Assessment of molecular dynamics time series descriptors in protein-ligand affinity prediction.

The advancement of computational methods in drug discovery, particularly through the use of machine learning (ML) and deep learning (DL), has significantly enhanced the precision of binding affinity predictions. Despite progress in computer-aided drug discovery (CADD) accurate prediction of binding affinity remains a challenge due to the complex, non-linear character of molecular interactions. Generalizability continues to limit these models, with performance discrepancies noted between training datasets and external test conditions. This study explores the integration of molecular dynamics (MD) simulations with ML to assess its predictive performance and limitations. In particular MD simulations offer a dynamic perspective by depicting the temporal interactions within protein-ligand complexes, potentially bringing additional information for affinity and specificity estimates. By generating and analyzing over 800 unique protein-ligand MD simulations, we evaluate the utility of MD-derived descriptors based on time series in enhancing predictive accuracies. The findings suggest specific and generalizable features derived from MD data and propose approaches to augment the current in silico affinity prediction methods.