Abstract
Interest in machine learning models based on protein-ligand interaction fingerprints as target-specific scoring functions has been increasing for their applications in drug discovery. Recognizing the critical role of decoys in model performance, our research investigates various decoy selection strategies to enhance machine learning models based on the Protein per Atom Score Contributions Derived Interaction Fingerprint PADIF. We explored three distinct workflows for decoy selection and compared these with real inactive compounds: random selection from extensive databases like ZINC15, leveraging recurrent non-binders from high-throughput screening (HTS) assays stored as dark chemical matter, and data augmentation by utilizing diverse conformations from docking results. Active molecules from ChEMBL, combined with these decoy approaches, were used to train and test different machine learning models based on PADIF. Our findings reveal that models trained with random selections from ZINC15 and compounds from dark chemical matter closely mimic the performance of those trained with actual non-binders, presenting viable alternatives for creating accurate models in the absence of specific inactivity data. Furthermore, all models demonstrated an enhanced ability to explore new chemical spaces for their specific target and improve the top active compound selection over classical scoring functions, thereby boosting the screening power in molecular docking. This study emphasizes the importance of selecting appropriate decoy sets in the development of machine learning models for virtual screening, suggesting pathways for future research in optimizing decoy usage and improving model accuracy and reliability in drug discovery applications.