Hierarchical Graph Representation of Pharmacophore Models

Arthur Garon, Oliver Wieder, Bareis Klaus, Thomas Seidel, Gökhan Ibis, Sharon D. Bryant, Isabelle Théret, Ducrot Pierre, Langer ThierryFrontiers in Molecular Biosciences 2020 · DOI

Abstract

For the investigation of protein-ligand interaction patterns, the current accessibility of a wide variety of sampling methods allows quick access to large-scale data. The main example is the intensive use of molecular dynamics simulations applied to crystallographic structures which provide dynamic information on the binding interactions in protein-ligand complexes. Chemical feature interaction based pharmacophore models extracted from these simulations, were recently used with consensus scoring approaches to identify potentially active molecules. While this approach is rapid and can be fully automated for virtual screening, additional relevant information from such simulations is still opaque and so far the full potential has not been entirely exploited. To address these aspects, we developed the hierarchical graph representation of pharmacophore models (HGPM). This single graph representation enables an intuitive observation of numerous pharmacophore models from long MD trajectories and further emphasizes their relationship and feature hierarchy. The resulting interactive depiction provides an easy-to-apprehend tool for the selection of sets of pharmacophores as well as visual support for analysis of pharmacophore feature composition and virtual screening results. Furthermore, the representation can be adapted to include information involving interactions between the same protein and multiple different ligands. Herein, we describe the generation, visualization and use of HGPMs generated from MD simulations of two x-ray crystallographic derived structures of the human glucokinase protein in complex with allosteric activators. The results demonstrate that a large number of pharmacophores and their relationships can be visualized in an interactive, efficient manner, unique binding modes identified and a combination of models derived from long MD simulations can be strategically prioritized for VS campaigns.

Introduction

In modern drug discovery, researchers utilize pharmacophores—abstract models of essential chemical features, such as hydrogen bonds or hydrophobic pockets, required for a drug to bind to a protein. Historically, these models were constructed from static "snapshots" of crystal structures. However, modern Molecular Dynamics (MD) simulations now allow us to capture the dynamic nature of protein-ligand interactions over time.

The challenge lies in the sheer volume of data: a single MD simulation can generate tens of thousands of potential pharmacophore models. Selecting the optimal models for virtual screening (VS) is a significant bottleneck, often requiring complex clustering or extensive training data that may not be available for novel targets. To address this, we introduce the Hierarchical Graph Representation of Pharmacophore Models (HGPM), a visualization tool that organizes thousands of simulation snapshots into an intuitive hierarchy, allowing researchers to identify unique binding modes and prioritize models for screening.

Methodology

The HGPM framework treats pharmacophore models as nodes within a graph, structured through a multi-step process:

  1. Feature Vectorization: Every unique interaction (e.g., a hydrogen bond with a specific residue) is assigned a position in a bit-string. Each pharmacophore snapshot is thus converted into a "feature vector" of 1s and 0s.
  2. Filtering and Node Creation: Snapshots with identical feature vectors are grouped into nodes. Infrequent patterns are filtered out to reduce noise.
Workflow for converting pharmacophore snapshots into feature vectors and filtered nodes.
  1. Hierarchical Linkage: Nodes are connected based on subset/superset relationships. If Model A contains all features of Model B plus one additional feature, Model A is designated as a "child" of Model B.
  2. Visualization Mapping: The graph is plotted with the X-axis representing specificity (the number of features) and the Y-axis using Multidimensional Scaling (MDS) to place similar models closer together. Node size indicates the frequency of that specific interaction pattern during the simulation.
The Hierarchical Graph Representation (HGPM) showing specificity vs. relative distance.

Validation and Results

The approach was validated using the human glucokinase (GK) protein, a critical target for diabetes. Simulations were initiated from two X-ray crystal structures (PDB IDs: 1v4s and 4no7), with three 100-nanosecond replicates each. The models were tested against a library of 20,756 molecules from ChEMBL, including known actives and decoys.

Key results include:

  • High Accuracy: Models prioritized via HGPM achieved Area Under the Curve (AUC) values exceeding 0.96 at a 1% threshold.
  • Efficiency: The hierarchical selection reduced the number of models required for consensus screening by 20 to 45 times compared to traditional methods without sacrificing accuracy.
  • Binding Mode Discovery: In the 4no7 structure, the HGPM revealed two distinct branches, indicating the ligand could bind in two structurally different ways—a detail often lost in standard cluster analysis.

Discussion and Future Directions

The HGPM naturally separates "core" interactions from transient, specialized ones, providing a clear strategy for virtual screening. Furthermore, the system is ligand-agnostic; by projecting interactions from different drugs onto the same graph, researchers can directly compare how different chemical series occupy the same protein pocket.

While highly effective, the method currently relies on manual noise filtering thresholds and focuses on residue contacts rather than precise 3D geometric angles. Future work will look toward integrating this tool into automated discovery pipelines, providing a "human-in-the-loop" visual verification for AI-driven drug design.

Summary updated

Abstract

For the investigation of protein-ligand interaction patterns, the current accessibility of a wide variety of sampling methods allows quick access to large-scale data. The main example is the intensive use of molecular dynamics simulations applied to crystallographic structures which provide dynamic information on the binding interactions in protein-ligand complexes. Chemical feature interaction based pharmacophore models extracted from these simulations, were recently used with consensus scoring approaches to identify potentially active molecules. While this approach is rapid and can be fully automated for virtual screening, additional relevant information from such simulations is still opaque and so far the full potential has not been entirely exploited. To address these aspects, we developed the hierarchical graph representation of pharmacophore models (HGPM). This single graph representation enables an intuitive observation of numerous pharmacophore models from long MD trajectories and further emphasizes their relationship and feature hierarchy. The resulting interactive depiction provides an easy-to-apprehend tool for the selection of sets of pharmacophores as well as visual support for analysis of pharmacophore feature composition and virtual screening results. Furthermore, the representation can be adapted to include information involving interactions between the same protein and multiple different ligands. Herein, we describe the generation, visualization and use of HGPMs generated from MD simulations of two x-ray crystallographic derived structures of the human glucokinase protein in complex with allosteric activators. The results demonstrate that a large number of pharmacophores and their relationships can be visualized in an interactive, efficient manner, unique binding modes identified and a combination of models derived from long MD simulations can be strategically prioritized for VS campaigns.
BibTeX
@inproceedings{garon2020hierarchical,
  author = {Arthur Garon and Oliver Wieder and Bareis Klaus and Thomas Seidel and Gökhan Ibis and Sharon D. Bryant and Isabelle Théret and Ducrot Pierre and Langer Thierry},
  title = {{Hierarchical Graph Representation of Pharmacophore Models}},
  booktitle = {Frontiers in Molecular Biosciences},
  year = {2020},
  doi = {10.3389/fmolb.2020.599059},
}
Figure