Accurate long-timescale modeling and interpretability of molecular dynamics data

Determine computational methodologies that accurately model the long-timescale dynamics of molecular systems from high-dimensional molecular dynamics trajectories and develop representation and summarization approaches that render the resulting large-scale simulation data comprehensible to human analysts.

Background

Molecular dynamics simulations generate high-dimensional time-series data and often struggle to directly capture long-timescale processes due to sampling limitations and the complexity of biomolecular free-energy landscapes. This makes it challenging both to model kinetics over long times and to extract human-interpretable insights from the data.

Markov state models and variational approaches have advanced the analysis of such data, yet the paper explicitly notes that accurately modeling long-timescale dynamics and making the deluge of MD data understandable remain unresolved challenges. The authors propose SPIB as a step toward addressing these issues by learning continuous embeddings and adaptive metastable states, but the broader problems persist as open questions in the field.