STRAND: Structure Refinement of RNA-Protein Complexes via Diffusion

Mohsen Al-zeqri, Jörg K.h. Franke, Frederic RungeGenBio @ ICML 2025 · OpenReview

TL;DR

Predicting the 3D structure of RNA-protein complexes is crucial but challenging. This paper introduces STRAND, a novel diffusion-based model that refines existing predictions of these complexes, significantly improving their accuracy. STRAND builds upon protein-protein docking methods but is specifically adapted for RNA-protein interactions, demonstrating notable improvements over state-of-the-art predictors like AlphaFold 3.

What Problem Does This Paper Address?

RNA-protein interactions are fundamental to countless biological processes, from gene regulation to viral replication. Understanding these interactions at a structural level is key to deciphering molecular mechanisms and developing new therapeutics. While significant progress has been made in predicting protein structures (e.g., AlphaFold 2) and even protein-protein complexes, accurately predicting the 3D structure of RNA-protein complexes remains a substantial challenge. This difficulty stems from the inherent flexibility of RNA molecules and the diverse nature of their binding interfaces. Existing deep learning models, while powerful, often struggle to achieve high accuracy for these complex structures, leaving a gap in our ability to model them reliably.

Methods

STRAND is a diffusion-based model designed for the refinement of pre-existing RNA-protein complex structures. It builds upon the principles of DiffDock-PP, a successful diffusion model for protein-protein docking, but adapts it for RNA-protein systems.

Key methodological aspects include:

  • Modular Extension: STRAND is designed as a modular component that can be plugged into existing RNA-protein complex prediction pipelines to enhance their output.
  • RNA Foundation Model Integration: It leverages an RNA foundation model, RNA-FM, to generate embeddings that capture the unique properties of RNA molecules. These embeddings are then integrated into the diffusion framework.
  • Diffusion Process Adaptation: Unlike traditional docking, STRAND focuses on refining the backbone structure. The diffusion process is applied to the protein component, treating the RNA as a fixed receptor.
  • Transformation Exploration: The authors systematically investigate the impact of different types of transformations during the diffusion process:
  • Translation: Moving the protein relative to the RNA.
  • Rotation: Rotating the protein around its center of mass.
  • Torsion: Adjusting rotatable bonds within the protein backbone.
  • Combinations: Exploring various combinations of these transformations (e.g., translation + rotation).
  • Intrinsic Diffusion Models: STRAND utilizes the framework of intrinsic diffusion models to handle the complex geometric constraints of molecular structures efficiently.
  • Neural Architecture: The model adapts the architecture of DiffDock-PP, introducing separate embedding layers for protein and RNA components and extending it to predict torsional scores for finer structural modeling.

Datasets

The study utilizes publicly available datasets for training and evaluation:

  • Protein Data Bank (PDB): The primary source for experimentally validated RNA-protein complexes.
  • Training Data: A comprehensive set of RNA-protein complexes from the PDB with deposition dates before September 30, 2021 (the AlphaFold 3 training cutoff). This dataset includes both monomeric and multimeric complexes, with one molecular type treated as the receptor and others as the ligand.
  • Evaluation Data: * X-ray Crystallography: A subset of 35 complexes derived from X-ray crystallography.
  • Non-X-ray: A subset of 46 complexes derived from other experimental techniques like NMR and cryo-EM.
  • ProRNA3D-single Evaluation: An additional set of 28 complexes from the PDB, combined with the X-ray dataset to form a total of 63 samples for evaluating refinements of ProRNA3D-single predictions.
  • AlphaFold 3 Predictions: Predictions from AlphaFold 3 are used as initial inputs for STRAND's refinement process.
  • ProRNA3D-single Predictions: Predictions from ProRNA3D-single are also used as initial inputs for STRAND's refinement.

