study these interactions, previous limitations in database availability—specifically the lack of protein structure data—meant most computational methods relied heavily on sequence data. Recently, AlphaFold has revolutionized the field of biology. Foreseeably, PRI prediction will also be promoted significantly in the upcoming years. This survey reviews the field, covering binding site and binding preference prediction, commonly used datasets, features, and models, while identifying potential challenges and opportunities in the post-AlphaFold era.
Introduction: The Biological Gears
Protein-RNA interactions (PRIs) serve as the fundamental "gears" of the cell, driving essential processes from gene expression to protein synthesis. Disruptions in these interactions are linked to severe pathologies, including cancer and neurodegenerative disorders. Historically, studying these interactions has been bottlenecked by the high cost and slow pace of experimental methods like CLIP-seq.
Computational methods initially filled this gap by relying on sequence data—the linear "letters" of DNA and RNA. However, molecules are complex 3D objects; a protein may possess the correct sequence but the wrong conformation to bind an RNA molecule. The field is currently undergoing a paradigm shift from simple sequence matching to complex 3D structural modeling, catalyzed by the widespread availability of high-quality protein structures via tools like AlphaFold.
The Evolution of Prediction Methods
The development of PRI prediction can be traced through several technological stages, moving from traditional machine learning to sophisticated deep learning architectures.
- Traditional Machine Learning: Early models utilized hand-crafted features such as hydrophobicity, charge, and solvent accessibility, processed through Support Vector Machines (SVMs) or Random Forests.
- Sequence-based Deep Learning: The advent of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks allowed for automatic motif extraction from raw sequences (e.g., DeepBind).
- Structure-based Deep Learning: The current frontier employs 3D CNNs (treating proteins as voxels), Graph Neural Networks (GNNs) (treating atoms as nodes), and Point Clouds to capture the actual geometry of interactions.
Methodological Tracks: Sites, Preferences, and Merged Approaches
Computational approaches generally follow three main tracks:
- Binding Site Prediction (Protein-centric): These models predict where on a protein's 3D surface an RNA molecule is likely to attach. Successful models like NucleicNet focus on the physicochemical environment of specific surface "patches" rather than the global protein structure.
- Binding Preference Prediction (RNA-centric): These models predict which RNA sequences or structures a specific protein prefers. Hybrid architectures (CNN + RNN) often excel here by capturing both local motifs and long-range dependencies.
- The Merged Approach: This represents the most promising future direction, considering both protein and RNA information (sequence and structure) simultaneously to model the interaction holistically.
Data Foundations
The efficacy of these models depends on robust datasets:
- Sequence Data: Sources like doRiNA, iCount, and AURA 2 provide experimental data derived from CLIP-seq.
- Structural Data: The Protein Data Bank (PDB) remains the gold standard for experimental structures, while the AlphaFold DB has expanded the horizon with millions of predicted structures.
- RNA Secondary Structure: Databases like bpRNA and RASP provide critical information on loops, stems, and pseudoknots, which are essential for binding.
Key Insights: Why Structure Matters
The "AlphaFold era" has transformed structure-based models from niche tools to scalable solutions. Key findings from the survey include:
- Structural Superiority: In binding site prediction, methods using 3D structural information consistently outperform sequence-only methods because they can directly "see" the binding pockets.
- Conservation of Structure: Structural features, such as electrostatic patches, are often more biologically conserved and informative than raw sequences.
- In Vivo Context: Models like PrismNet demonstrate that incorporating in vivo RNA structure data—how RNA folds inside a living cell—significantly improves accuracy over "test tube" (in vitro) data.
Challenges and Future Directions
Despite recent progress, several hurdles remain:
- Atomic Resolution: Most current models operate at the "residue" level; moving to atomic-level resolution is necessary to understand the underlying chemistry.
- Molecular Dynamics: Proteins and RNA are flexible, dynamic entities. Future models must incorporate molecular dynamics to account for conformational changes during binding.
- Ab Initio Prediction: The ultimate goal is end-to-end prediction—folding both molecules and docking them in a single computational step.
- Interpretability: As models like GNNs and 3D CNNs grow in complexity, developing "Explainable AI" is crucial to ensure that predictions lead to actionable biological insights.
Summary updated