Drug/nondrug classification using Support Vector Machines with various feature selection strategies

Area of Science:

• Computational Chemistry
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Virtual screening of small molecules is crucial in modern drug discovery.
• Efficient classification methods are needed to screen large compound libraries.
• Support Vector Machines (SVM) are powerful machine learning tools for classification.

Purpose of the Study:

• To evaluate Support Vector Machines (SVM) for classifying active and inactive molecules in early-phase drug discovery.
• To compare the performance of SVM with different feature selection strategies.

Main Methods:

• Utilized Support Vector Machines (SVM) for molecule classification.
• Applied Pearson's correlation coefficient for feature filtering and reduction.
• Investigated SVM performance with feature selection methods: SVM-Recursive Feature Elimination, Wrapper Method, and Subset Selection.

Main Results:

• Feature selection methods generally improved SVM performance compared to a single SVM.
• Subset Selection emerged as the most effective feature selection strategy.
• SVM proved to be a useful tool for classification tasks in real-life drug discovery.

Conclusions:

• SVM, particularly with Subset Selection, is a valuable method for accelerating early-phase drug discovery.
• The study validates SVM's utility in efficiently classifying large sets of chemical compounds.