Guided Ensemble Stacking Method for Predicting Biological Activities of Compounds

Machine learning (ML)-driven quantitative structure-activity relationship (QSAR) modeling has gained significant attention for predicting compound biological activity based on compounds' structural, chemical, and physical properties because of the advancement of ML techniques. However, traditional ML-QSAR models often suffer from biases due to algorithm selection and limitations in training data. Additionally, these approaches root in deducing biological activities purely from compounds' structural information and disregard their pharmacokinetic (PK) properties, a key factor contributing to the 15% failure rate in clinical trials, limiting their applicability in drug discovery. To address these challenges, we propose a guided ensemble-based ML approach that integrates a supervised data preparation strategy with an ensemble stacking method, leveraging the strengths of multiple ML algorithms. By incorporating PK properties, our approach enhances prediction reliability. Specifically, we developed two ensemble stacking models: The classification model predicts the biological activity type, "inhibition" versus "activation," based on compound features, while the regression model predicts bioactivity values. The classification model achieved an accuracy exceeding 0.85, while the regression model attained an R² above 0.77, demonstrating superior performance over traditional QSAR models. These results highlight the potential of our approach in improving drug discovery pipelines by enhancing predictive accuracy and addressing key QSAR limitations.