Ligand-based graph neural network, molecular dynamics and biological evaluation for identification of potential FGFR1 kinase inhibitors.

Currently approved FGFR1 kinase inhibitors still suffer from several disadvantages, including significant toxicities and adverse effects. To identify promising FGFR1 inhibitors, we developed a ligand-based graph neural network (GNN) virtual screening model and conducted high-throughput screening of FGFR1 kinase inhibitors using this model. Compared with traditional models, our model reveals better performance (Accuracy of 0.9186, AUC of 0.964). The model was employed to conduct preliminary screening of a 14-million-compound library. Subsequently, multi-stage screening was performed, including two rounds of molecular docking and two stages of molecular dynamics simulations. This approach identified five compounds capable of forming stable hydrogen bond interactions with FGFR1 and strong binding free energy (between − 28.23 and − 34.24 kcal/mol), which exhibited adequate binding stability in 300 ns molecular dynamics simulation. Four of the compounds were then subjected to biological evaluation, which revealed that they exhibited varying degrees of inhibitory activity against FGFR1 kinase. Notably, two of these compounds demonstrated excellent inhibitory activity ([Formula: see text]M and 3.3 [Formula: see text]M, respectively). In summary, this study developed a high-throughput virtual screening model for FGFR1 kinase inhibitors based on graph neural networks, providing novel lead compounds for the research of FGFR1 kinase inhibitors.