Positive Predictive Value Surfaces as a Complementary Tool to Assess the Performance of Virtual Screening Methods.

Area of Science:

• Computational chemistry
• cheminformatics
• Drug discovery

Background:

• Receiver Operating Characteristic (ROC) curve metrics are standard for virtual screening benchmarking.
• Predicting the actual probability of a hit being active (Positive Predictive Value - PPV) is crucial but challenging due to unknown library yields.
• Traditional metrics do not fully address the practical need for predicting true hit rates in virtual screening.

Purpose of the Study:

• To evaluate PPV surfaces from simulated screening as a complementary tool to ROC curves.
• To benchmark virtual screening methods and optimize score cutoff values using PPV surfaces.
• To enhance the practical application of computational methods in drug discovery.

Main Methods:

• Utilized retrospective virtual screening experiments.
• Employed four diverse datasets for Quantitative Structure-Activity Relationship (QSAR) classifier inference.
• Assessed inhibitors for Trypanosoma cruzi trypanothione synthetase, Trypanosoma brucei N-myristoyltransferase, GABA transaminase, and anticonvulsant activity.

Main Results:

• PPV surfaces effectively compare machine learning model performance in virtual screening.
• The approach aids in selecting appropriate score thresholds for virtual screening.
• Ensemble learning models demonstrated superior predictivity and robustness compared to other methods.

Conclusions:

• PPV surfaces are valuable for evaluating virtual screening tools and setting score thresholds for prospective screens.
• Ensemble learning consistently improves predictivity and robustness in virtual screening applications.
• The study highlights the practical utility of PPV surfaces in computational drug discovery.