Using machine learning to improve ensemble docking for drug discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Ensemble docking accounts for receptor flexibility in molecular docking for virtual screening.
• Current methods struggle to effectively use ensemble docking scores for classifying compounds as active or inactive.
• Adding more structures to ensembles can decrease, not increase, classification performance.

Purpose of the Study:

• To improve the performance of ensemble docking for virtual screening by developing more effective machine learning models.
• To address the limitations of previous machine learning approaches, such as the independence assumption in naïve Bayesian models.

Main Methods:

• Applied several machine learning methods: k-nearest neighbor, logistic regression, support vector machine, and random forest.
• Relaxed the independence assumption between docking scores from different receptor structures.
• Evaluated the performance of these enhanced models in classifying compounds.

Main Results:

• Significant improvements were observed in compound classification performance using the developed machine learning models.
• Relaxing the independence assumption led to better utilization of ensemble docking scores.
• The new methods demonstrated enhanced accuracy in distinguishing between active and inactive compounds.

Conclusions:

• Machine learning methods, particularly those that do not assume independence of docking scores, offer a significant improvement for ensemble docking.
• These advanced models provide a more effective way to leverage ensemble docking data for virtual screening and drug discovery.
• The findings suggest a promising direction for enhancing the reliability and performance of virtual screening techniques.