Choosing feature selection and learning algorithms in QSAR.

Area of Science:

• Quantitative Structure-Activity Relationship (QSAR) studies
• Cheminformatics
• Machine Learning in Drug Discovery

Background:

• Feature selection is crucial for QSAR model development.
• Previous work assessed feature selection method performance.
• The interaction between feature selection and learning methods was unexplored.

Purpose of the Study:

• To investigate if specific feature selection methods enhance prediction accuracy when combined with certain machine learning algorithms.
• To empirically evaluate the synergy between feature selection techniques and various QSAR learners.

Main Methods:

• Employed four feature selection methods: wrapper, ReliefF, MARS, and elastic nets.
• Utilized eight machine learning algorithms: MARS, elastic net, random forest, SVM, neural networks, multiple linear regression, PLS, and kNN.
• Conducted in silico experiments on real QSAR datasets.

Main Results:

• State-of-the-art learners (random forest, SVM, neural networks) did not show improved prediction accuracy with feature selection.
• No specific feature selection method demonstrated superior performance with any particular learner.
• The combination of feature selection and learning methods did not yield significant gains in predictive power.

Conclusions:

• Feature selection offers no discernible benefit for advanced QSAR models.
• The choice of feature selection method is not critical when using powerful machine learning algorithms.
• Future QSAR research should focus on other areas for predictive accuracy enhancement.