Apo2ph4: A Versatile Workflow for the Generation of Receptor-based Pharmacophore Models for Virtual Screening

Jörg Heider, Jonas Kilian, Aleksandra Garifulina, Steffen Hering, Thierry Langer, Thomas SeidelJournal of Chemical Information and Modeling 2023 · DOI

TL;DR

This paper introduces apo2ph4, a novel computational workflow that generates receptor-based pharmacophore models for virtual screening without needing prior knowledge of active ligands. This addresses a major bottleneck in drug discovery by enabling rapid screening against new targets, demonstrating significant success in identifying potential drug candidates.

What Problem Does This Paper Address?

Drug discovery often relies on virtual screening (VS) to sift through vast libraries of chemical compounds to find potential drug candidates. A key tool in VS is the pharmacophore model, which represents the essential 3D features a molecule needs to interact with a biological target. Traditionally, generating high-quality pharmacophore models requires either the 3D structure of a ligand bound to its target (structure-based) or knowledge of at least one active ligand (ligand-based).

However, obtaining these prerequisites can be a significant hurdle, especially for novel targets where no active ligands have been identified. This leaves researchers with limited options, often resorting to less efficient or less accurate methods. The paper tackles this challenge by proposing a method to generate reliable pharmacophore models solely from the 3D structure of the unbound (apo) receptor.

Methods

The apo2ph4 workflow consists of four main steps:

  1. Binding Site Selection: Identifying potential binding pockets within the apo-protein structure.
  2. Fragment Docking: Docking a diverse set of small, drug-like fragments (from the Prestwick Chemical library) into the identified binding site. This step explores various potential interactions within the pocket.
  3. SB Pharmacophore Generation: For each docked fragment pose, a structure-based (SB) pharmacophore model is generated. This captures the interactions between the fragment and the surrounding protein residues.
  4. Scoring, Clustering, and Filtering: The generated pharmacophore models are then clustered and scored based on feature density (how many fragments suggest similar features in similar locations). This process consolidates the information from multiple fragment docks into a single, high-quality pharmacophore model suitable for virtual screening.

The workflow leverages existing tools like AutoDock Vina for docking and LigandScout for pharmacophore generation, but integrates them into a novel pipeline with custom scripts for feature scoring and clustering.

Datasets

  • Prestwick Chemical Fragment Library: A diverse set of 1456 lead-like fragments, with a subset of 200 molecules used for docking to balance computational cost and diversity.
  • LIT-PCBA Dataset: This dataset, specifically designed to be unbiased for benchmarking VS techniques, was used to evaluate apo2ph4's performance across 15 different targets. It includes true inactive compounds, differentiating it from datasets that rely solely on computer-generated decoys.
  • GABAĀ Receptor Dataset: A cryo-EM structure of the GABAA receptor in complex with etomidate was used for a specific case study.
  • MolPort Compound Library: A large commercial library of approximately 7 million compounds was used for virtual screening against the GABAA receptor.

Key Insights and Ideas

  • Ligand-Independent Pharmacophore Generation: The core innovation is demonstrating that high-quality pharmacophore models can be generated without any prior knowledge of active ligands, solely from the receptor's apo structure. This significantly broadens the applicability of pharmacophore modeling.
  • Fragment Docking as a Feature Generator: Using fragment docking as a way to probe the binding site and generate diverse interaction hypotheses is a clever approach. It leverages the chemical diversity of fragments to explore the pocket's features.
  • Feature Density for Robustness: Scoring and clustering pharmacophore features based on their density across multiple fragment poses provides a robust way to identify the most likely and important interaction points, filtering out noise.
  • Generalizability Across Targets: The successful application of apo2ph4 to the diverse LIT-PCBA dataset highlights its versatility and applicability to a wide range of protein targets.

Key Results

  • LIT-PCBA Dataset Performance: apo2ph4 achieved significant enrichment of active compounds for 13 out of 15 targets in the LIT-PCBA dataset, with enrichment factors ranging from 1.7 to 24.2. This demonstrates its effectiveness in identifying promising hits.
  • GABAĀ Receptor Case Study: A pharmacophore model generated by apo2ph4 for the GABAA receptor was used to screen a large compound library. After rigorous filtering, 20 compounds were selected for in vitro testing, and an impressive 19 out of 20 (95%) were found to significantly potentiate GABA currents. This is a strong validation of the workflow's predictive power.
  • Comparison to SB Models: For the M2 muscarinic acetylcholine receptor, apo2ph4-generated pharmacophores showed good agreement with those derived from a known ligand-bound structure, with key features matching closely.

Limitations and Future Work

  • Metal Ion Coordination: The current apo2ph4 workflow cannot automatically account for coordinating metal ions in binding sites, requiring manual intervention for such cases.
  • HBD Feature Discarding: The authors note difficulty in finding a robust method to discard "unnecessary" hydrogen bond donor (HBD) features, which can sometimes be generated too frequently.
  • Future Applications: The authors are exploring the use of apo2ph4 for generating "toxicophores" to predict potential off-target effects and neurotoxicity of drug candidates. They also mention ongoing development of novel ligands for previously unknown binding pockets identified using this workflow.

Summary updated

BibTeX
@article{heider2023apo,
  author = {Jörg Heider and Jonas Kilian and Aleksandra Garifulina and Steffen Hering and Thierry Langer and Thomas Seidel},
  title = {{Apo2ph4: A Versatile Workflow for the Generation of Receptor-based Pharmacophore Models for Virtual Screening}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2023},
  doi = {10.1021/acs.jcim.2c00814},
}
Figure