Targeting HsDHODH: Shape and machine-learning guided discovery and structural validation of SARS-CoV-2 antivirals

The frequent emergence of novel RNA viruses and the rapid development of viral resistance highlight the urgent need for new antiviral therapies. In this study, we employed an integrated computational pipeline combining shape-based models, machine-learning (ML)-based models, and hotspot analysis to identify novel human dihydroorotate dehydrogenase (HsDHODH) inhibitors. A virtual screening (VS) campaign of the H3D chemical library identified H3D-002856 as the primary hit (IC₅₀ = 2.92 ± 0.07 μM), which subsequently guided a similarity-based hit expansion (Tanimoto coefficient≥0.7) of 17 structural analogs. This two-stage workflow collectively yielded four anthranilate-based compounds H3D-002856, H3D-003181, H3D-001915, and H3D-003186 with enzymatic IC₅₀ values below 5.0 μM. Our crystal structures of HsDHODH in complex with H3D-002856, H3D-003181, and H3D-003186 confirmed the predicted binding modes and revealed a conserved interaction pattern across the series. The four most potent compounds successfully inhibited SARS-CoV-2 replication in vitro in Calu-3 cells, demonstrating EC₅₀ values ranging from 1.3 to 3.9 μM, and a favorable selectivity index, supported by minimal cytotoxicity (CC₅₀ values ≥ 100 μM). These findings highlight the value of integrated computational-experimental workflows in the discovery of HsDHODH inhibitors, establishing this scaffold as a viable starting point for further development with potential applicability against SARS-CoV-2. Furthermore, by using accessible, shape-based and low-code ML-based modeling platforms, this workflow provides a reproducible and practical template that can be readily adopted by medicinal chemists to accelerate lead identification.