Tetrapeptide inhibitors of BACE-1 revealed by combined data-driven screening and physics-based free-energy refinement.

Alzheimer's disease remains a major global health challenge, and β-site amyloid precursor protein cleaving enzyme 1 (BACE-1) represents a critical therapeutic target as it catalyzes the rate-limiting step in amyloid-β production. While computational studies have extensively explored small-molecule BACE-1 inhibitors, large-scale screening combined with physics-based refinement has rarely been applied to peptide scaffolds. Here, we report an integrated machine learning (ML)-to- free energy perturbation (FEP) pipeline, combining XGBoost-based screening, molecular docking, replicate explicit-solvent molecular dynamics (MD), and absolute FEP calculations, as a multi-stage alternative to single-scoring approaches for tetrapeptide lead discovery against BACE-1. Screening a library of 16,000 tetrapeptides with an XGBoost model prioritized four candidates (HWRE, HWER, WHRR, and HWRQ) for structure-based evaluation. Molecular docking confirmed favorable positioning within the catalytic cleft, while replicate MD simulations revealed multivalent binding through hydrogen bonds, salt bridges to the catalytic dyad (Asp32/Asp228), and hydrophobic contacts with conserved pocket residues; interaction-fingerprint and hotspot analyses provided residue-level guidance for optimization. FEP calculations delivered quantitative thermodynamic ranking, separating three strong predicted binders (HWRE, WHRR, HWRQ; ΔG