RaptorX-Angle: real-value prediction of protein backbone dihedral angles through a hybrid method of clustering and deep learning

Yujuan Gao, Sheng Wang, Minghua Deng, Jinbo XuBMC Bioinformatics 2018 · DOI

TL;DR

This paper introduces RaptorX-Angle, a novel method that combines k-means clustering with deep learning to predict protein backbone dihedral angles with improved accuracy. By transforming the problem into a discrete label assignment task and then leveraging deep learning, RaptorX-Angle achieves state-of-the-art performance, offering a more reliable way to constrain protein structure predictions.

What Problem Does This Paper Address?

Predicting the three-dimensional structure of proteins from their amino acid sequence is a fundamental challenge in biology. While direct prediction is difficult, a hierarchical approach involving intermediate predictions has proven effective. One crucial intermediate step is predicting protein backbone dihedral angles (phi and psi), which describe the local conformation of each amino acid residue. These angles significantly narrow down the possible overall protein structures. However, directly predicting these real-valued angles from sequence alone is challenging. Existing methods often struggle with the inherent periodicity of angles and the accumulation of prediction errors, limiting their direct utility in building accurate 3D models. This paper aims to address these limitations by developing a more accurate and robust method for dihedral angle prediction.

Methods

RaptorX-Angle employs a hybrid approach that cleverly combines clustering and deep learning.

  1. Angle Representation: Instead of directly using the angles, they are converted into a 4D vector (cos($\phi$), sin($\phi$), cos($\psi$), sin($\psi$)). This representation handles the circular nature of angles and avoids issues with periodicity.
  2. K-Means Clustering: The angle vectors from a large dataset are clustered using k-means. This partitions the conformational space into a predefined number of "basins" or clusters, each representing a typical local conformation.
  3. Deep Learning Classification: The problem is reframed as a classification task. For each residue, the model predicts the probability of it belonging to each of the k-means clusters. A Deep Convolutional Neural Network (DCNN) with a Residual Network (ResNet) architecture is used for this classification. ResNet is chosen for its ability to train very deep networks effectively, capturing complex local dependencies in the sequence.
  4. Real-Value Prediction: The final real-valued dihedral angles are predicted by taking a weighted average of the cluster centers, where the weights are the predicted probabilities from the deep learning model. This "mixing" approach leverages the information from all relevant clusters, rather than just assigning a single label.

Datasets

  • PDB25: A large, non-redundant dataset of protein structures (10820 chains) was used. A subset of 5070 chains was used for training (TR5046 after filtering), 1267 for validation (VL1267), and 1267 for testing (TS1267).
  • CASP targets: 85 CASP11 and 40 CASP12 targets were used for further evaluation. Redundancy between training and CASP targets was checked using MMseqs2.

Key Insights and Ideas

  • Transforming a Regression to Classification: The core innovation is reframing the real-valued angle prediction as a classification problem (predicting cluster probabilities). This simplifies the learning task and allows for a more robust prediction by combining information from multiple clusters.
  • Leveraging Cluster Probabilities: Instead of assigning a single cluster label, using the probabilities of belonging to each cluster to predict the final angle is crucial. This "mixing" approach smooths predictions and accounts for uncertainty.
  • Periodicity Handled by Vectorization: Converting angles to sine and cosine vectors is an effective way to handle their circular nature, a common challenge in this field.
  • Deep Residual Networks for Local Context: The use of ResNet allows the model to effectively learn from the sequence context, as dihedral angles are known to be influenced by neighboring residues.
  • Confidence Scores: The paper explores estimating confidence scores for predictions, which is vital for their application in downstream tasks like structure prediction.

Key Results

  • State-of-the-Art Accuracy: RaptorX-Angle achieved the best performance on the TS1267 dataset, outperforming SPIDER2 by approximately 0.5° and 1.4° in Mean Absolute Error (MAE) for $\phi$ and $\psi$ angles, respectively.
  • Improved Performance on CASP Targets: The method also showed superior performance on CASP11 and CASP12 test targets compared to SPIDER2 and other methods.
  • Feature Importance: The study found that combining sequence information (PSSM and PSFM) with predicted secondary structure (SS) and solvent accessibility (ACC) yielded the best results.
  • Linear Relationship Between Error and Standard Deviation: A notable finding is the approximately linear relationship between the real prediction error and the estimated bounds (standard deviation derived from cluster probabilities). This suggests that the method's confidence estimates are well-calibrated.
  • Computational Efficiency: RaptorX-Angle is computationally efficient, especially when predicting angles for many proteins in parallel, making it practical for large-scale applications.

Limitations and Future Work

  • One-Dimensional Features: The current RaptorX-Angle uses one-dimensional features and a 1D CNN, which may not fully capture long-range interactions within the protein sequence.
  • Error Accumulation: The paper acknowledges that accumulation of prediction errors can still limit the direct use of dihedral angles for 3D model construction.
  • Future Directions:
    • Investigate two-dimensional features and 2D CNNs to capture long-range interactions.
  • Develop more robust techniques to handle prediction errors and incorporate angle restraints and confidence scores into protein tertiary structure prediction pipelines.
  • Explore the application of dihedral angle prediction for structure alignment and fold recognition.

Summary updated

BibTeX
@article{gao2018raptorx,
  author = {Yujuan Gao and Sheng Wang and Minghua Deng and Jinbo Xu},
  title = {{RaptorX-Angle: real-value prediction of protein backbone dihedral angles through a hybrid method of clustering and deep learning}},
  journal = {BMC Bioinformatics},
  year = {2018},
  doi = {10.1186/s12859-018-2065-x},
}
Figure