Data-Driven Modeling of Biomolecular Dynamics under Physical and Experimental Constraints

Doctoral thesis, 2026

Biophysical experiments and biomolecular simulations report on the same processes that govern biological function, yet each provides only a partial view.
Experiments offer unbiased but sparse ensemble measurements, while simulations provide atomic detail but suffer from force-field biases and finite sampling. Integrating both into a single model is essential for a complete picture of biomolecular dynamics. This thesis develops the mathematical foundations for this integration through constrained statistical inference using the transfer operator framework and observables as the common language between experiments and simulations, comprising five papers in total. A survey chapter reviews machine learning approaches to molecular dynamics and identifies the absence of experimental consistency mechanisms as a common limitation. An empirical study then predicts dynamics-sensitive experiments from simulation data, revealing significant discrepancies that establish the need for integrative approaches. To address this, dynamic Augmented Markov Models (dynAMMo) are introduced, incorporating experimental measurements as constraints to recover slow exchange processes even from incomplete simulations. The Minimum-Excess-Work framework pushes the integration upstream, offering a method that incorporates stationary observables into probability flow generative models. Finally, applying dynAMMo to an engineered lipocalin demonstrates that these methods yield mechanistic insights that neither simulations nor experiments alone could provide. Together, these contributions progressively shift from comparing experimental and computational data, to correcting kinetic models post hoc, to guiding the generative process itself.