Machine Learning-Assisted Design of Peptidomimetic Inhibitors Targeting SARS-CoV-2 Main Protease: An In Silico and In Vitro Validation

The persistent emergence of SARS-CoV-2 variants and the potential for resistance against frontline therapeutics underscore the urgent necessity for novel antiviral agents. The main protease (Mpro) of SARS-CoV-2 remains one of the most clinically validated therapeutic targets due to its indispensable role in viral polyprotein processing and the absence of human homologues with identical substrate specificity. In this study, we implemented an integrated machine learning-assisted design pipeline to construct and prioritize a focused library of peptidomimetic inhibitors tailored to the substrate-binding pocket of Mpro. Utilizing a deep graph neural network trained on bioactivity and crystallographic datasets, we identified high-affinity peptidomimetic backbones equipped with diverse electrophilic warheads. Top-ranked candidates underwent extensive in silico profiling, including ensemble docking, 500-ns all-atom molecular dynamics simulations, and MM-GBSA binding free energy assessments. Two prioritized candidates, designated PM-4 and PM-7, demonstrated remarkable structural stability within the catalytic dyad (His41-Cys145) microenvironment. Subsequent solid-phase synthesis and in vitro validation using a continuous fluorogenic resonance energy transfer (FRET) enzymatic assay revealed that PM-7 inhibits recombinant SARS-CoV-2 Mpro with an IC50 of 0.38 ± 0.04 µM. Furthermore, cellular cytotoxicity evaluations in human HEK293T cells indicated negligible cytotoxicity (CC50 > 100 µM). These findings highlight the utility of combining deep learning with rigorous biophysical simulations to accelerate peptidomimetic lead discovery, offering a promising scaffold for next-generation coronavirus therapeutics.