A machine learning-based method to improve docking scoring functions and its application to drug repurposing.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery

Background:

• Traditional docking scoring functions exhibit limited accuracy in predicting binding affinity.
• Existing methods often use fixed weights and assume additive interactions, neglecting nonlinear dependencies.
• Gene family-specific weights and nonlinear interaction modeling are crucial for accurate affinity prediction.

Purpose of the Study:

• To develop improved molecular docking scoring functions using machine learning.
• To enhance the correlation between predicted and known binding affinities.
• To create a novel scoring function for Mycobacterium tuberculosis (M.tb) InhA inhibitors.

Main Methods:

• Employed support vector machines (SVMs) to train prediction models.
• Utilized energy terms from molecular docking and high-throughput screening data (BindingDB, DUD).
• Developed regression and classification models, including a novel multiple-planar SVM for imbalanced data.

Main Results:

• SVM-based models significantly improved the correlation between predicted and known binding affinities compared to the original eHiTS scoring function.
• The developed models demonstrate the effectiveness of nonlinear methods in scoring function development.
• A new scoring function for M.tb InhA was trained, suggesting potential repurposing of phosphodiesterase inhibitors.

Conclusions:

• Nonlinear machine learning approaches, like SVMs, offer a powerful alternative to traditional scoring functions.
• The methodology can be generalized to other gene families with available structural and activity data.
• This work advances computational drug discovery by providing more accurate binding affinity predictions and novel therapeutic strategies.