Molecular dynamics frames for machine-learning scoring functions: evaluating binding affinity prediction and virtual screening

This dataset is associated with the study []. We examine how structures derived from molecular dynamics (MD) simulations influence machine learning scoring functions. 

The dataset contains a curated collection of MD frames selected using four protocols capturing different types of conformational variability: cluster, ligand_2A, pocket_low and combined. They are described in the study. 

The dataset augments the crystallographic training data with MD frames and is used to train and machine learning scoring functions for binding affinity prediction and for screening enrichment evaluation. The dataset allows to study how frame selection and structural distribution affects predictive performance of different models.

The archive is organized by PDBID protein-ligand complex identifiers. Every complex directory contains four subdirectories, one for each frame-selection protocol. Each protocol subdirectory contains up to ten representative MD frames.