Combining global and local measures for structure-based druggability predictions.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of drug target druggability is vital for efficient pharmaceutical research and development.
• Existing methods often rely on global pocket properties, which can be limited by ligand-induced binding site changes and boundary definition issues.

Purpose of the Study:

• To develop and validate an automated algorithm, DoGSiteScorer, for predicting protein pocket druggability.
• To integrate both global and local pocket properties for improved druggability prediction accuracy.

Main Methods:

• Utilized a support vector machine (SVM) trained on a comprehensive druggability dataset (DD) and its non-redundant version (NRDD).
• Incorporated global pocket descriptors (size, shape, hydrophobicity) and local pocket properties analyzed via nearest neighbor searches and distance-dependent histograms of atom pairs.
• Identified discriminant features, such as the ratio of short-range hydrophilic-hydrophilic to lipophilic-lipophilic pairs.

Main Results:

• The SVM model achieved 90% accuracy in classifying druggable versus undruggable targets using global descriptors on the NRDD.
• Incorporating local pocket properties improved accuracy, with 88% conformity between nearest neighbors and the pocket structure's druggability type.
• Found that druggable pockets tend to have fewer short-range hydrophilic-hydrophilic pairs and more short-range lipophilic-lipophilic pairs.

Conclusions:

• DoGSiteScorer provides a robust, automated approach to druggability prediction by combining global and local pocket features.
• The inclusion of local pocket properties addresses limitations of global descriptors, particularly concerning pocket flexibility and ligand binding.
• This method enhances the reliability of descriptor-based druggability prediction, aiding in prioritizing disease-modifying targets for drug development.