DeepPurpose based deep learning approach for drug repurposing targeting VEGFR2 in hepatocellular carcinoma

Hepatocellular carcinoma (HCC) is one of the most prevalent and lethal malignancies worldwide due to its molecular heterogeneity, complex signaling pathways, and limited therapeutic options. Existing treatments often show suboptimal efficacy and significant toxicity, highlighting the urgent need for innovative strategies to identify safer and more effective therapeutics. Drug repurposing offers a cost- and time-efficient alternative by leveraging compounds with established safety profiles. In this study, we developed a drug repurposing pipeline based exclusively on the DeepPurpose deep learning framework to identify potential inhibitors targeting Vascular Endothelial Growth Factor Receptor 2 (VEGFR2), a key protein implicated in HCC progression. The model was trained on a manually curated dataset of experimentally validated drug-target interactions incorporating SMILES representations and protein sequences with reported bioactivity values. The optimized model achieved a correlation coefficient of 0.81 and a mean squared error (MSE) of 0.45, demonstrating reliable predictive performance. Virtual screening of FDA-approved and clinical-stage compounds identified top candidates, which were further evaluated through drug-likeness assessment, ADMET profiling, molecular docking, MM-GBSA binding free energy calculations, molecular dynamics simulations, and DFT analysis. Edotecarin (107), Rebeccamycin (90), and APY-29 (55) exhibited docking scores of - 11.062, - 10.067, and - 9.866 kcal/mol, with MM-GBSA energies of - 66.56, - 65.78, and - 67.45 kcal/mol, respectively. Initial 10 ns MD simulations were performed for the top 5 compounds, followed by extended 100 ns simulations for the top 3 most stable candidates. Among them, Compound 55 (APY-29) demonstrated superior dynamic stability and was therefore further subjected to an extended 500 ns molecular dynamics simulation, which confirmed sustained protein-ligand stability and persistent intermolecular interactions throughout the trajectory. Overall, the MD simulations confirmed stable complexes with RMSD values ranging from 1.395 to 1.748 Å. Furthermore, DFT analysis of Compound 55 revealed a HOMO-LUMO energy gap of 4.284 eV, supporting favourable electronic stability and reactivity characteristics. This integrative DeepPurpose-based workflow effectively bridges deep learning prediction with molecular-level validation for HCC drug repurposing.

The online version contains supplementary material available at 10.1007/s40203-026-00669-6.