Probabilistic Pocket Druggability Prediction via One-Class Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate in silico druggability prediction is crucial for efficient drug discovery campaigns.
• Traditional two-class classification methods for druggability are limited by the ambiguous nature of the 'non-druggable' class.
• A one-class approach focusing on confirmed druggable pockets offers a more robust strategy.

Purpose of the Study:

• To propose and validate a novel one-class computational method for predicting protein target druggability.
• To leverage the import vector domain description (IVDD) algorithm for enhanced druggability assessment.
• To address and mitigate biases inherent in traditional druggability prediction labeling.

Main Methods:

• Utilized the import vector domain description (IVDD), a one-class probabilistic kernel machine.
• Employed customized DrugPred descriptors computed via NanoShaper as input features for the IVDD algorithm.
• Applied a one-class classification strategy focusing exclusively on known druggable pockets.

Main Results:

• Demonstrated the feasibility and effectiveness of the proposed IVDD-based one-class approach for druggability prediction.
• Successfully mitigated biases associated with pocket labeling in computational drug discovery.
• The method shows promise in accurately identifying druggable pockets, improving prediction reliability.

Conclusions:

• The one-class IVDD approach provides a more appropriate and effective strategy for in silico druggability prediction compared to traditional two-class methods.
• This method enhances the reliability of druggability assessments by focusing on unambiguous data (druggable pockets).
• The findings support the integration of this approach into drug discovery pipelines to improve target selection and reduce bias.