Selective virtual screening of hypoxia inducible factor agonists and antagonists using artificial intelligence and linear response theory.

Abstract

Hypoxia inducible factor (HIF) is a heterodimeric transcription factor that mediates a signaling cascade under low oxygen concentration and prevents cell death. Under normoxia, enhanced HIF activity promotes tumorigenesis, and it is considered as a key oncogenic driver in renal cell carcinoma. Enhancing HIF activity is a therapeutic approach to treat renal anaemia, thus making both HIF activators (agonists) and inhibitors (antagonists) a potential therapeutic target under different disease conditions. While known HIF agonists and antagonists have opposing effect on the HIF activity, both binds to same cavity in the C-terminal domain of HIF (PASB domain). We identify the atomic level mechanism behind differential effect of agonists/antagonists on HIF activity using molecular dynamics simulations and linear response theory. Agonists and antagonists trigger different allosteric pathway from the PASB domain to the DNA-binding domain of HIF transcription factor. The ligand induced external forces on the protein residues explains the observed allosteric effect and used as a descriptor to predict the activation/inhibition potential of a given ligand. A large-scale structure-based virtual screening of 700 million molecules from ZINC database is performed to identify the potential new ligands of HIF. Binding energies of 10 million ligands is predicted using Autodock Vina and used to train a directed-message passing neural network (D-MPNN) to predict binding energies of the remaining molecules, hence enabling the screening of the larger portion of molecules at a reasonable computational cost. The top candidate molecules are further classified as potential inhibitors based on the forces exerted by the ligands on protein residues.