(142c) Design of Optimal Experimental Probes for Protein Dynamics Using Machine Learning and Variational Approach to Modeling Conformational Kinetics

In this talk, we plan to answer a broad question about design of experiments and validation of protein conformational ensembles, what is the minimum experimental information required for describing the conformational dynamics of a protein? Markov state models of protein dynamics describe the transitions between the conformational states of proteins using a transition probability matrix built from the raw simulation data. The eigenvalues and eigenvectors/eigenfunctions of the transition matrix provide the timescales and the population shift between states due to a particular dynamic process respectively. Considering only the m-slowest processes, an ideal set of experimental observations would unequivocally describe these m-slowest processes using only m-experimental observables. Our machine learning protocol involves using the variational approach for building Markov state models that provides approximation of the first m eigenfunctions of the molecular dynamics propagator using all possible experimental probe positions on the protein and picks the optimal set of probe locations that best describe the m-slowest processes. In this work, we combine two central ideas from machine learning and chemical physics - hyperparameter selection via cross-validation and variational approaches for linear operator eigenproblems; to create a new method for predicting observables that optimally discriminate between alternative models for molecular kinetics constructed from MD simulations. This protocol could also be used with a broad range of experimental techniques or a combination of experimental techniques. We present results from application of this protocol to three systems of biological importance, conformational change in membrane peptide transporter PepT_SO and β2-Adrenergic receptor, and folding of WW domain.