Knowledge-Based Methods To Train and Optimize Virtual Screening Ensembles.

Area of Science:

• Computational drug discovery
• Molecular modeling
• Bioinformatics

Background:

• Ensemble docking is vital for virtual screening due to target conformational flexibility.
• Current methods struggle to select optimal conformational subsets, leading to false positives.

Purpose of the Study:

• To present three knowledge-based methods for constructing structural ensembles for virtual screening.
• To optimize ensemble selection using receiver operating characteristic (ROC) curve metrics (AUC and EF).

Main Methods:

• Developed three methods with varying computational scaling (O(2^N), O(N^2), O(N)) for ensemble selection.
• Optimized ensembles by maximizing the area under the ROC curve (AUC) or enrichment factor (EF).
• Applied methods to androgen receptor (AR), CDK2, and PPAR-δ targets using crystal structures and simulation data.

Main Results:

• All three methods demonstrated similar performance across training and test sets for the tested targets.
• The methods effectively selected conformational subsets to improve virtual screening.
• Computational scaling varied, with a linear O(N) method offering efficiency.

Conclusions:

• The presented knowledge-based methods provide effective strategies for optimizing ensemble docking.
• These approaches enhance virtual screening by reducing false positives and improving drug candidate identification.
• The methods are broadly applicable across various drug targets.