Protein dynamics and its prediction using machine learning

In the first part of this dissertation, chemical exchange and the use of R1ρ rotating frame relaxation experiment for its study are discussed. Beginning with the Bloch-McConnell equations, new expressions are derived for the spin relaxation rate constant in the rotating frame, R1ρ for chemical exchange between two or more sites that have distinct magnetic environments and Larmor frequencies. The 2-site results are accurate provided that the spin relaxation decay is dominated by a single exponential damping constant and are applicable to a wider range of conditions than existing theoretical descriptions. The n-site results are accurate when the population of one of the sites is much greater than that of others. The second part of this dissertation discusses the prediction of protein backbone flexibility using machine learning. Using a data set of 16 proteins, a neural network has been trained to predict backbone 15N generalized order parameters from the three-dimensional structures of proteins. The average prediction accuracy, as measured by the Pearson correlation coefficient between experimental and predicted values of the square of the generalized order parameter is 71.4%. The network parameterization contains six input features. Predicted order parameters for non-terminal amino acid residues depends most strongly on local packing density and the probability that the residue is located in regular secondary structure.