Drug repurposing of fluvastatin as a DP2 antagonist: an ensemble learning strategy with in vitro cellular functional...

We present a framework to repurpose FDA-approved drugs as DP2 antagonists for asthma, using a two-stage strategy. First, we trained and evaluated eight models-five machine learning algorithms and three graph neural networks-for their complementarity and performance. Then, we create a dynamically weighted ensemble predictor from the top-performing models, trained on DP2 inhibitor activity data (pIC₅₀). This ensemble showed strong predictive ability, with R ² values of 0.752, 0.722, and 0.684 for different data splits. SHAP analysis identified key molecular features for DP2 activity, including a molecular weight of 420-550 Da and a specific fingerprint bit. After screening a library of 2019 approved drugs, three candidates were found and tested through assays like cytotoxicity, cAMP inhibition, and cell migration. Pitavastatin demonstrated inadequate activity at non-toxic concentrations, and levocabastine's inhibition was not concentration-dependent, precluding the determination of its IC₅₀. Fluvastatin, however, was a potent DP2 antagonist with an IC₅₀ of 12.72 nM and achieved 94.9% inhibition of DK-PGD₂-induced cell migration at 243 nM. All code and datasets generated in this work are publicly accessible. Contribution: this study presents a reproducible, open-source workflow that combines split-adaptive model selection with targeted functional assays for drug repurposing. In contrast to traditional static ensembles that amalgamate all available models without discrimination, our approach selectively integrates complementary models using dynamic weighting tailored to each data splitting strategy. Additionally, SHAP interpretability offers actionable insights into structure-activity relationships for lead optimization. In this study, the developed workflow demonstrates effective performance in the discovery of DP2 antagonists, offering a valuable reference for AI-driven drug repositioning targeting this receptor.