AI-Assisted Identification and Validation of Non-Covalent Inhibitors of SARS-CoV-2 Main Protease (Mpro) via Generative Design and Physics-Based Simulations

The emergence of Coronavirus 2 from serious acute respiratory syndrome (SARS-CoV-2) underscored the urgent need for effective antiviral therapies. The SARS-CoV-2 main protease (Mpro) remains a critical target due to its essential role in cleaving replicase polyproteins into functional nonstructural proteins required for viral replication. Its unique substrate specificity, evolutionary conservation among coronaviruses, and evolutionary divergence from human cysteine proteases make it a high-priority antiviral target. Here, we describe an integrated in silico platform, MedGraph™, which combines AI-based generative modeling, structure-guided docking, molecular dynamics (MD) simulations, and free energy perturbation (FEP) calculations to identify potential non-covalent Mpro inhibitors. Five de novo ligands were prioritized using a GPU-accelerated docking protocol into the substrate-binding cleft of Mpro (PDB: 7RFW), focusing on engagement with the catalytic dyad (Cys145-His41), followed by machine learning–aided multiparameter optimization and 100 ns MD simulations for dynamic assessment. To evaluate chemical novelty, the CHEESE tool was used for Morgan fingerprint-based and electrostatic-based similarity analyzes against Enamine and ZINC15 libraries, confirming the uniqueness of the prioritized scaffolds. Benchmarking against the COVID Moonshot dataset further confirmed low scaffold overlap, with maximum Tanimoto similarity values ≤0.35, high lighting exploration of underrepresented regions of chemical space. The resulting candidates therefore combine pharmacophoric anchoring with scaffold novelty, providing both therapeutic promise and intellectual property differentiation. Independently, identified hits demonstrated strong and stable binding, favorable thermodynamic profiles, and complementarity with conserved Mpro residues. The distinct chemotypes observed suggest synthetic novelty and potential tractability. This study demonstrates the utility of MedGraph™ in coupling AI-based generative modeling with physics-driven simulations to accelerate antiviral discovery and providing a generalizable framework for the iterative optimization of structurally diverse and pharmacologically tractable inhibitors.