Statistical analysis of ADMET correlations.

To achieve a comprehensive evaluation of candidate drugs in terms of both biological activity and ADMET properties, this study proposes a two-stage predictive framework based on Quantitative Structure–Activity Relationship (QSAR) modeling integrated with machine learning techniques, elucidating the quantitative relationships between molecular structure and pharmacological properties. A novel Dual-Filter Feature Selection (DFFS) method integrates statistical analysis and feature importance scores derived from machine learning models. The averaged rankings are used to obtain a robust set of molecular descriptors. In the first stage, 20 key two-dimensional molecular descriptors were selected via DFFS from ERα antagonists. RF, XGBoost, LightGBM, and gcForest—were employed for activity prediction. Experimental results indicated LightGBM achieved the best performance, with MRE of 0.0775. The comparative experiment demonstrates that under the same LightGBM regression framework, DFFS outperformed its individual components—Mutual Information and XGBoost—as well as the high-dimensional features generated by ChemBERTa. In the second stage, based on 40 descriptors selected by DFFS, a stacking model was constructed to perform multitask prediction of ADMET properties, ensuring that high-activity candidate compounds also exhibit favorable profiles in absorption, distribution, metabolism, excretion, and toxicity. The AUC scores for all five ADMET models exceeded 0.95. To elucidate the molecular mechanisms and interpret the model decisions, we applied Phi coefficient analysis to assess inter-property correlations and SHAP analysis to identify key molecular features governing compound activity. Furthermore, molecular docking was performed to evaluate the binding affinity of highly active compounds towards the target protein, thereby providing quantitative validation of the predicted biological activities.