Comparative modelling by restraint-based conformational sampling

Nicholas Furnham, Paul Iw de Bakker, Swanand Gore, David F. Burke, Tom L. BlundellBMC 2008 · DOI

TL;DR

This paper introduces a new application of the RAPPER software for comparative protein modeling, which generates more accurate and representative structural models by incorporating experimental and knowledge-based restraints. By treating protein structures as ensembles rather than single conformers, this approach offers a more realistic representation of protein flexibility and can inform modeling efforts with sparse experimental data.

What Problem Does This Paper Address?

Determining the three-dimensional structure of proteins is crucial for understanding their function, but experimental methods like X-ray crystallography and NMR have limitations. While comparative modeling, which uses known protein structures (templates) to predict the structure of a related protein (target), is a powerful computational approach, existing methods often struggle to incorporate diverse sources of information. Specifically, there's a need for automated methods that can effectively integrate experimental data (like X-ray electron density) and knowledge derived from homologous structures. Furthermore, proteins are inherently dynamic and exist as ensembles of conformations, yet most modeling approaches produce single, static models, which can be an oversimplification. This paper aims to address these limitations by developing a restraint-based conformational sampling method that can leverage multiple sources of structural information and generate ensembles of models.

Methods

The core of this work is the adaptation of the RAPPER (restraint-based conformational space search engine) for comparative modeling. RAPPER works by defining spatial restraints that guide the conformational sampling process. The key innovations and approaches include:

  • Knowledge-Based Restraints from Homologues: Instead of relying solely on direct structural alignment, RAPPER derives restraints from the geometrical features and main-chain/side-chain conformations of homologous templates. This includes using information about main-chain curvature, torsion, and side-chain angle conservation.
  • Combinatorial Template Use: The study explores using various combinations of homologous templates (from one to four) to build models, allowing for an assessment of how template diversity and quality impact model accuracy.
  • Integration of Advanced Restraint Definitions:
    • CHORAL/ANDANTE Predictions: The paper integrates predictions from CHORAL (for main-chain curvature and torsion) and ANDANTE (for side-chain rotamers) to create more refined and accurate restraints. This helps to reduce the influence of inaccurate template information.
  • Probability Density Functions (PDFs): A novel approach uses PDFs derived from homologous structures, weighted by sequence identity, to define restraints. This allows for the exploitation of more distantly related homologous structures.
  • Ensemble Generation: RAPPER generates an ensemble of 50 models, which are then averaged to produce a representative model. This ensemble approach captures protein flexibility and uncertainty, providing a more realistic representation than a single model.
  • Comparison with MODELLER: The performance of RAPPER is benchmarked against MODELLER, a widely used comparative modeling program, using various metrics like all-atom RMSD.

Datasets

The study utilized a dataset of twenty-four protein families chosen from the HOMSTRAD database. These families were selected to represent the four main SCOP classes (all α, all β, α+β, and α/β). For each family, five members were chosen based on maximizing the range of relative percentage identity (PID) and ensuring high-resolution structures (greater than 2Å). One member was designated as the target, and the remaining four served as templates.

Key Insights and Ideas

  • Ensembles are More Informative: Generating an ensemble of models, rather than a single structure, provides a more appropriate representation of a structural prediction. This ensemble captures conformational variability and uncertainty, which is crucial for understanding protein dynamics and function.
  • Restraint Quality is Paramount: The accuracy of comparative models is highly dependent on the quality and type of restraints used. Incorporating knowledge-based restraints derived from homologous structures, especially when enhanced by predictive algorithms like CHORAL and ANDANTE, significantly improves model accuracy.
  • Leveraging Distant Homologues: By formulating homologous structural information as probability density functions (PDFs), the method can effectively exploit information from more distantly related homologues, expanding the applicability of comparative modeling.
  • Addressing Local Inaccuracies: The study highlights how specific regions, like C-terminal loops, can contribute disproportionately to the overall RMSD, often due to factors like crystal packing. The ensemble approach can help to better represent these variable regions.

Key Results

  • Improved Accuracy: The application of RAPPER with optimized restraint definitions (especially using CHORAL/ANDANTE predictions and PDF-derived restraints) achieved accuracies comparable to leading modeling tools like MODELLER.
  • Ensemble Comparison: The RAPPER-generated ensemble for a-parvalbumin showed similar compactness and diversity to the experimentally determined NMR ensemble, suggesting that comparative models can capture relevant conformational features.
  • Quantitative Improvements: While specific numerical improvements vary across different families and methods, the paper presents tables (e.g., Table 4, Table 5) showing all-atom RMSD values that demonstrate the relative performance of different RAPPER configurations and MODELLER. For instance, RAPPER using CHORAL/ANDANTE predictions generally showed lower RMSD values compared to simpler RAPPER approaches and sometimes even MODELLER.
  • Identification of Limitations: The study identified specific cases (e.g., Phospholipase A2, Response Regulator Receiver domain) where RAPPER's default behavior led to inaccuracies due to insertions or incorrect orientation of secondary structures, suggesting areas for future improvement.

Limitations and Future Work

  • Inaccurate Alignments: The paper acknowledges that errors in sequence alignment can lead to inaccurate restraints and thus affect model quality. While they used structure-based alignments from HOMSTRAD, this remains a potential source of error.
  • Restraint Definition Challenges: For regions with low sequence identity or significant structural divergence, deriving accurate restraints from homologues remains challenging. The paper suggests that incorporating secondary structure predictions could help mitigate this.
  • Future Directions: The authors propose exploring the use of RAPPER for modeling other macromolecules and complexes, as well as its application in conjunction with sparse experimental data (e.g., low-resolution X-ray data, NMR, SAXS) to inform conformer generation. They also mention ongoing work on RAPPER-TK for modeling larger systems.

Summary updated

Abstract

The application of RAPPER to comparative modelling provides an effective means of exploring the conformational space available to a target sequence. Enhanced methods for generating positional restraints can greatly improve structure prediction. Generation of an ensemble of solutions that are consistent with both target sequence and knowledge derived from the template structures provides a more appropriate representation of a structural prediction than a single model. By formulating homologous structural information as sets of restraints we can begin to consider how comparative models might be used to inform conformer generation from sparse experimental data.
BibTeX
@inproceedings{furnham2008comparative,
  author = {Nicholas Furnham and Paul IW de Bakker and Swanand Gore and David F. Burke and Tom L. Blundell},
  title = {{Comparative modelling by restraint-based conformational sampling}},
  booktitle = {BMC},
  year = {2008},
  doi = {10.1186/1472-6807-8-7},
}
Figure