A Replicate-Aware Biophysical Triage Workflow for De-Risking Protein-Ligand Hypotheses Before Experimental Validation

Computational protein-ligand studies often treat ligand retention during molecular dynamics as evidence of preserved binding quality. This assumption is risky in drug-repurposing and resistance settings, where a ligand may remain pocket-associated while losing canonical contact fidelity, adopting a weakened pose, or showing replicate-dependent behavior. Here, we present a replicate-aware biophysical triage workflow, referred to in this manuscript as Gen2, for de-risking protein-ligand hypotheses before experimental validation. Gen2 uses predefined quality-control and trajectory-adjudication principles to evaluate whether a protein-ligand hypothesis remains structurally interpretable across replicate molecular dynamics simulations. We applied the workflow to two primary benchmark modules. In SARS-CoV-2 main protease, apo 6Y2E served as the ligand-free baseline, 6W63/X77 as a noncovalent positive-control anchor, and 6Y2E-lopinavir as a weak or nonvalidated comparator. X77 showed replicate-supported active-site retention and was accepted as the positive-control anchor, whereas lopinavir remained active-site associated but failed to preserve robust, replicate-consistent canonical late-pose quality. In BCL2-venetoclax systems, WT 6O0K behaved as a preserved positive anchor, F104L 6O0M showed bounded impairment without ligand escape, and G101V 6O0L showed replicate-asymmetric resistance-linked perturbation of neighboring groove contacts rather than direct G101 contact loss. Across both systems, ligand retention and binding-quality preservation were separable outcomes. The two modules were selected as orthogonal stress tests of the same methodological problem, not as evidence that the workflow is a universal potency or clinical-response predictor. Gen2 is not an affinity predictor, potency predictor, clinical-resistance model, or replacement for biochemical or cellular validation. Its intended role is to classify computational protein-ligand hypotheses as preserved, weakened, asymmetric, nonideal, setup-limited, or unsuitable for prioritized experimental escalation.