NextTopDocker: the largest-to-date docking power benchmark reveals that deep learning performs generally much worse than logistic regression models

Predicting three-dimensional binding orientations of drug-like molecules remains challenging in structure-based drug design. Following AlphaFold’s success, end-to-end machine-learning (ML) docking frameworks have emerged, leveraging deep learning for pose generation. Yet, their performance is often assessed on small, outdated benchmarks containing only a few hundred protein-ligand complexes. We present “NextTopDocker,” a large, up-to-date, open-access data set for docking power assessment comprising 14,038 training and 5,201 test entries across 3,173 unique protein targets, constructed from the Protein Data Bank using a cold-ligand split. Developed entirely with open-source tools, it includes crystallographic structures, Smina-generated docking poses, and ligand-similarity-aware training subsets. With “NextTopDocker,” we benchmarked four state-of-the-art end-to-end ML docking frameworks, DeepDock, Interformer, SurfDock, and Uni-Mol Docking v.2, against classical (Smina) and hybrid (GNINA 1.3 and logistic regression using Smina and GNINA 1.3 scores) baselines. Despite sophisticated architectures, only Interformer matched the docking power of logistic regression on Smina poses, while others failed without post-docking physics-based refinement. Most raw ML-generated poses exhibited steric clashes and/or implausible geometries, underscoring the lack of intrinsic physical realism in current ML-based docking. Embedding physics-informed constraints directly into pose generation, rather than relying on downstream optimization, is essential for achieving valid, autonomous molecular docking. “NextTopDocker” is available at https://github.com/caominhtr/NextTopDocker and https://zenodo.org/records/17492994 .