Artificial intelligence models trained on data from molecular dynamics trajectories

Artificial intelligence (AI) and machine learning (ML) models trained on molecular dynamics (MD) trajectories are increasingly used for predicting various properties reported from MD simulations. The reliability of such models depends critically on the quality and reproducibility of the underlying simulations. A frequent assumption in MD research is that longer single simulations necessarily yield more accurate results. In practice, unless mechanistic insights are the primary goal, multiple shorter, independent simulations are likely to provide a more representative sampling of highly complex potential energy surfaces, often characterized by numerous conformational states and local minima. These considerations are especially important for AI models, where biased or poorly sampled training data can propagate errors and reduce predictive power. This commentary discusses best practices in MD simulation design for training AI models that depend on MD data.