Thema der Dissertation:
Enhanced sampling methods for molecular systems: multiscale and data-driven techniques
Enhanced sampling methods for molecular systems: multiscale and data-driven techniques
Abstract: Simulations of molecular systems have led to significant discoveries in molecular biology. The high accuracy of these simulations enables us to understand biological functions on a molecular scale. While the applications for such simulations are countless, in practice it is only possible to simulate small systems due to computational limitations; reaching biologically relevant time- and length-scales is still beyond feasibility, even for the most powerful computers. This constraint is commonly known as the sampling problem. This thesis aims to provide new tools that help molecular simulations reach biologically relevant scales. It is split into two parts:
The first part provides new methods for rate computations in reactive systems, which can consist e.g. of a protein-ligand binding, oligomerization, or protein-protein association. The first method combines Markov state models of molecular kinetics with particle-based reaction-diffusion (PBRD) to generate a coarse-grained simulation of interacting molecules. Furthermore, a method is introduced to provide realistic parameters for PBRD simulations. It enables tuning the microscopic parameters of PBRD simulations such that experimentally obtained rates can be matched in the dilute limit.
The second part provides new methods based on Markov chain Monte Carlo. These can be utilized to speed up the computation of stationary observables. In biological systems, it is often observed that high barriers in the free energy landscape dramatically slow down the sampling process. To speed up computations, a whole range of methods has been developed. The latest advancements are facilitated by the recent rise of machine learning research, which provides new promising tools to approach the sampling problem from completely different angles. In this spirit a new method is introduced that aims for directly proposing transitions between regions of high populations in phase space, thus directly jumping over energetic barriers. A second proposed method is based on the recently developed Boltzmann Generators and aims to combine these with parallel tempering in order to speed up sampling significantly. To this end, a machine learning technique is employed which generates samples close to the Boltzmann distribution at different temperatures. In both of these methods, the convergence to the correct distribution is ensured by enforcing detailed balance.
Zeit & Ort
10.12.2021 | 10:00