<p>The limitations of molecular docking and molecular dynamics (MD) simulations, including rigid structural assumptions and high computational costs, hinder their ability to predict binding affinities accurately. This study introduces Binding-Net^Ana, a novel machine learning framework using graph convolutional neural networks to rescore protein-ligand complexes based on structural parameters. By leveraging graph-based representations and molecular features, the method demonstrated improved accuracy in predicting binding affinities (Kd and Ki) over traditional methods. Results showed that XGBoost and Random Forest outperformed other models, achieving a test correlation coefficient of 0.771 and 0.742 respectively, showcasing the potential of machine learning for large-scale virtual screening and hit identification. Protein-ligand complexes were retrieved from the PDBBind-CN database and processed into graph representations, incorporating features such as solvent-accessible surface area, atomic interactions, and Euclidean distances. Machine learning models, including Random Forest and XGBoost, were trained on these graph-embedded features and evaluated using metrics such as mean absolute error (MAE), root mean squared error (RMSE), and correlation coefficient (R). Additional evaluations included molecular docking and MD simulations were used to validate binding affinities and structural stability. A case study on the estrogen-related receptor gamma protein highlighted the comparative performance of machine learning and traditional physics-based methods.</p>