Publication date: Jun 16, 2026
Coronavirus main protease (Mpro) is a conserved antiviral target, yet most generative AI pipelines optimize surrogate-predicted affinity without enforcing electronic plausibility or cross-target consistency. We present PRISM-Gen (Physics-guided Robust Inhibitor Selection Method – Generative Module), a multi-fidelity framework coupling fragment-tree molecular generation with a three-tier electronic screening cascade – GFN2-xTB semi-empirical descriptors, Gaussian Electronic Moderation (GEM) scoring, and B3LYP/6-31 G* DFT validation – followed by conservative worst-case docking across SARS-CoV-2, SARS-CoV-1, and MERS-CoV Mpro. Applied to 4136 generated candidates, the pipeline identifies 36 broad-spectrum-consistent inhibitor candidates whose top-ranked members exhibit predicted worst-case binding energies comparable to those of the non-covalent reference inhibitor ensitrelvir under identical docking conditions, while sharing no Bemis-Murcko scaffolds with nirmatrelvir, ensitrelvir, or GC376. Stage-wise statistical validation confirms that each tier exerts non-redundant, orthogonal selection pressure. Retrospective analysis demonstrates that replacing GEM’s continuous moderation with a conventional hard electronic cutoff would eliminate 48. 0% of candidates, including 55. 6% of the final 36 molecules, disproportionately depleting scaffold diversity. These results establish that continuous, physics-informed electronic moderation integrated within a multi-fidelity generative pipeline can recover structurally novel chemotypes that binary exclusion filters would irreversibly discard – a design principle applicable to generator-agnostic molecular discovery workflows.