Potency-directed similarity searching using support vector machines.

Area of Science:

• Chemoinformatics
• Machine Learning
• Drug Discovery

Background:

• Support vector machine (SVM) modeling is widely used in chemoinformatics.
• Advanced SVM applications, such as multitask learning for ligand-target prediction, are emerging.
• Current methods can be improved for identifying potent compounds in large datasets.

Purpose of the Study:

• To introduce a novel SVM approach for integrating compound potency into similarity searching.
• To enhance database selection by enriching sets with potent drug candidates.
• To improve hit identification in high-throughput screening (HTS).

Main Methods:

• Development of a structure-activity kernel function for SVM.
• Implementation of a potency-oriented SVM linear combination approach.
• Application of fingerprint descriptors and potency-directed SVM searching on HTS datasets.

Main Results:

• Potency-directed SVM searching successfully applied to four HTS datasets.
• Comparison of different SVM strategies demonstrated effectiveness.
• Potency-directed SVM searching meets or exceeds standard SVM recall rates.
• Significantly more potent hits were detected compared to standard methods.

Conclusions:

• The novel potency-directed SVM approach effectively enriches database selection with potent compounds.
• This method improves hit detection in HTS compared to traditional SVM calculations.
• The structure-activity kernel and linear combination strategy offer a valuable tool for drug discovery.