Abstract
The rapid evolution of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) underscores the need for antivirals that are resilient to resistance. Current Food and Drug Administration (FDA)-approved therapies primarily target single viral mechanisms, leaving gaps in efficacy. Here, we developed a Deep Learning-based Activity Screening Model (DLASM), which integrates graph convolutional network with machine learning to identify SARS-CoV-2 inhibitors, using experimental 3-chymotrypsin-like (3CL) main protease assay data. The optimized DLASMs virtually screened 170,000 compounds from diverse in-house collections and yielded novel hits, several of which not only inhibited the 3CL protease but also blocked viral entry by interfering with heparan sulfate-mediated host interactions. These activities were validated through multiple assays, including 3CL enzymatic inhibition, SARS-CoV-2 pseudotyped particle entry, α-synuclein fibril uptake as a proxy for endocytosis, live virus cytopathic effect, heparan sulfate-dependent entry assay, and a 3D human lung mucociliary tissue model. Molecular docking studies elucidated binding modes at the 3CL protease active site, while molecular dynamics simulations provided insights into compound-heparan sulfate interactions. The identified compounds represent early-stage hits with moderate potency that demonstrate dual-mechanism antiviral activity. Together, these findings establish dual-target inhibition as a promising antiviral strategy, offering not only enhanced potency but also reduced risk of resistance. Moreover, our DLASM framework provides a generalizable pipeline for identifying chemically diverse scaffolds and for broader applications beyond SARS-CoV-2.