AI-Assisted Reprioritization of HBV Core Benzimidazole Candidates: From a Conserved Motif to a Validated CAM-Pocket Top Cluster

This record deposits the manuscript and reproducibility package for Paper C of the HBV VISLM series: a prospective, candidate-agnostic, AI-assisted computer-aided drug design (CADD) workflow for reprioritizing benzimidazole candidates at the hepatitis B virus core protein (HBcAg) capsid-assembly-modulator / heteroaryldihydropyrimidine (CAM/HAP) inter-dimer pocket. The workflow begins from a conserved N-terminal HBcAg motif, YKEFGA[S/T] at core residues 6–12, identified in the companion motif-analysis study. Rather than assuming that this conserved motif is a small-molecule target, the workflow first tests motif druggability prospectively. Drug-anchored fpocket analysis shows that the motif is not a druggable small-molecule pocket, with a druggability score of approximately 0.001 compared with 0.733 for the validated 5GMZ CAM pocket. Structure-based design is therefore redirected from the motif itself to the experimentally validated CAM/HAP pocket. Eleven benzimidazole candidates are then docked candidate-agnostically with AutoDock Vina under fixed seeds. Historical YM-numbering is mapped onto the anonymous candidate identifiers only after ranking and identity locking. The resulting top cluster is stress-tested on a Y132-revertant 5GMZ B+C receptor model through five orthogonal computational layers: native-ligand redocking, GNINA consensus scoring, single-trajectory MM-GBSA rescoring, short explicit-solvent molecular dynamics, and an OpenFE relative binding free-energy (RBFE) network. The defensible output is a validation-ready benzimidazole top cluster, not a single decisive winner. cand_01, historically designated YM-005 after identity lock, is the Y132A-baseline numerical leader. cand_03 is the numerical leader on the Y132-revertant receptor model and in the RBFE network. The candidate cluster remains below the native CAM reference in the control contexts where native CAM is included, and the RBFE network is an extended pilot tree without cycle-closure. No experimental binding, capsid-assembly, antiviral, cytotoxicity, or off-target data are presented. All conclusions are therefore computational hypotheses requiring experimental validation. This deposit includes the manuscript with embedded figures, a figures-plus-supplementary display PDF, machine-readable supplementary tables and methods, the locked discovery protocol, per-gate audit notes, analysis and figure-rendering scripts, receptor and docked-pose structures, single-source-of-truth result files for docking, MM-GBSA, MD, and RBFE analyses, environment specifications, and a per-file SHA-256 manifest. Trained model weights from the companion VISLM studies and large MD trajectory files are intentionally not included in this record; the deposited poses, scripts, fixed seeds, and result files provide the reproducibility basis for the reported rankings and figures. This is the third deposit in a three-part HBV VISLM series and is cross-linked to the companion records: Paper A, doi:10.5281/zenodo.20529386, and Paper B, doi:10.5281/zenodo.20571631.