Enabling Automatic Generation of Protein-Ligand Complex Data Sets with Atomistic Detail.

Area of Science:

• Computational chemistry and cheminformatics
• Structural biology
• Drug discovery and development

Background:

• Predicting protein-ligand bioactivities is critical for drug discovery but challenging.
• Supervised machine learning models are competitive in structure-based drug design but require high-quality data.
• Integrating bioactivity and protein structure data from repositories like ChEMBL and PDB is difficult.

Purpose of the Study:

• To create an automated and continuously growing dataset of annotated protein-ligand complexes.
• To facilitate the development of structure-based molecular design methods.
• To provide a foundation for novel datasets for training and testing machine learning models.

Main Methods:

• Utilized ActivityFinder for automated linking of protein sequence/structure with bioactivity data.
• Integrated ActivityFinder with tools for structure quality estimation and property calculation.
• Constructed the StrAcTable dataset based on ChEMBL (Version 35) and PDB data.

Main Results:

• Created StrAcTable, an automatically generated dataset of 20,063 annotated protein-ligand complexes.
• StrAcTable includes detailed quality metrics for data matching, macromolecular structure, ligands, and bioactivity.
• The automated approach ensures sustainable growth and accessibility of the dataset.

Conclusions:

• StrAcTable offers a valuable resource for constructing training and validation datasets in structure-based drug design.
• The dataset supports the development and testing of both machine learning and traditional molecular design methods.
• Automated data generation is key for building comprehensive and up-to-date resources for computational drug discovery.