Answers are grounded in this paper and cite their sources.
Computational methods for virtual screening and lead optimization are benchmarked on specifically designed data sets to predict compound potency. Here we present a workflow and two resulting new benchmark data sets, BALLROOM A and BALLROOM S, from combining PDB and ChEMBL data. Due to high data quality standards, the formation of molecule pairs measured under the same conditions only, and avoiding pairs with close to equal potency, we achieve a reliable data collection for physics-based bioactivity prediction based on protein–ligand complex structures. When testing state-of-the-art scoring functions for docking, no classical method statistically outperforms a simple molecular-weight-based prediction, consistent with previous studies such as D3R. While challenging for physics-based methods, the prediction task can be easily addressed by simple nearest-neigbor scorers showing that the benchmark cannot be used for machine learing based scoring functions due to significant train-test inference.
No takes yet. Share an insight, caveat, or question.
Gutermuth et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: