Comparing pharmacophore models derived from crystal structures and from molecular dynamics simulations

Marcus Wieder, Ugo Perricone, Thomas Seidel, Stefan Boresch, Thierry Langer

TL;DR

This paper investigates whether using molecular dynamics (MD) simulations to refine protein-ligand complex structures improves the quality of pharmacophore models for drug discovery. They found that MD-refined models can indeed offer better discrimination between active and decoy compounds compared to models built directly from crystal structures, suggesting MD simulations are a valuable refinement step.

What Problem Does This Paper Address?

Structure-based pharmacophore modeling is a powerful tool in drug discovery, helping to identify potential drug candidates by representing the essential features of a ligand that interact with a protein target. However, these models are typically derived from static crystal structures. There are concerns that these crystal structures might not perfectly represent the dynamic interactions occurring in vivo, due to factors like ligand conformation, non-physiological contacts, and solvent effects. This can lead to pharmacophore models that are not entirely accurate or predictive. The authors aim to address this by exploring whether using the final structure from a molecular dynamics (MD) simulation of a protein-ligand complex can generate more robust pharmacophore models.

Methods

The core of this study involves comparing two types of pharmacophore models:

  1. Initial Pharmacophore Models: These were generated directly from protein-ligand complex structures obtained from the Protein Data Bank (PDB).
  2. MD-Refined Pharmacophore Models: These were generated from the final structure of a 20 ns molecular dynamics (MD) simulation of the same protein-ligand complexes.

The researchers used the LigandScout software for generating the pharmacophore models and performing virtual screening. To evaluate the models, they employed two key metrics:

  • Receiver Operating Characteristic (ROC) curves: These visualize the trade-off between the true positive rate (correctly identifying actives) and the false positive rate (incorrectly identifying decoys) at various thresholds.
  • Enrichment Factor (EF): This quantifies how much better a model is at finding active compounds compared to random screening.

They also performed MD simulations using the CHARMM package, including system setup, parameterization, and the simulation itself. RMSD (Root Mean Square Deviation) calculations were used to assess the stability of the protein and ligand during the MD simulations.

Datasets

The study utilized six protein-ligand systems from the DUD-E (Directory of Useful Decoys, Enhanced) database. This database is specifically designed for benchmarking virtual screening methods, providing known active compounds and computationally generated decoys that share similar physicochemical properties but differ in topology. The PDB codes for the systems used were: 1J4H, 3BQD, 2HZI, 3L3M, 1UYG, and 3EL8. The number of actives and decoys for each system is detailed in Table 1.

Key Insights and Ideas

  • MD Simulations Reveal Structural Differences: The study clearly demonstrates that the final structures from MD simulations often differ from the initial crystal structures, both in terms of ligand conformation and protein-ligand interactions. This highlights the dynamic nature of these complexes.
  • Pharmacophore Models Change with MD Refinement: The pharmacophore models derived from MD-refined structures were not identical to those from crystal structures. They often differed in the number and types of features (e.g., hydrogen bond acceptors/donors, hydrophobic regions). This suggests that MD captures different interaction nuances.
  • MD Refinement Can Improve Predictive Power: In several cases, the MD-refined pharmacophore models showed a better ability to distinguish between active and decoy compounds in virtual screening. This is a crucial finding, as it implies that incorporating dynamic information can lead to more effective drug discovery tools.
  • No Simple Correlation with RMSD: Interestingly, the researchers found no clear correlation between the RMSD values (indicating structural stability or deviation) of the protein or ligand and the performance of the resulting pharmacophore models. This suggests that simply looking at RMSD alone isn't enough to predict model quality.

Key Results

  • Feature Differences: For all six analyzed systems, the pharmacophore models differed between the initial crystal structure and the MD-refined structure. The number of features often changed, with some increasing and others decreasing.
  • Improved Virtual Screening Performance: For several systems (e.g., 1UYG, 2HZI, 3BQD, 3EL8), the MD-refined pharmacophore models demonstrated improved performance in virtual screening, as evidenced by better ROC curves and higher enrichment factors, particularly when considering early enrichment (top-ranked compounds).
  • Varied Impact: While MD refinement was beneficial in many cases, it did not universally improve performance for all systems. For some, the initial models performed comparably or even slightly better, suggesting that the benefit of MD refinement can be system-dependent.
  • No Universal Guideline: The study concluded that it was not possible to establish universal guidelines linking MD refinement performance to specific protein/ligand structural characteristics or RMSD values.

Limitations and Future Work

  • Limited Number of Systems: The study analyzed six protein-ligand systems. While diverse, this is a relatively small sample size, and the findings might not generalize to all protein-ligand interactions.
  • "Simple" MD Refinement: The authors used a basic MD refinement protocol (using the final frame). More sophisticated MD-based approaches, such as ensemble averaging or free energy calculations, might yield different or even better results.
  • No Manual Refinement: The virtual screening was performed with "default" pharmacophore models without manual refinement. The authors acknowledge that manual refinement could further bias comparisons.
  • DUD-E Database Nuances: The authors note that the DUD-E database, while valuable, may contain decoys that still bind to the protein, and decoys can sometimes be structurally similar to actives, which can complicate evaluation.

The authors suggest that additional work is needed to find more specific guidelines for optimizing MD simulations in the context of pharmacophore modeling. They also highlight that MD simulations can reveal crucial interaction information that can lead to better pharmacophore models, potentially targeting specific binding sites or transitional conformations.

Summary updated

Abstract

Pharmacophore modeling is a widely used technique in computer-aided drug discovery. Structure-based pharmacophore models of a ligand in complex with a protein have proven to be useful for supporting in silico hit discovery, hit to lead expansion, and lead optimization. As a structure-based approach it depends on the correct interpretation of ligand-protein interactions. There are legitimate concerns about the fidelity of the bound ligand and about non-physiological contacts with parts of the crystal and the solvent effects that influence the protein structure. A possible way to refine the structure of a protein-ligand system is to use the final structure of a given MD simulation. In this study we compare pharmacophore models built using the initial protein-ligand structure obtained from the protein data bank (PDB) with pharmacophore models built with the final structure of a molecular dynamics simulation. We show that the pharmacophore models differ in feature number and feature type and that the pharmacophore models built from the last structure of a MD simulation shows in some cases better ability to distinguish between active and decoy ligand structures.
BibTeX
@inproceedings{wieder2025comparing,
  author = {Marcus Wieder and Ugo Perricone and Thomas Seidel and Stefan Boresch and Thierry Langer},
  title = {{Comparing pharmacophore models derived from crystal structures and from molecular dynamics simulations}},
  booktitle = {Monatshefte für Chemie},
  year = {2025},
  url = {https://openreview.net/forum?id=DO6ShOPMhc},
}
Figure