Exploring Alternative Strategies for the Identification of Potent Compounds Using Support Vector Machine and...

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Support vector regression (SVR) is a key method for predicting compound potency.
• SVR extends quantitative structure-activity relationships (QSAR) by handling complex structure-activity relationships (SARs).
• Accurate potency prediction is crucial for virtual compound screening and database enrichment.

Purpose of the Study:

• To evaluate novel strategies for compound potency prediction in virtual screening.
• To compare direct SVR, two-stage SVM-SVR, and SVR with active/inactive training data.
• To determine the optimal approach for balancing prediction accuracy and identifying potent compounds.

Main Methods:

• Direct support vector regression (SVR) for potency prediction.
• Two-stage modeling combining support vector machine (SVM) and SVR.
• SVR models trained with both active and inactive compounds.

Main Results:

• SVR models trained with active and inactive compounds maximized recall but reduced accuracy for high potency values.
• Direct SVR predictions were preferred for accuracy in high potency values.
• Combined SVM-SVR modeling offered the best balance between prediction accuracy and enrichment of potent compounds.

Conclusions:

• The study extends compound potency prediction methods for virtual screening.
• Combined SVM-SVR modeling provides a superior strategy for identifying potent compounds.
• Findings aid in optimizing virtual screening workflows for drug discovery.