Small Libraries of Protein Fragments Model Native Protein Structures Accurately

Rachel Kolodny, Patrice Koehl, Leonidas Guibas, Michael LevittJournal of Molecular Biology 2002 · DOI

TL;DR

This paper introduces a novel method for representing protein structures using small, discrete libraries of "fragments." By intelligently clustering known protein backbone structures, they create compact libraries that can accurately model native protein conformations, achieving resolutions previously requiring much larger models. This work offers a more efficient way to represent and analyze protein structures, with implications for protein folding, modeling, and design.

What Problem Does This Paper Address?

Understanding and predicting the three-dimensional structure of proteins is a fundamental challenge in biology. Proteins are complex molecules, and their function is intimately tied to their shape. Traditional approaches often simplify protein structures by focusing on local backbone conformations. Early work by Corey & Pauling and Ramachandran identified key local structures (like alpha-helices and beta-sheets) based on the limited possible angles of amino acid residues. Later, researchers like Unger et al. and Micheletti et al. proposed representing protein backbones as sequences of "fragments" – short, contiguous stretches of amino acids (typically 4-10 residues long) that exhibit recurring structural patterns.

The challenge lies in creating representations that are both accurate (closely matching real protein structures) and economical (requiring a small number of "states" or possibilities per residue). Previous discrete models often struggled to achieve high accuracy without becoming computationally very complex. This paper aims to bridge this gap by developing a method to construct "small libraries" of representative protein fragments that can accurately model native protein structures with significantly reduced complexity.

Methods

The core of this paper is a novel clustering technique called simulated annealing k-means used to build representative libraries of protein fragments.

  1. Fragment Extraction: They start by extracting consecutive, non-overlapping fragments of specific lengths (4, 5, 6, and 7 residues) from a large dataset of reliable protein structures.
  2. Clustering: The simulated annealing k-means algorithm is applied to these fragments. This method iteratively groups similar fragments into clusters. The "centroids" of these clusters (the fragments that best represent each cluster) form the library. The simulated annealing aspect helps the algorithm escape local minima and find more robust clusterings, especially when dealing with unevenly distributed data (like the overrepresentation of alpha-helices).
  3. Library Evaluation: The quality of these fragment libraries is assessed using two criteria:
  4. Local-fit: Measures how well individual fragments from a test protein can be approximated by fragments in the library (using cRMS deviation).
  5. Global-fit: Measures how well the entire 3D structure of a test protein can be reconstructed using the library (also using cRMS deviation).
  6. Approximation Construction:
    • Local-fit: This is a straightforward process where each fragment of a target protein is matched to the most similar fragment in the library.
  7. Global-fit: This is more complex. They use a greedy algorithm where each new fragment is added by superimposing its first three Cα atoms onto the last three Cα atoms of the previously added fragment. This ensures a discrete, one-to-one mapping between a sequence of library fragment codes and a 3D structure. To improve robustness, they employ a heap-based approach (keeping track of the best Nkeep approximations) to explore a slightly larger portion of the approximation space.

Datasets

  • Training Set: A dataset of 200 unique protein domains from the SCOP database (version 1.57) with the highest SPACI scores (indicating high structural reliability). These domains were used to extract fragments for building the libraries.
  • Test Set: A set of 145 proteins, previously used by Park & Levitt (1995), serving as an independent benchmark for evaluating the accuracy of the constructed libraries. This set includes proteins with diverse structural motifs.

Key Insights and Ideas

  • Fragment Libraries as Discrete Models: The core idea is that a protein's 3D structure can be represented as a "string" of discrete fragment codes, similar to how an amino acid sequence defines a protein. This discretization dramatically reduces the conformational space.
  • Simulated Annealing K-Means for Robust Clustering: The choice of clustering algorithm is crucial. Simulated annealing k-means proved superior to standard k-means and hierarchical clustering, particularly in handling the uneven distribution of fragment types and being less sensitive to initial random seeds.
  • Fragment Length vs. Complexity Trade-off: Longer fragments capture more local correlation between residues, leading to better accuracy for a given complexity. However, creating libraries of longer fragments requires significantly more data.
  • Local-fit Predicts Global-fit: The accuracy of local-fit approximations strongly correlates with the accuracy of global-fit approximations. This means libraries that are good at approximating individual fragments are also good at reconstructing the overall structure.
  • Efficiency Gains: The developed libraries and methods achieve high accuracy (approaching 1 Å) with significantly lower complexity (e.g., 10-15 states per residue) compared to previous models that required over 50 states per residue for similar accuracy. This has a massive impact on the size of the conformational space that needs to be explored.

Key Results

  • Accuracy: The fragment libraries can approximate native protein structures with Global-fit cRMS deviations ranging from 2.9 Å to 0.76 Å.
  • Complexity: These approximations are achieved with complexities ranging from 2.66 to 15 states per residue.
  • Improved Accuracy with Complexity: For a fixed fragment length, increasing library complexity (more fragments) leads to better accuracy.
  • Improved Accuracy with Fragment Length: For a fixed complexity, longer fragments generally provide better fits.
  • Comparison to Previous Work: The models developed here are substantially better in terms of accuracy and economy compared to previous discrete models (e.g., Park & Levitt's model). For example, achieving 1 Å accuracy required a complexity of ~12 for a 4-residue fragment or ~10 for a 5-residue fragment, whereas Park & Levitt's model needed >50 states per residue.

Limitations and Future Work

  • Data Requirements for Longer Fragments: The ability to construct reliable libraries for longer fragments (e.g., 7 residues) is limited by the availability of large datasets of high-resolution protein structures. The authors anticipate that with the rapid pace of structure determination, such libraries will become feasible.
  • Greedy Algorithm for Global-fit: While the global-fit approximation uses a greedy algorithm, which is efficient, it doesn't guarantee the absolute optimal solution. The heap-based approach improves this by exploring a slightly larger space.
  • Future Applications: The authors suggest their fragment libraries and methods will be useful for:
  • Generating better decoy sets for ab initio protein folding.
  • Generating accurate loop conformations in homology modeling.
  • Analyzing strings of conformational states, analogous to analyzing amino acid sequences.
  • Low-resolution structure determination by NMR or X-ray crystallography.

Summary updated

BibTeX
@article{kolodny2002small,
  author = {Rachel Kolodny and Patrice Koehl and Leonidas Guibas and Michael Levitt},
  title = {{Small Libraries of Protein Fragments Model Native Protein Structures Accurately}},
  journal = {Journal of Molecular Biology},
  year = {2002},
  doi = {10.1016/s0022-2836(02)00942-7},
}
Figure