Prediction of Binding Sites of Peptide Recognition Domains: An Application on Grb2 and SAP SH2 Domains

William A. McLaughlin, Tingjun Hou, Wei WangJournal of Molecular Biology 2006 · DOI

TL;DR

This paper introduces a novel computational method that combines binding free energy estimations and sequence analysis to predict protein-protein interaction sites, specifically focusing on SH2 domains. By integrating these two approaches, the method successfully identifies known binding motifs and predicts new interaction partners, offering a powerful alternative to traditional experimental methods.

What Problem Does This Paper Address?

Understanding how proteins interact is fundamental to deciphering cellular signaling pathways and the mechanisms behind various diseases. Modular protein domains, like the Src homology 2 (SH2) domain, play a crucial role as "molecular matchmakers," mediating protein-protein interactions in response to extracellular signals. SH2 domains, in particular, are known to bind to phosphorylated tyrosine residues on their interaction partners.

However, accurately identifying these binding sites and predicting interaction partners has been a significant challenge. Traditional methods, such as peptide library screening, can be time-consuming and may introduce biases due to the artificial nature of the libraries. While computational methods exist, they often rely on either binding free energy estimations or sequence motif analysis alone. This paper addresses the need for a more robust and comprehensive approach by integrating both these orthogonal pieces of information. The goal is to improve the accuracy and scope of predicting SH2 domain binding sites and, by extension, understanding their roles in cellular processes.

Methods

The authors developed a computational method that integrates two key approaches:

  1. Binding Free Energy Estimation: They used the molecular mechanics/Poisson-Boltzmann solvent-accessible surface area (MM/PBSA) method to estimate the binding free energy of peptides to SH2 domains. This method accounts for various energetic contributions, including solvation effects. To improve discrimination between binders and non-binders, they focused on unphosphorylated peptides, as the contribution of the phosphate group to binding energy was found to be similar for both known binders and background peptides, and its calculation had a relatively high error.
  2. Sequence Motif Analysis: After estimating binding energies, peptides were clustered based on both their sequence characteristics and their estimated binding energies. This combined approach, referred to as "sequence-energy clustering," was found to be more effective than using either sequence or energy alone. From the cluster containing known binders, a sequence motif was extracted and represented as a Hidden Markov Model (HMM).

These HMMs were then used to scan large human protein sequence databases (SWISS-PROT) to identify potential interaction partners and their binding sites. A conservation filter was also applied to the top-scoring candidates to increase the confidence in predicted interactions.

Datasets

The study utilized two primary datasets, each focusing on a specific SH2 domain:

  • Grb2 SH2 Domain Dataset: This dataset consisted of 1415 peptides, including 15 known binders and 1400 randomly selected peptides from the human proteome.
  • SAP SH2 Domain Dataset: This dataset comprised 1810 peptides, including 11 known binders, 1400 randomly selected peptides, and an additional 399 conserved peptides derived from proteins related in sequence to SLAM, a known SAP SH2 domain interactor.

These datasets were generated by extracting peptides of specific lengths and with tyrosine residues at defined positions from solved protein-peptide complex structures available in the Protein Data Bank.

Key Insights and Ideas

  • Synergy of Energy and Sequence: The core insight is that combining binding free energy estimations with sequence analysis provides a more powerful predictive tool than either method alone. This synergy helps to overcome the limitations of each individual approach.
  • Unphosphorylated Peptides for Discrimination: Focusing on unphosphorylated peptides for energy calculations improved the ability to distinguish between known binders and random peptides, simplifying the downstream analysis.
  • Sequence-Energy Clustering is Key: The "sequence-energy clustering" approach was crucial for grouping peptides with similar binding characteristics, leading to the extraction of more accurate and generalizable binding motifs.
  • HMMs as Powerful Scanners: Hidden Markov Models (HMMs) derived from these clusters proved effective in scanning large proteomes to identify potential interaction partners, even those with sequences not identical to the initial known binders.
  • Predicting Novel Roles: By identifying novel interaction partners, the study was able to suggest new potential biological roles for the Grb2 and SAP SH2 domains, expanding our understanding beyond experimentally confirmed functions.

Key Results

  • Improved Discrimination: The MM/PBSA method, when applied to unphosphorylated peptides, showed a statistically significant difference in binding energies between known binders and random peptides for both Grb2 and SAP SH2 domains (p-values as low as 2.31x10⁻⁹ for Grb2 and 7.37x10⁻⁶ for SAP).
  • Effective Clustering: The sequence-energy clustering scheme yielded the highest overlap of known binders within a designated "binding cluster" for both datasets. For SAP, cluster 5 contained the majority of known binders, with a mean binding energy of -44.55 ± 5.33 kcal/mol. For Grb2, cluster 4 was the binding cluster, with a mean energy of -16.86 ± 6.31 kcal/mol.
  • Accurate Motif Reconstruction: The HMMs derived from the binding clusters successfully reconstructed experimentally determined binding motifs for both Grb2 and SAP SH2 domains, demonstrating the method's validity.
  • Successful Database Searches: Database searches using the binding cluster HMMs placed known binders at significantly higher ranks compared to control HMMs (p-values of 2.78x10⁻⁶ for SAP and 1.62x10⁻¹ for Grb2), indicating that the binding cluster HMMs captured relevant binding information beyond just the known sequences.
  • Prediction of Novel Interactions: The method predicted several novel interaction candidates, including a potential binding site on the UFO protein for the Grb2 SH2 domain, which had been experimentally validated but missed in previous surveys.

Limitations and Future Work

  • Conformational Entropy: The study acknowledges that conformational entropy was not included in the binding energy calculations due to computational cost and previous findings suggesting it might not significantly improve discrimination in similar systems. This is a potential area for future refinement.
  • Computational Cost: While the method is more efficient than full genomic free energy calculations, estimating binding energies for thousands of peptides is still computationally intensive.
  • Further Refinement of Methods: The authors suggest that more accurate binding affinity calculations and more efficient sampling methods could further improve the approach.
  • Incorporating Additional Information: Future work could benefit from incorporating information about protein interactions in vivo, such as cellular compartmentalization, and the ability of kinases to bind substrates prior to SH2 domain binding.
  • Broader Application: The authors anticipate that with advancements in structural genomics and homology modeling, this method can be applied to a wider range of protein-peptide interactions beyond SH2 domains.

Summary updated

BibTeX
@article{mclaughlin2006prediction,
  author = {William A. McLaughlin and Tingjun Hou and Wei Wang},
  title = {{Prediction of Binding Sites of Peptide Recognition Domains: An Application on Grb2 and SAP SH2 Domains}},
  journal = {Journal of Molecular Biology},
  year = {2006},
  doi = {10.1016/j.jmb.2006.01.005},
}
Figure