Nonlinear machine learning of macromolecular folding and self-assembly

High performance computation and sophisticated machine learning algorithms have emerged as new tools for studying biological, physical and chemical systems at the atomistic scale. In this thesis, I report several applications of molecular dynamics simulation and machine learning in the study of the macromolecular folding and assembly. In the first aspect, I employ molecular simulation and non-linear manifold learning to explore the dynamics and configuration of linear and ring polymers. Integrating statistical mechanics with dynamical systems theory, I establish a means to determine single molecule folding funnels from univariate time series in experimentally accessible observables. In the second aspect, I utilize coarse grained molecular simulation to explore the self-assembly of hundreds of asphaltene molecules over micro-second time scales to reveal the aggregation phase behavior as a function of temperature, pressure and solvent conditions. I then employ graph matching and non-linear manifold learning to obtain asphaltene folding and assembly free energy landscapes. This thesis establishes new fundamental understanding of the folding and assembly of macromolecules, builds connections between computer simulation and experimental measurements, and provides new routes to the rational design of functional molecular materials