Scoring predictive models using a reduced representation of proteins: model and energy definition

Federico Fogolari, L. Pieri, Agostino Dovier, Luca Bortolussi, G. Giugliarelli, Alessandra Corazza, Gennaro Esposito, Paolo ViglinoBMC 2007 · DOI

TL;DR

This paper introduces PC2CA, a novel, simplified protein representation using only two interaction centers per amino acid, coupled with a knowledge-based energy function. This approach effectively scores protein models, identifying native-like structures with high accuracy across various benchmark datasets and a blind CASP7 quality assessment, demonstrating its potential for rapid conformational space exploration.

What Problem Does This Paper Address?

Predicting the 3D structure of proteins from their amino acid sequence is a fundamental challenge in biology. While many computational methods exist, accurately assessing the quality of these predicted structures is crucial. Existing scoring functions, which estimate the energy of a protein conformation, often rely on complex, all-atom representations of proteins. This can be computationally expensive and may not always perform optimally in distinguishing native-like structures from a vast sea of incorrect "decoys." The authors aim to develop a simpler, yet effective, scoring function by employing a reduced protein representation and a corresponding knowledge-based energy model, making it more computationally tractable for large-scale applications like protein structure prediction and quality assessment.

Methods

The core of this work lies in the development of a reduced protein representation called PC2CA (PseudoCovalent structure with 2 Centers of interaction per Amino acid). In this model, each amino acid is represented by two interaction centers: one for the backbone (at the Cα atom) and one for the sidechain (at the center of mass of sidechain atoms, except for glycine which has only one backbone center).

This simplified representation is then used to derive a knowledge-based statistical effective energy function. The energy function is built upon the Boltzmann principle, where the probability of a given conformation is related to its energy. The authors extract "potentials of mean force" from a large dataset of experimentally determined protein structures (Top500H). These potentials capture the preferred interactions (e.g., distances, angles, torsions) between the interaction centers in the PC2CA model. The energy function includes terms for:

  • Bonded interactions: Pseudo-bonds between consecutive Cα atoms and between Cα and sidechain centers.
  • Angular interactions: Angles involving these centers.
  • Torsional interactions: Pseudo-torsional energies to maintain chirality and local conformational preferences.
  • Non-bonded interactions: Pairwise interactions between centers, accounting for excluded volume and finite protein size.

A key innovation is the inclusion of a term that captures correlations between adjacent local conformations, aiming to better reproduce secondary structure elements. The derivation of these potentials involves discretizing the relevant variables (distances, angles) into bins and calculating the frequency of occurrence in the reference dataset.

Datasets

  • Training/Reference Dataset: The Top500H dataset, a non-redundant set of 500 high-resolution X-ray crystallography structures, was used to derive the statistical potentials.
  • Evaluation Datasets:
    • Decoys'R'us database: This database contains multiple decoy sets, each representing a single native structure with numerous alternative conformations generated by various methods. The paper specifically mentions testing on sets like 4state_reduced, fisa, fisa_casp3, hg_structal, ig_structal, ig_structal_hires, lattice_ssfit, Imds, semfold, and vhp_mcmd. These datasets vary in size and the methods used to generate decoys, providing a comprehensive testbed.
  • CASP7 Quality Assessment: The authors also evaluated their model in the blind CASP7 quality assessment category, a community-wide experiment for assessing protein structure prediction models.

Key Insights and Ideas

  • Simplicity Can Be Powerful: The PC2CA representation, with just two interaction centers per residue, is significantly simpler than all-atom models. The fact that it can achieve competitive scoring performance highlights that a high level of atomic detail isn't always necessary for effective energy evaluation.
  • Local Correlations Matter: The inclusion of a term for correlations between adjacent local conformations proved crucial. This term significantly improved the model's ability to reproduce local conformational propensities, a finding confirmed by a test where this term was zeroed out.
  • Reference State Choice is Critical: The paper emphasizes the importance of defining an appropriate reference state for deriving statistical potentials, particularly one that accounts for the finite size of proteins and the varying nature of amino acids.
  • Coarse-Grained Models Can Compete: The PC2CA model, despite its coarse-grained nature, performs comparably to or better than many more complex models in distinguishing native-like structures from decoys, as evidenced by its performance in CASP7.

Key Results

  • High Accuracy in Decoy Recognition: The PC2CA energy function successfully identified native-like structures (RMSD < 2.0 Å) in almost all tested decoy sets from the Decoys'R'us database.
  • CASP7 Performance: In the blind CASP7 quality assessment experiment, the PC2CA model demonstrated strong performance, comparing favorably with other established quality assessment programs. The average loss in GDT_TS (a measure of structural similarity) was 10.1, which is lower than the average loss of 10.7 for all quality assessment methods tested.
  • Correlation with RMSD: For many decoy sets, a good correlation between the calculated energy and the RMSD from the native structure was observed, indicating the energy function's ability to rank structures appropriately.
  • Importance of Correlation Term: Removing the correlation term between adjacent local conformations reduced the accuracy of predicting secondary structure from 0.57 to 0.51, underscoring its significance.

Limitations and Future Work

  • Pseudo-Minimization: The paper acknowledges that the "pseudo-minimization" used to relax conformations is a rough approximation of real energy minimization. This might lead to some dependence on the simplification procedure and building details.
  • Model Generation vs. Scoring: While the potential is effective for scoring, its direct use for generating protein conformations (e.g., via Monte Carlo simulations) might require adjustments. The authors suggest that the potential could have lower minima for conformations not explored by current decoy generation algorithms.
  • Potential for Improvement in Model Generation: For applications in model generation, the authors suggest modifying the potential to prevent bond breakage (e.g., by replacing flat high-energy regions with increasing potentials) and potentially adjusting weights of different terms.
  • Future Work: The authors are developing applications for model generation using this potential and are exploring its use for fast exploration of conformational space. They also note that the correlation found between torsional and bending energy terms in native structures might need further treatment.

Summary updated

Abstract

The model and the energy definition proposed have been tested on all the multiple decoys' sets in the Decoys'R'us database. The energetic model is able to recognize, for almost all sets, native-like structures (RMSD less than 2.0 A). These results and those obtained in the blind CASP7 quality assessment experiment suggest that the model compares well with scoring potentials with finer granularity and could be useful for fast exploration of conformational space. Parameters are available at the url: http://www.dstb.uniud.it/~ffogolari/download/.
BibTeX
@inproceedings{fogolari2007scoring,
  author = {Federico Fogolari and L. Pieri and Agostino Dovier and Luca Bortolussi and G. Giugliarelli and Alessandra Corazza and Gennaro Esposito and Paolo Viglino},
  title = {{Scoring predictive models using a reduced representation of proteins: model and energy definition}},
  booktitle = {BMC},
  year = {2007},
  doi = {10.1186/1472-6807-7-15},
}
Figure