Year: 2026
Type: preprint
Abstract: Motivation: Machine-learning scoring functions are typically trained on static crystallographic structures, although protein-ligand recognition is inherently dynamic. It remains unclear whether MD derived conformations provide transferable information or instead shift the training data away from experimentally observed binding geometries. Results: We generated 20 ns molecular dynamics trajectories for 2,502 protein–ligand complexes and evaluated five scoring function architectures trained on crystallographic structures and MD-derived frames using four different sampling protocols. Across the CASF-2016 and CASF-2013 affinity benchmarks, MD-only training generally reduced predictive performance, whereas augmenting crystallographic structures with MD data improved four of the five architectures. Averaged across models, benchmarks, and sampling protocols, augmentation increased PCC by 0.072 and reduced RMSE by 0.165 relative to MD-only training. A nearest-neighbor analysis showed that the benefits of augmen-tation were not restricted to complexes similar to frames, arguing against a simple similarity-based explanation. The results suggest improvements were connected with exposing models to controlled structural perturbations around experimentally observed binding poses. Overall, MD conformations were most effective as augmentation rather than replacement for crystallographic training data. Availability: MD frames used for model training are available at: (https://zenodo.org/ records/20443988).
Source: ChemRxiv
Authors: Jakub Poziemski, Pawel Siedlecki
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7213981516