Scoring of protein–protein docking models utilizing predicted interface residues

Gabriele Pozzati, Petras J. Kundrotas, Arne ElofssonProteins 2022 · DOI

TL;DR

This paper explores using predicted protein-protein interaction interfaces to improve the scoring of protein-protein docking models. They found that the BIPSPI predictor, which considers pairs of structures, performed best, leading to a significant increase in acceptable docking models ranked first. This work highlights the potential of interface prediction to refine computational protein-protein docking.

What Problem Does This Paper Address?

Understanding how proteins interact is crucial for deciphering biological processes and developing therapeutics. Protein-protein docking is a computational method used to predict the 3D structure of these complexes. However, docking algorithms generate a vast number of potential models (decoys), and accurately identifying the correct, near-native structure among them is a major challenge. Current scoring functions, while improving, still struggle to consistently rank the correct models highly. This paper investigates whether predicting the specific residues involved in the protein-protein interface can serve as an effective scoring mechanism to filter these decoys, thereby improving the success rate of protein docking.

Methods

The researchers employed a two-stage approach. First, they generated a large number of low-resolution, rigid-body docking models using the GRAMM software. This "scan stage" produces hundreds of thousands of potential poses. Second, they re-scored these poses using a custom scoring function (Equation 1) that incorporates probabilities of residues being at the interface, as predicted by various interface prediction methods. This scoring function essentially uses predicted interface residues as constraints. They also explored an alternative docking protocol using the LzerD server, which can directly incorporate restraints. To evaluate the interface predictors themselves, they used standard metrics like True Positive Rate (TPR) and Precision (PPV), as well as the Area Under the Precision-Recall Curve (AUC).

Datasets

The primary dataset used was the DOCKGROUND benchmark set 4, specifically the unbound section, which provided 220 protein pairs with available unbound and bound experimental structures. To ensure diversity and avoid redundancy, sequences were clustered against the Benchmark 5 dataset at 20% sequence identity, resulting in 175 unique dimers. An additional testing set was derived from the CAPRI Score_set. The native interfaces were extracted from the bound structures based on changes in solvent accessible surface area (SASA).

Key Insights and Ideas

  • Interface prediction as a powerful filter: The core idea is that accurately predicting which residues form the interaction interface can significantly guide the docking process and help distinguish correct models from incorrect ones.
  • BIPSPI stands out: Among the tested interface predictors, BIPSPI, which uniquely considers pairs of structures simultaneously, demonstrated superior performance in scoring docking models.
  • Precision matters for constraints: The study found that the precision of interface predictions was a critical factor when generating docking constraints, especially for focusing the search.
  • Top 10 is better than Top 1: While some models were correctly ranked first, analyzing the top 10 models generally yielded better results, suggesting that a broader set of high-scoring predictions can be more informative.
  • Rotational ambiguity is a challenge: Even with good interface predictions, rotational freedom between the receptor and ligand can lead to incorrect orientations, highlighting a limitation of current constraint-based methods.

Key Results

  • Using BIPSPI predictions, over 12% of first-ranked docking models were acceptable (DockQ > 0.23), a significant improvement over baseline methods.
  • The success rate (SR) for acceptable models within the top 10 ranked models reached 81% when using native interface residues as constraints, with an average DockQ score of 0.45.
  • BIPSPI was consistently better than other predictors, achieving a higher success rate for first-ranked models compared to other methods.
  • Simulated interface predictions showed that both reducing precision (false positives) and reducing recall (false negatives) negatively impacted docking success, with recall reduction being more detrimental.
  • The LzerD server, when used with interface restraints, did not consistently produce acceptable docking models, and in some cases, restraints even worsened the results, suggesting that restraint geometry and quality are crucial.

Limitations and Future Work

  • Overtraining concerns: The authors noted that some interface predictors performed worse than reported in previous publications, suggesting potential overtraining on specific datasets.
  • Rotational ambiguity: A significant limitation identified is the inherent difficulty in resolving the relative orientation of receptor and ligand when using independently predicted interface constraints, leading to incorrect models even with correct interface patches.
  • Complex-wise variability: While BIPSPI performed well overall, there were specific complexes where it failed, and other predictors succeeded, indicating a need for more robust or complementary approaches.
  • Future directions: The authors suggest that the methodology can be extended to evaluate other interface or contact prediction methods. They also hint at the potential of combining different predictors to enhance cumulative docking success. Further research into more complex factors, such as the geometric relationship of restraints to the real interface, is needed.

Summary updated

Abstract

Scoring docking solutions is a difficult task, and many methods have been developed for this purpose. In docking, only a handful of the hundreds of thousands of models generated by docking algorithms are acceptable, causing difficulties when developing scoring functions. Today's best scoring functions can significantly increase the number of top-ranked models but still fail for most targets. Here, we examine the possibility of utilizing predicted interface residues to score docking models generated during the scan stage of a docking algorithm. Many methods have been developed to infer the regions of a protein surface that interact with another protein, but most have not been benchmarked using docking algorithms. This study systematically tests different interface prediction methods for scoring >300.000 low-resolution rigid-body template free docking decoys. Overall we find that contact-based interface prediction by BIPSPI is the best method to score docking solutions, with >12% of first ranked docking models being acceptable. Additional experiments indicated precision as a high-importance metric when estimating interface prediction quality, focusing on docking constraints production. Finally, we discussed several limitations for adopting interface predictions as constraints in a docking protocol.
BibTeX
@inproceedings{pozzati2022scoring,
  author = {Gabriele Pozzati and Petras J. Kundrotas and Arne Elofsson},
  title = {{Scoring of protein–protein docking models utilizing predicted interface residues}},
  booktitle = {Proteins},
  year = {2022},
  doi = {10.1002/prot.26330},
}
Figure