AnCore: A Six-Stage Structure-Based Virtual Screening Pipeline Integrating Two-Phase Redocking, Reciprocal Rank Fusion Consensus Scoring, and 300 K Langevin Dynamics Relaxation

Background. Structure-based virtual screening is routine in early drug discovery, but its desktop tools have not kept pace. Many rely on legacy Python or X11 toolkits that run poorly on modern macOS, handle one ligand at a time, and stop at a static docking score. Static scores ignore thermal and entropic effects. As a result, property-matched decoys with opportunistic hydrogen bonds can outrank true ligands. Methods. We built AnCore Scientific Studio, a native macOS workbench written in Rust/Tauri. It uses a WKWebView and 3Dmol.js interface and a six-stage Python engine. It accepts RCSB or local PDB receptors and multi-molecule SDF, MOL, or SMILES ligands. The stages are receptor curation and 5.0 Å pocket mapping; a two-phase redocking gate (BFGS basin relaxation plus focused-box Monte Carlo sampling, RMSD ≤ 2.0 Å); ligand preparation (ETKDGv3, MMFF94s); consensus screening with AutoDock Vina, Vinardo, and AutoDock4 through Reciprocal Rank Fusion (k = 10); 3D interaction fingerprinting with 2D pharmacophore mapping; and 300 K Langevin triage (50 frames, 5.0 ns effective window; mean RMSD ≤ 2.0 Å, hotspot hydrogen-bond occupancy ≥ 65.0%). We tested the protocol on five targets (1IEP, 3ERT, 1UYG, 1M17, 1STP) with 50 actives and property-matched ZINC decoys. We then applied it to a new target, 4WKQ. Results. The redocking gate met the 2.0 Å limit for all five targets (mean RMSD 0.419 Å). It also fixed an 11.71 Å pose inversion in 1IEP. Stage 4 gave strong early enrichment on four targets (EF20% = 2.00×; AUC 0.680–1.000). Streptavidin (1STP) was a scoring trap. The decoy ZINC_Decoy_S3 (−7.95 kcal/mol) outranked D-biotin (−7.27 kcal/mol; AUC 0.360). At 300 K, the decoy drifted out of the pocket (mean RMSD 3.61 Å, hydrogen-bond occupancy 43.9%). D-biotin stayed locked in place (0.46 Å, 97.9%). Across the suite, Stage 6 rejected 25/25 decoys. It raised the mean ROC-AUC from 0.728 to 1.000 and lifted top-1 accuracy from 80% to 100%. Conclusions. AnCore links static docking with dynamic validation in one reproducible, step-by-step macOS workbench.