Multi-task MSSP model can accurately predict selective Src inhibitors

Abstract

Src family kinases (SFKs), non-receptor tyrosine kinases, crucially contribute to invasion, tumor progression, epithelial-mesenchymal transition, angiogenesis, and metastasis. Thus, Src inhibitors offer a promising avenue for cancer therapy. This study introduced a multitask MSSP deep learning model to predict molecule inhibitory activity across multiple Src subtypes. Comparative assessment against four traditional machine learning methods—Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), eXtreme Gradient Boosting (XGBoost)—established the superior performance of the multitask MSSP model. It demonstrated the best comprehensive performance, achieving F1-Score and AUC values of 0.906 and 0.975, respectively. An online web server, "SRC-Predictor," was created to aid the practical application of the multitask MSSP model, predicting compounds' potential inhibitory activity against Src. Finally, compounds ranking in the top twenty based on model predictions were selected for experimental validation. Literature search for these compounds revealed limited research on four of them concerning Src. Molecular docking identified Doramapimod as exhibiting better affinity towards Src compared to reference compounds. It significantly inhibited Lyn kinase activity and influenced the secretion levels of inflammatory factors in LPS-induced macrophages. Experimental validation confirmed that our study provides a novel approach for identifying and screening lead compounds as Src inhibitors.