Support Vector Machine Classification and Regression Prioritize Different Structural Features for Binary Compound...

Area of Science:

• Computational Chemistry
• Chemoinformatics
• Machine Learning

Background:

• Support Vector Machine (SVM) and Support Vector Regression (SVR) are key machine learning algorithms in computational chemistry and chemoinformatics.
• These methods are widely used for identifying active compounds and modeling structure-activity relationships.
• Molecular fingerprints are common descriptors for SVM and SVR, representing chemical structures and properties.

Purpose of the Study:

• To compare Support Vector Machine (SVM) and Support Vector Regression (SVR) calculations on identical compound datasets.
• To identify which molecular fingerprint features are responsible for predictions in SVM and SVR models.
• To analyze feature contributions and improve the interpretability of SVM/SVR models.

Main Methods:

• Systematic feature weight analysis was performed on SVM and SVR models.
• Comparison of feature sets contributing to predictive performance for both algorithms.
• Feature mapping was employed to interpret individual predictions.

Main Results:

• Significant differences were observed in the fingerprint features utilized by SVM and SVR models.
• The overlap between feature sets critical for SVM and SVR predictive performance was minimal.
• Features were identified that exerted opposite effects on SVM and SVR predictions.
• Feature weight analysis and mapping enabled the interpretation of specific predictions.

Conclusions:

• Molecular fingerprint contributions to SVM and SVR models are distinct and not interchangeable.
• Understanding feature importance enhances the interpretability of these machine learning models.
• This analysis provides insights into the differing roles of features in classification (SVM) and regression (SVR) tasks.