A Workflow to Create a High-Quality Protein-Ligand Binding Dataset for Training, Validation, and Prediction Tasks

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Scoring functions (SFs) are crucial for predicting protein-ligand binding energies.
• High-quality structural and binding assay data are essential for developing accurate SFs.
• Existing datasets like PDBbind contain structural artifacts that can compromise SF reliability.

Purpose of the Study:

• To identify and address structural artifacts in protein-ligand datasets.
• To develop a semi-automated workflow (HiQBind-WF) for curating these datasets.
• To create a new, high-quality dataset (HiQBind) for SF development and validation.

Main Methods:

• Developed a series of algorithms for a semi-automated curation workflow (HiQBind-WF).
• Applied HiQBind-WF to curate non-covalent protein-ligand datasets.
• Integrated binding free energy data from BioLiP, Binding MOAD, and BindingDB with PDB structures to create HiQBind.

Main Results:

• Identified common structural artifacts in widely-used protein-ligand datasets.
• Demonstrated the potential of these artifacts to reduce the accuracy and generalizability of SFs.
• Successfully created the HiQBind dataset and workflow, ensuring reproducibility and minimizing human intervention.

Conclusions:

• The HiQBind-WF and HiQBind dataset offer a more reliable resource for developing and testing SFs.
• Addressing data quality issues is critical for advancing protein-ligand binding energy prediction.
• The open-source nature of HiQBind promotes transparency and collaboration in drug discovery research.