CHARMM-GUI Covalent Ligand Docker as a Web-based Molecular Docking Platform for Covalent Ligands

Technical Summary: CHARMM-GUI Covalent Ligand Docker (CGUI-CLD)

Problem Statement
Covalent inhibitors offer distinct advantages in drug discovery, including prolonged binding and reduced resistance, yet their computational study faces significant hurdles. While molecular docking is a standard tool for predicting ligand poses, covalent docking is complicated by the structural transformation of the ligand during complex formation (from a pre-reaction state to a post-reaction adduct). Current covalent docking tools often require extensive manual preparation to generate these adduct structures, a process that is time-consuming, error-prone, and requires specialized expertise. Furthermore, many existing tools are restricted by commercial licensing or lack user-friendly interfaces for handling diverse warhead chemistries and target residues.

Methodology
The authors present CHARMM-GUI Covalent Ligand Docker (CGUI-CLD), a web-based platform designed to automate and streamline the covalent docking workflow using AutoDock4. The system integrates several key components:

1. Workflow Integration: CGUI-CLD is built upon the existing PDB Reader & Manipulator module. Users can upload protein structures (PDB) and ligands (SDF). The system utilizes the flexible side-chain method of AutoDock4, where the target residue side chain and the ligand are treated as flexible while the protein backbone remains rigid.
2. Automated Ligand Preparation: A core feature is a knowledge-based library supporting the seamless transformation of ligands into docking-ready adduct forms.
   • Pre-reaction Input: Users can upload a non-covalent scaffold or a pre-reaction ligand. The system uses a built-in library (supported by RDKit) to detect warheads and apply specific reaction chemistries to generate the covalent linkage.
   • Post-reaction Input: Users can directly upload or draw the final adduct structure if the specific warhead is not in the library.
   • Interface: A pop-up interface with Marvin JS allows users to select target residues, highlight warhead atoms, and visualize the generated linkage.
3. Library Scope: The built-in library currently supports 66 unique warheads and 8 amino acid residues, covering 104 specific warhead-residue combinations.
4. High-Throughput Integration: The platform connects with CHARMM-GUI High-Throughput Simulator (CGUI-HTS) to automatically generate molecular dynamics (MD) simulation systems and input files for the docked complexes.
5. Benchmarking Protocol: The system was evaluated using a dataset of 207 protein-ligand complexes from Scarpino et al. The workflow involved:
   • Generating up to 10 poses per complex.
   • Performing 10-ns MD simulations (NPT ensemble, 303.15 K, CHARMM36m force field) to refine poses.
   • Calculating Ligand RMSD against native structures and binding free energies using MM/GBSA (OBC2 model).
   • Using Receiver Operating Characteristic (ROC) analysis to evaluate metrics (docking scores, MD RMSD, MM/GBSA) as classifiers for pose accuracy.

Key Contributions

• Automation of Adduct Generation: CGUI-CLD significantly reduces the manual workload required to prepare covalent ligands by automating the transformation from pre-reaction to post-reaction states for a wide range of chemistries.
• Accessibility: It provides a free, web-based interface that lowers the expertise barrier for covalent docking, eliminating the need for command-line implementation or complex local setup.
• Integrated Validation: By linking docking directly to MD simulation setup, the platform allows for immediate structural refinement and stability assessment of docking poses.
• Extensive Library: With support for 66 warheads and 8 amino acids, it offers broader coverage than many existing free tools, addressing the rapid evolution of covalent warhead discovery.

Results

• Docking Accuracy: In the benchmark of 207 complexes, the Top-1 AutoDock4 pose achieved a success rate (RMSD < 2 Å) of 45.5%. When considering the best RMSD pose among all generated poses, the success rate increased to 69.6%.
• Pose Sampling: The system generated 865 total poses. While only 18.4% of Top-1 poses were within 1 Å of the native structure, 80.3% were within 3 Å.
• MD Refinement: Short (10-ns) MD simulations showed that incorrect poses (native RMSD > 2.5 Å) tended to be more unstable (higher RMSD drift) than correct poses.
• Pose Classification: ROC analysis indicated that Ligand MD RMSD (AUC = 0.673) and the combination of MD RMSD + Relative MM/GBSA (AUC = 0.693) were more effective classifiers for distinguishing correct from incorrect poses than docking scores alone (AUC = 0.629).
• Case Studies: The platform successfully reproduced native poses for Penicillin G (RMSD 0.71 Å), Tirabrutinib (RMSD 1.70 Å in cross-docking), and a trans-aziridine inhibitor, demonstrating its utility in diverse biological contexts.

Significance
The paper positions CGUI-CLD as a critical tool to address the "tedious preparation" bottleneck in covalent drug discovery. By automating the generation of covalent adducts and integrating docking with MD refinement, the platform aims to "significantly reduce a massive workload" and "advance covalent ligand research." The authors emphasize that this is the first free online platform capable of full ligand preparation for covalent docking. They note that while tools like AlphaFold3 show promise, CGUI-CLD provides a specialized, structure-based workflow that is currently essential for virtual screening, particularly given the limitations of existing tools in handling diverse warheads without custom scripting. The authors modestly suggest that the workflow's ligand transformation function could be generalized for use with other covalent docking programs in the future.