TL;DR
This paper introduces trRosettaRNA, a novel deep learning method that significantly improves de novo RNA 3D structure prediction. By leveraging transformer networks for predicting intermediate geometric constraints, trRosettaRNA achieves higher accuracy than existing methods, particularly in predicting physically realistic side-chain atoms, and can predict structures for previously uncharacterized RNA families.
What Problem Does This Paper Address?
Predicting the three-dimensional (3D) structure of RNA molecules is a fundamental challenge in molecular biology. Unlike proteins, where experimental structures are abundant, RNA structures are less well-characterized (around 6,000 in the PDB vs. ~190,000 for proteins). This scarcity is partly due to RNA's inherent flexibility and the difficulty of experimental determination. The 3D structure is crucial because it dictates RNA's biological function. Existing prediction methods fall into two categories: template-based (reliant on known homologous structures) and de novo (predicting from scratch). While template-based methods are accurate when suitable templates exist, they are limited by the availability of known structures and the difficulty of aligning RNA sequences. De novo methods, particularly those simulating folding from scratch, struggle with the vast conformational space and inaccurate force fields, especially for larger RNAs. Recent advances in deep learning for protein structure prediction (like AlphaFold2) have inspired new approaches for RNA, but accurately predicting RNA 3D structures, especially for novel families, remains an open problem.
Methods
The core of trRosettaRNA is a deep learning pipeline inspired by the success of trRosetta for proteins and AlphaFold2 for proteins. It follows a two-step procedure:
- 1D and 2D Geometries Prediction: A transformer network, dubbed "RNAformer," takes a multiple sequence alignment (MSA) and predicted secondary structure as input. This network predicts various geometric properties, including inter-nucleotide distances, orientations, and contact probabilities. This is a key innovation, as it moves beyond simpler contact prediction to capture more nuanced geometric information.
- Full-Atom 3D Structure Folding: The predicted 1D and 2D geometries are then used as restraints in an energy minimization process to fold a full-atom RNA model. This step refines the structure based on the learned geometric constraints.
The RNAformer architecture itself is novel, adapting the Evoformer architecture from AlphaFold2. It uses gated self-attention mechanisms to iteratively update both MSA and pair representations, allowing for rich interaction modeling between different parts of the RNA sequence and its homologous sequences.
Datasets
- Test Sets:
- RNA-Puzzles: A benchmark dataset comprising 30 RNA targets from the RNA-Puzzles experiments. This set is used to evaluate performance against established methods and recent deep learning approaches.
- PDB-derived set: 101 non-redundant RNA structures collected from the Protein Data Bank (PDB) that were released after January 2017 (to avoid overlap with training data). This set provides a more general evaluation of de novo prediction capabilities.
- Training Sets:
- PDB-derived training set: All RNA chains released before January 2022 from PDB, after processing (e.g., separating multi-chain structures, replacing modified nucleotides). This forms the primary training data.
- Self-distillation training set: Constructed using sequences from the bpRNA database and Rfam MSAs. This dataset, inspired by AlphaFold2's approach, aims to augment the limited experimental RNA structure data by using predictions from a pre-trained model as labels.
Key Insights and Ideas
- Transformer Networks for RNA Geometry: The paper demonstrates that transformer networks, successfully applied to protein structure prediction, are highly effective for predicting intermediate 1D and 2D geometric constraints in RNA. This is a significant step beyond predicting simple contact maps.
- MSA and Secondary Structure Synergy: The combination of MSA information (capturing evolutionary conservation) and predicted secondary structure is crucial for accurate geometric prediction. The model shows a strong correlation between secondary structure precision and prediction accuracy.
- Side-Chain Accuracy Matters: While overall RMSD is important, trRosettaRNA shows a notable advantage in predicting more accurate and physically realistic side-chain atoms compared to other deep learning methods. This suggests a deeper understanding of the atomic-level interactions.
- Predicting Novel Topologies: The application to Rfam families with unknown structures reveals that trRosettaRNA can predict structures with novel topologies, which are not represented in existing PDB databases. This is vital for expanding our understanding of RNA structural diversity.
Key Results
- Outperforms Conventional Methods: On a dataset of 101 RNAs, trRosettaRNA achieved an average RMSD of 4.0 Å, significantly outperforming RNAComposer (18.8 Å) and SimRNA (19.2 Å). 74.2% of its predictions were within 4 Å RMSD.
- Competitive with State-of-the-Art Deep Learning: On RNA-Puzzles targets, trRosettaRNA achieved an average RMSD of 5.5 Å, comparable to or better than recent deep learning methods like DeepFoldRNA (5.1 Å) and RoseTTAFoldNA (average RMSD of 3.8 Å on a subset of RNAs, though with some outliers).
- Superior Side-Chain Prediction: trRosettaRNA showed better accuracy for side-chain atoms (e.g., N3(1) atom) compared to DeepFoldRNA.
- Efficient Prediction: For typical RNAs (<200 nucleotides), the prediction takes less than 30 minutes.
- Application to Uncharacterized RNAs: trRosettaRNA successfully predicted accurate structures (eRMSD < 4 Å) for 263 Rfam families with no known experimental structures, including 27 families with novel topologies.
Limitations and Future Work
- Outliers and Complex Cases: While generally accurate, trRosettaRNA, like other methods, can struggle with highly complex RNA targets, such as those in free conformations or interacting with proteins, as seen in some RNA-Puzzles cases.
- Data Limitations: The performance is still influenced by the quality and quantity of available experimental RNA structures for training. The paper notes that the 263 predicted Rfam structures represent only a "small proportion" of current Rfam families, suggesting room for improvement as more data becomes available.
- Future Work: The authors believe that continued algorithmic improvements and the deposition of more experimental RNA structures will lead to higher accuracy for more families in the future. The availability of the trRosettaRNA server and package encourages further research and application.
Summary updated