Predictiveness curves in virtual screening.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Predictiveness curves are established metrics in clinical epidemiology for evaluating diagnostic and prognostic tools.
• Virtual screening methods are crucial for identifying potential drug candidates.
• Existing evaluation metrics for virtual screening may not fully capture predictive performance.

Purpose of the Study:

• To adapt and apply predictiveness curves to the field of virtual screening.
• To provide a graphical and intuitive tool for comparing the predictive power of different virtual screening methods.
• To introduce metrics for quantifying early recognition of active compounds in virtual screening.

Main Methods:

• Utilized logistic regression models to calculate activity probabilities from virtual screening scores.
• Applied predictiveness curves, analogous to ROC curves, to analyze score distributions.
• Introduced total gain and partial total gain for quantifying recognition and early recognition.

Main Results:

• Predictiveness curves effectively describe the dispersion of scores, unlike ROC curves.
• Demonstrated the utility of predictiveness curves in quantifying the predictive performance on subsets of molecular datasets.
• Illustrated the application of predictiveness curves using virtual screening data from the Directory of Useful Decoys with multiple methods.

Conclusions:

• Predictiveness curves offer a comprehensive evaluation of virtual screening score predictiveness.
• These curves serve as a valuable complement to existing tools like ROC curves for analyzing virtual screening results.
• Predictiveness curves can aid in defining optimal score thresholds for experimental compound selection in drug discovery.