Development of selected mesoscopic physical models with the aid of machine learning methods and their applications in studies of molecular systems

This dissertation is concerned with the development and application of unsupervised machine learning methods in the field of theoretical biophysics and bioinformatics. The machine learning approach offers a powerful framework for extracting and purifying valuable information from large, multi-dimensional sets of data generated in simulations and experiments of biomolecular systems. It is not, however, the case that ready-made machine learning methods offer infallible means of dealing with all sorts of complex, and partially chaotic data encountered in computational biophysics and structural biology. Large portion of this work is devoted to the adaptation of unsupervised machine learning techniques to our particular purposes. In this dissertation, we employed unsupervised machine learning strategies dealing with two problems arising in theoretical biophysics and bioinformatics. The first problem was the identification of quasi-rigid structural parts in proteins, whereas the second one was devoted to discovery of internal cooperation of molecular subsystems that propels a conformational transition. Both problems involved dynamical properties of molecular systems, and the analyses presented in this dissertation allowed for a simplified description of these phenomena. We demonstrate how the unsupervised machine learning approach can help in explaining intricacies hidden within seemingly chaotic molecular dynamics simulation data. The methods developed in this thesis increase our ability to understand complex molecular phenomena. But we also point out potential problems associated with applying unsupervised machine learning algorithms in the field of molecular biophysics.