An automated, open-source pipeline for anti-virulence compound discovery: multi-target molecular docking with Z-score specificity filtering

Background. Salmonella infections pose a severe threat to global public health and livestock production, while the antimicrobials used to control them drive the emergence of antimicrobial resistance (AMR). Anti-virulence compounds are an attractive alternative, and structure-based virtual screening is an efficient route to find them. In practice, however, docking screens rank candidates by absolute predicted docking score, so their output may be dominated by uninhibited ligands that bind many unrelated protein surfaces. Open software that selects on binding specificity rather than on absolute docking score is not readily available to microbiologists. Results. We present an automated, open-source pipeline that couples multi-target molecular docking to a two-tier statistical specificity filter. The pipeline retrieves and curates compound libraries from PubChem by chemical abstracts service (CAS) numbers, with automatic desalting and size-based exclusion of polymers. Also, it builds three-dimensional ligands at a user-defined protonation state with Open Babel. It docks every ligand against every member of a protein panel with Smina, centering each search box on a pocket detected by fpocket and converts the resulting absolute docking score matrix into per-compound Z-scores, retaining only pairs that are both a strong absolute docking score (≤ −8.0 kcal mol⁻¹) and selective by standardized relative docking score (Z ≤ −1.5, cross-panel standard deviation > 1.5). Applied to 381 generally recognized as safe (GRAS) food additives and the Salmonella-associated structures, it returned 22 compound-target pairs. Riboflavin ranked highest, with S = −10.3 kcal mol⁻¹ and Z = −2.6 against the type III secretion system (T3SS) chaperone SigE. Conclusions. Ranking on standardized relative docking score rather than on absolute docking score gives a transparent way to remove uninhibited binders from a virtual screen. The pipeline is modular, configurable for any ligand library and protein panel, and openly released. The candidates it returns are computational predictions and require experimental validation.