Guided Ensemble Stacking Method for Predicting Biological Activities of Compounds.

Area of Science:

• Computational Chemistry
• Pharmacology
• Machine Learning

Background:

• Machine learning (ML)-driven quantitative structure-activity relationship (QSAR) models predict compound activity using structural properties.
• Traditional ML-QSAR models face limitations due to algorithm bias, data constraints, and neglect of pharmacokinetic (PK) properties, impacting drug discovery success rates.

Purpose of the Study:

• To develop an advanced ML-QSAR approach integrating supervised data preparation and ensemble stacking.
• To enhance prediction reliability by incorporating pharmacokinetic properties into QSAR modeling.
• To improve the accuracy and applicability of QSAR models in drug discovery pipelines.

Main Methods:

• Developed a guided ensemble-based ML approach combining supervised data preparation and ensemble stacking.
• Created two ensemble stacking models: a classification model for activity type (inhibition/activation) and a regression model for bioactivity values.
• Incorporated compound structural and pharmacokinetic (PK) properties for enhanced prediction.

Main Results:

• The classification model achieved over 0.85 accuracy in predicting biological activity type.
• The regression model attained an R-squared value above 0.77 for predicting bioactivity values.
• The proposed models demonstrated superior performance compared to traditional QSAR approaches.

Conclusions:

• The integrated ML-QSAR approach significantly improves predictive accuracy by incorporating PK properties.
• This method addresses key limitations of traditional QSAR, offering enhanced reliability for drug discovery.
• The enhanced QSAR models show potential for optimizing drug development pipelines.