Integrating Machine Learning-Based Pose Sampling with Established Scoring Functions for Virtual Screening

Abstract Classical docking methods have dominated the field of structure-based virtual screening (VS) for decades. Recently, several machine learning (ML)-based docking approaches have been introduced, presenting a promising avenue for advancing VS technologies. In this work, we report on the integration of DiffDock-L, one of the most promising ML-based pose sampling methods, into VS workflows by combining it with the established Vina and Gnina scoring functions. We assess the integrated approach regarding its VS effectiveness, pose sampling quality, and complementarity to classical docking methods, represented by AutoDock Vina. Our results on the DUD-Z benchmark data set show that pose sampling with DiffDock-L and AutoDock Vina yields comparable performance. In contrast, the choice of the scoring function has a decisive impact on VS success. In general, DiffDock-L generates physically plausible and biologically relevant poses in most cases, confirming it as a viable alternative to classical docking algorithms.