ChemPointNet: A 3D Geometric Deep Learning Framework for Virtual Screening of Potential EGFR Tyrosine Kinase Inhibitors

The epidermal growth factor receptor (EGFR) tyrosine kinase serves as a pivotal regulatory protein in the EGFR signaling pathway, driving tumorigenesis and cancer progression. Enzyme activation initiates through an autophosphorylation process mediated by its catalytic domain, with ATP functioning as the phosphate group donor. Aberrant phosphorylation of receptor tyrosine residues leads to persistent activation of downstream signaling pathways, resulting in uncontrolled cellular proliferation and malignant transformation. In this study, we propose ChemPointNet, a novel 3D structure-based classification technique leveraging PointNet architecture - a deep learning network demonstrating exceptional performance in 3D point cloud object classification. Our approach employed atomic three-dimensional coordinates of chemical compounds as input data for training the ChemPointNet model. The optimized model was subsequently applied to screen potential EGFR tyrosine kinase inhibitors from the SPECS compound library. Molecular docking analysis of the predicted hits identified 21 molecules exhibiting high binding affinity. Subsequent enzymatic activity assays revealed 17 highly active candidates suitable for structural optimization. Molecular dynamics simulations were further conducted to elucidate the binding mechanisms of the most promising hits. This integrated computational-experimental framework establishes an efficient strategy for drug discovery targeting oncogenic kinases.