NEPRE: a Scoring Function for Protein Structures based on Neighbourhood Preference

Siyuan Liu, Xilun Xiang, Haiguang LiubioRxiv 2018 · DOI

TL;DR

This paper introduces NEPRE, a novel scoring function for protein structure prediction that leverages the statistical preferences of amino acid neighborhoods. By analyzing known protein structures, NEPRE quantifies how likely different amino acid types are to interact and in what relative positions, leading to improved accuracy in identifying correct protein structures from a pool of predictions.

What Problem Does This Paper Address?

Predicting the 3D structure of proteins is a fundamental challenge in biology. While methods for generating potential protein models (decoys) have advanced, accurately selecting the correct native structure from these numerous possibilities remains difficult. Experimental methods for determining protein structures are slow and costly, making computational prediction crucial, especially as new protein sequences are discovered at a rapid pace. Existing scoring functions often rely on physics-based principles or simpler statistical models. This paper addresses the need for a more refined scoring function that captures the subtle, yet important, local interactions between amino acids, which are key determinants of a protein's final folded state.

Methods

The core of NEPRE is a statistical energy function derived from analyzing the "neighborhood preference" of amino acids. This involves:

  • Defining Neighborhoods: For each amino acid in a protein, its neighbors are defined as residues within a certain distance cutoff.
  • Local Coordinate System: A local coordinate system is established for each amino acid to precisely describe the relative positions of its neighbors.
  • Statistical Analysis: A large dataset of experimentally determined protein structures (PDB) is used to count the observed frequencies of different amino acid types interacting with each other at specific relative positions.
  • Energy Function Derivation: These observed frequencies are compared to expected frequencies (assuming random interactions) to derive an energy score using Boltzmann's law. This score reflects the preference for certain amino acid pairings and orientations.
  • Two Implementations:
    • NEPRE-F: Uses a fixed distance cutoff for all amino acid types to define neighborhoods.
  • NEPRE-R: Uses distance cutoffs that are dependent on the radii of the interacting amino acid types, aiming for a more nuanced definition of neighborhood.

The algorithm was implemented and tested on various decoy sets to evaluate its performance in ranking structures.

Datasets

The study utilized several established decoy datasets for evaluation:

  • I-TASSER(a) and I-TASSER(b): Datasets generated by the I-TASSER program.
  • 3DRobot: Datasets generated by the 3DRobot program.
  • Rosetta: Datasets generated by the Rosetta program.
  • Modeller: Datasets generated by the Modeller program.
  • CASP12 Decoy Sets: Ten high-resolution structures from the Critical Assessment of protein Structure Prediction (CASP) experiments were used for further validation.

The statistical analysis for deriving the neighborhood preferences was performed on a non-redundant dataset composed of 14,647 PDB structures.

Key Insights and Ideas

  • Amino Acid Neighborhoods Matter: The paper strongly supports the idea that not only which amino acids interact but also where and how they are positioned relative to each other is crucial for protein structure.
  • Cysteine-Cysteine Preference: A striking example is the strong preference of cysteine for interacting with other cysteines, highlighting specific residue-residue interactions.
  • Orientation is Key: The authors suggest that the orientation of neighboring amino acids is more critical for packing in protein structures than just their proximity.
  • Fixed Cutoff Outperforms Type-Dependent: Surprisingly, the simpler NEPRE-F (fixed cutoff) often performed better than the more complex NEPRE-R (type-dependent cutoff), suggesting that a uniform definition of neighborhood might be more robust or that the type-dependent radii were not optimally utilized.
  • Optimal Cutoff Identified: For NEPRE-F, a cutoff of 6Å was found to be optimal for identifying native structures across multiple datasets.

Key Results

  • Improved Ranking: NEPRE demonstrated significant success in ranking protein structure decoys, with the native structure often receiving the lowest energy score.
  • High Success Rates: On various decoy sets, NEPRE achieved high success rates in identifying the native structure. For instance, with a 6Å cutoff, NEPRE-F successfully ranked the native state as the lowest energy in 220 out of 390 decoy sets.
  • CASP12 Performance: On CASP12 decoy sets, NEPRE (especially NEPRE-F with a 6Å cutoff) showed good performance, correctly identifying native structures in several cases and selecting structures with very low RMSD to the native state in others.
  • Computational Efficiency: The execution time for NEPRE is within 3-4 seconds for proteins, making it feasible for integration into larger protein modeling pipelines.

Limitations and Future Work

  • NEPRE-R Performance: The authors acknowledge that NEPRE-R's performance was surprisingly inferior to NEPRE-F, and they hypothesize that this might be due to the average radii not fully capturing complex neighboring relationships or the uniform definition of neighborhood in NEPRE-F being more effective. Further detailed analysis is needed to resolve this.
  • Residue-Level Neighborhood: The current neighborhood definition is at the residue level. Future work could explore more fine-grained atomic-level interactions.
  • Further Optimization: While NEPRE-F's cutoff was optimized, the algorithm itself was not extensively fine-tuned, suggesting potential for further improvements.
  • Integration into Modeling: The authors suggest that NEPRE's efficiency makes it suitable for integration into model generation programs to improve structure sampling.

Summary updated

Abstract

ABSTRACT Protein structure prediction relies on two major components, a method to generate good models that are close to the native structure and a scoring function that can select the good models. Based on the statistics from known structures in the protein data bank, a statistical energy function is derived to reflect the amino acid neighbourhood preferences. The neighbourhood of one amino acid is defined by its contacting residues, and the energy function is determined by the neighbhoring residue types and relative positions. A scoring algorithm, Nepre, has been implemented and its performance was tested with several decoy sets. The results show that the Nepre program can be applied in model ranking to improve the success rate in structure predictions.
BibTeX
@misc{liu2018nepre,
  author = {Siyuan Liu and Xilun Xiang and Haiguang Liu},
  title = {{NEPRE: a Scoring Function for Protein Structures based on Neighbourhood Preference}},
  year = {2018},
  doi = {10.1101/463554},
  howpublished = {bioRxiv},
}
Figure