Screening and classification of anti-angiogenic VEGFR2 inhibitors with supervised machine learning, deep learning and...

Vascular Endothelial Growth Factor Receptor 2 (VEGFR2) is a critical therapeutic target in cancer due to its role in pathological angiogenesis and tumor progression. Despite available FDA-approved VEGFR2 tyrosine kinase inhibitors (TKIs), challenges like resistance, off-target toxicities, and low discovery success rates persist, necessitating more efficient and scalable approaches to identify novel inhibitors. We trained five machine learning (ML) and two deep learning (DL) models on a dataset of 17,750 compounds derived from three public databases (e.g., PubChem, ChEMBL, and BindingDB). The best-performing model, Support Vector Machine (SVM) using an ECFP-like fingerprint (91.7% test accuracy), was used to select top-predicted compounds for subsequent ADMET analysis and molecular docking. We also introduce a custom ML/DL program predicting VEGFR2 inhibition probability based on SMILES input. Models demonstrated strong performance, achieving an average Matthews Correlation Coefficient (MCC) of 0.775; ECFP and RDKit fingerprints yielded the highest predictive performance. ADMET analysis indicated favorable pharmacokinetic properties (e.g., good passive GI absorption, low BBB permeation) for 11 of 17 top-ranked novel small molecules, suitable for oral peripheral administration. Molecular docking confirmed strong binding affinities to the VEGFR2 ATP-binding site (ranging -7.9 to -11.4 kcal/mol) for all 17 candidates. Notably, molecule 4 (-11.4 kcal/mol) and molecule 8 (-11.3 kcal/mol) exhibited affinities superior to FDA-approved inhibitors Sorafenib and Ponatinib. Molecular dynamics simulation results (structural and thermodynamics analysis) confirm these data. These findings identify promising novel VEGFR2 inhibitors with favorable drug-like properties. To facilitate the identification of potential VEGFR2 kinase inhibitors, we developed VEGFR2pred, a user-friendly graphical interface that integrates pretrained machine learning models, including RF, SVM, and XGB, and two best-performing molecular representations based on our results: ECFP and RDKit. To enhance accessibility and reproducibility, the complete tool, along with source code and installation instructions, has been made publicly available on GitHub.