Multi-objective QSAR prediction of ERα antagonists via SHAP-based interpretation

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacology and drug discovery
• Machine learning in bioinformatics

Background:

• Evaluating drug candidates requires assessing both biological activity and Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties.
• Existing methods may not fully capture the complex relationships between molecular structure and pharmacological profiles.
• There is a need for integrated predictive frameworks to streamline drug development.

Purpose of the Study:

• To develop and validate a two-stage predictive framework for comprehensive drug candidate evaluation.
• To integrate Quantitative Structure-Activity Relationship (QSAR) modeling with machine learning for predicting drug activity and ADMET properties.
• To introduce and assess a novel Dual-Filter Feature Selection (DFFS) method for identifying key molecular descriptors.

Main Methods:

• A two-stage machine learning approach combining QSAR modeling with feature selection.
• Development of a Dual-Filter Feature Selection (DFFS) method integrating statistical analysis and machine learning feature importance.
• Application of LightGBM for activity prediction and a stacking model for multitask ADMET property prediction.
• Utilized molecular docking and SHAP analysis for mechanistic insights and model interpretation.

Main Results:

• The DFFS method successfully selected key molecular descriptors for QSAR modeling.
• LightGBM demonstrated superior performance in predicting ERα antagonist activity (MRE of 0.0775).
• The stacking model achieved high accuracy for ADMET property prediction, with AUC scores exceeding 0.95 for all tasks.
• DFFS outperformed individual feature selection methods and ChemBERTa-generated features.

Conclusions:

• The proposed two-stage predictive framework offers a robust approach for evaluating drug candidates.
• The DFFS method is effective in identifying relevant molecular descriptors for QSAR and machine learning models.
• The integrated framework facilitates the identification of high-activity compounds with favorable ADMET profiles, advancing drug discovery efforts.