Decoding the limits of deep learning in molecular docking for drug discovery

Structure-based molecular docking, a cornerstone of computational drug design, is undergoing a paradigm shift fueled by deep learning (DL) innovations. However, the rapid proliferation of DL-driven docking methods has created uncharted challenges in translating in silico predictions to biomedical reality. Here, we delve into the performance and prospects of traditional methods and state-of-the-art DL docking paradigms—encompassing generative diffusion models, regression-based architectures, and hybrid frameworks—across five critical dimensions: pose prediction accuracy, physical plausibility, interaction recovery, virtual screening (VS) efficacy, and generalization across diverse protein–ligand landscapes. We reveal that generative diffusion models achieve superior pose accuracy, while hybrid methods offer the best balance. Regression models, however, often fail to product physically valid poses, and most DL methods exhibit high steric tolerance. Furthermore, our analysis reveals significant challenges in generalization, particularly when encountering novel protein binding pockets, limiting the current applicability of DL methods. Finally, we explore failure mechanisms from a model perspective and propose optimization strategies, offering actionable insights to guide docking tool selection and advance robust, generalizable DL frameworks for molecular docking.