Developing optimal non-linear scoring function for protein design

Changyu Hu Ii, Xiang Li, Jie LiangBioinformatics 2004 · DOI

TL;DR

This paper introduces a novel nonlinear scoring function for protein design that significantly outperforms linear methods in distinguishing native protein sequences from a vast number of decoys. This advancement is crucial for accurately predicting protein structures and engineering new proteins with desired functions.

What Problem Does This Paper Address?

Protein design aims to identify amino acid sequences that fold into a specific target protein structure while avoiding alternative, incorrect folds. A critical component of this process is a "scoring function" that can accurately assess the "fitness" of a given sequence for a particular structure. Existing methods often rely on linear scoring functions, which, despite extensive optimization, struggle to perfectly discriminate native sequences from a massive pool of potential "decoy" sequences (sequences that are similar but fold incorrectly). This limitation hinders the accuracy and scope of protein design efforts. The authors tackle this by developing a more sophisticated nonlinear scoring function capable of capturing complex relationships in protein sequence-structure compatibility.

Methods

The core of this work lies in developing a nonlinear scoring function based on a mixture of Gaussian kernel functions. Instead of a simple weighted sum of pairwise interactions (as in linear models), this approach maps protein representations into a high-dimensional space where discrimination becomes more feasible.

Key methodological aspects include:

  • Kernel Function: A Gaussian kernel, $K(x, y) = e^{-\|x-y\|^2 / 2\sigma^2}$, is used to transform protein contact vectors into a higher-dimensional space.
  • Optimization: The authors employ a Support Vector Machine (SVM)-like approach, specifically optimizing a Lagrange dual form of quadratic programming. This method aims to find a hyperplane with the largest margin separating native proteins from decoys in the transformed high-dimensional space. This is a departure from previous methods that focused on maximizing score gaps.
  • Iterative Training: To handle the massive number of decoys, an iterative training strategy is used, progressively refining the scoring function by focusing on difficult-to-discriminate decoys.

Datasets

  • Training Set: A dataset of 440 native protein structures was compiled from the WHATIF database, filtered for sequence identity (<30%), resolution (<2.1Å), and R-factor (<0.21).
  • Decoy Set: For each native protein, a massive set of approximately 14 million sequence decoys was generated using a gapless threading method. This method involves threading a sequence from a larger protein onto the structure of a smaller protein, creating challenging decoys.
  • Test Set: An independent test set of 194 proteins (not present in the training set) was used for blind testing. A further set of 201 proteins was also used, including those with higher interchain contacts.

Key Insights and Ideas

  • Linearity is Insufficient: A fundamental finding is that no linear scoring function can perfectly discriminate a large set of native proteins from millions of sequence decoys. This highlights a critical limitation of previous approaches.
  • Nonlinearity is Key: The nonlinear nature of the Gaussian kernel function is essential for capturing the complex fitness landscape of protein sequences and structures, enabling better discrimination.
  • SVM-like Optimization: Adapting principles from statistical learning theory, particularly the concept of maximizing the margin in a high-dimensional space, proves highly effective for optimizing the scoring function.
  • Decoy Generation Matters: The use of challenging sequence decoys generated by threading is crucial for training a robust scoring function.

Key Results

  • Perfect Discrimination (Training Set): The developed nonlinear scoring function achieved perfect discrimination of the 440 native proteins from the 14 million sequence decoys in the training set.
  • High Accuracy (Blind Test): In a blind test on an independent set of 194 proteins, the nonlinear scoring function correctly identified 93.3% (181 out of 194) of native sequences.
  • Superior Performance: This performance significantly outperforms optimal linear scoring functions (80.9% and 73.7% success rates) and statistical potentials (58.2%) reported in the literature.
  • Robustness: The method also showed good performance on a set of 201 proteins and demonstrated the ability to discriminate large proteins from decoys.

Limitations and Future Work

  • Computational Cost: The nonlinear scoring function requires more computation than linear functions, although the authors note it is still manageable.
  • Generalization: While the method shows promise, the authors suggest further improvements by incorporating more realistic structural decoys (beyond those generated by threading) and potentially exploring other kernel functions or optimization criteria.
  • Protein Representation: The current approach is based on contact maps. Future work could extend this to other protein representations, such as those including explicit hydrogen bonds or higher-order interactions.
  • Misclassification Analysis: The authors identified that misclassifications often occur when native proteins have extensive interchain interactions or are part of complexes, suggesting that these factors could be better incorporated into future models.

Summary updated

BibTeX
@article{ii2004developing,
  author = {Changyu Hu II and Xiang Li and Jie Liang},
  title = {{Developing optimal non-linear scoring function for protein design}},
  journal = {Bioinformatics},
  year = {2004},
  doi = {10.1093/bioinformatics/bth369},
}
Figure