Robust optimization of SVM hyperparameters in the classification of bioactive compounds

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Support Vector Machine (SVM) is a key machine learning tool for virtual screening in drug discovery.
• Optimizing SVM hyperparameters, such as C and gamma, is crucial for effective virtual screening.
• Efficient hyperparameter optimization is needed to maximize predictive power and identify drug candidates.

Purpose of the Study:

• To investigate Bayesian and random search for optimizing SVM hyperparameters.
• To compare these methods against grid search and heuristic approaches for classifying bioactive compounds.

Main Methods:

• Bayesian optimization of SVM hyperparameters.
• Random search optimization of SVM hyperparameters.
• Comparison with grid search and heuristic hyperparameter selection.

Main Results:

• Bayesian optimization demonstrated superior efficiency and speed in hyperparameter tuning.
• Random search outperformed grid search and heuristic methods in classification performance.
• Bayesian optimization achieved the lowest number of iterations for optimal predictive performance.

Conclusions:

• The Bayesian approach offers superior accuracy and speed for SVM optimization in bioactivity assessment.
• Random search is a viable alternative when Bayesian optimization is not feasible.
• Optimized SVM models enhance the accuracy of classifying bioactive chemical compounds.