Discovery of EP4 antagonists with image-guided explainable deep learning workflow

In target drug design, the manual creation of a poor initial compounds library, the time-consuming wet-laboratory experimental screening method, and the weak explainability of their activity against compounds significantly limit the efficiency of discovering novel therapeutics. Here we propose an image-guided, interpretability deep learning workflow, named LeadDisFlow, to enable rapid, accurate target drug discovery. Using LeadDisFlow, we identified four potent antagonists with single-nanomolar antagonistic activity against PGE₂ receptor subtype 4 (EP4), a promising target for tumor immunotherapy. Remarkably, the most potent EP4 antagonist, ZY001, demonstrated an IC₅₀ value of 0.51 ± 0.02 nM, along with high selectivity. Furthermore, ZY001 effectively impaired the PGE₂-induced gene expression of a panel of immunosuppressive molecules in macrophages. The workflow facilitates the discovery of potent EP4 antagonists that enhance anti-tumor immune response, and provides a convenient and quick approach to discover promising therapeutics for a specific drug target.