Key Insights and Ideas

  • Refinement is Key: The paper strongly suggests that while state-of-the-art predictors like AlphaFold 3 are already quite good, there's significant room for improvement through dedicated refinement strategies. STRAND demonstrates that even small adjustments (translation, rotation) can yield substantial gains.
  • Translation and Rotation are Powerful: The most impactful transformations for refining RNA-protein complexes appear to be translation and rotation. This implies that initial predictions often misplace the protein relative to the RNA, rather than having severe internal structural errors.
  • Torsion's Limited Impact (in this context): Refining torsion angles alone showed less improvement. This might be because torsion adjustments primarily affect internal protein flexibility, while the major errors in RNA-protein complex prediction often lie in the global positioning of the protein relative to the RNA.
  • Modular Approach Works: STRAND's success as a modular extension highlights the potential for building specialized refinement tools that can enhance existing general-purpose structure predictors.
  • Data Diversity Matters: Evaluating on both X-ray and non-X-ray datasets reveals that STRAND's improvements are consistent, though the degree of improvement can vary depending on the initial prediction quality and experimental validation method.

Key Results

  • Significant Improvement over AlphaFold 3: STRAND consistently improved AlphaFold 3 predictions across various metrics. For X-ray data, STRANDtr+rot achieved the lowest mean cRMSD (2.97Å) compared to AlphaFold 3's 3.09Å. For non-X-ray data, STRANDtr+rot also showed improvements.
  • Enhanced Prediction Accuracy: STRAND was able to improve the percentage of predictions below critical cRMSD thresholds (e.g., 2Å and 5Å) for a substantial fraction of complexes.
  • Effective Refinement of ProRNA3D-single: STRAND also demonstrated significant improvements when refining predictions from ProRNA3D-single, with STRANDtr+rot showing the largest gains.
  • Confidence Model for Sample Selection: The authors developed a confidence model to automatically select the best refined pose from multiple STRAND outputs, achieving competitive results and reducing the need for manual selection.
  • Best Performing Models: The STRANDtr+rot model (combining translation and rotation) generally showed the best performance, indicating the importance of global positional adjustments.

Limitations and Future Work

  • Backbone Only Refinement: Currently, STRAND refines only the backbone atoms of the protein. The authors acknowledge that full-atom refinement would be preferable.
  • Sequence Embedding Length Limit: The model is limited by the sequence embedding length of the foundation models (currently 1022 residues).
  • Confidence Model Training Data: The confidence model was trained on a limited fraction of the data due to computational constraints, suggesting potential for further improvement with more data.
  • Torsion Refinement Challenges: The authors note that torsion angle refinement might require further development, particularly regarding noise schedules.
  • Future Directions:
    • Develop a full-atom refinement version of STRAND.
  • Explore optimizing the confidence model for cRMSD directly, rather than ligand RMSD.
  • Investigate scaling up the training data for the confidence model.
  • Develop a version of STRAND that diffuses over the RNA instead of the protein.

Summary updated

Abstract

RNA-protein interactions play crucial roles in cellular processes, from gene regulation to viral replication. While recent advances in structure prediction have revolutionized our ability to model macromolecular complexes, achieving accurate predictions of RNA-protein binding poses remains challenging. In this work, we present STRAND, a diffusion-based model for monomeric RNA-protein complex refinement that builds upon the success of DiffDock-PP in protein-protein docking. Unlike traditional docking, we develop STRAND as a modular extension to existing RNA-Protein complex prediction tools to improve their backbone predictions. We study the effect of different transformations by training models to learn either translation, rotation, torsion, or combinations of these during the diffusion process and initialize the backward process with a complex prediction at test time. Our experiments with AlphaFold 3 and ProRNA3D-single reveal that STRAND can improve the backbones of a large fraction of RNA-protein complex predictions.
BibTeX
@inproceedings{alzeqri2025strand,
  author = {Mohsen Al-zeqri and Jörg K.H. Franke and Frederic Runge},
  title = {{STRAND: Structure Refinement of RNA-Protein Complexes via Diffusion}},
  booktitle = {GenBio @ ICML},
  year = {2025},
  url = {https://openreview.net/forum?id=NEealHqriE},
}
Figure