Recently, many artificial intelligence (AI)-powered protein-ligand docking and scoring methods have been developed, showing high speed and accuracy. However, they often neglected the physical plausibility of the docked complexes and their performance in virtual screening (VS) projects. Therefore, we conducted a comprehensive benchmark analysis of four AI-powered, four physics-based docking tools, and two AI-powered re-scoring methods.We initially constructed DTEBV-D, on which the re-docking experiments reveal that KarmaDock and CarsiDock surpassed all physics-based tools on docking accuracy while all physics-based tools significantly outperformed AI-based methods on structural rationality. The VS results on DTEBV-D highlights the effectiveness of RTMScore as a re-score function and Glide-based methods achieved the highest enrichment factors (EFs) among all docking tools. We additionally constructed DRSM-D that more closely resembles real VS scenarios,where the employed AI-based tools obviously outperformed Glide. Finally, we proposed a hierarchical VS strategy that could efficiently and accurately enrich active molecules in real large-scale VS projects.

Please use the version v3 of the dataset, ensuring that the naming is consistent with the text. The v3 version can be found at https://zenodo.org/records/14874127.