Preview

PDF, English Download (20MB) | Lizenz: Predicting Hydrogen Atom Transfer in Collagen by Riedmiller, Kai underlies the terms of Creative Commons Attribution - ShareAlike 4.0

Abstract

Molecular Dynamics (MD) simulation is an established method for studying bi- ological systems and materials at the molecular level. For example, it can predict the effects of mutations on the folding behavior of proteins, or elucidate the binding mechanisms of drugs to receptors. To achieve this, a molecular system is modeled as beads connected via springs or sticks. Goal of this thesis is to develop a method to deepen our understanding of mechanoradicals undergoing hydrogen atom transfer (HAT) reactions. Usually, no new chemical bonds can be formed during an MD simulation, prohibiting the analysis of HAT and other reactions. In this thesis, a method called KIMMDY is implemented, enabling reactive MD through the use of kinetic Monte Carlo steps. While it has previously been used for homolysis reactions, here its scope is extended to also allow HAT during an MD simulation. First, a data set of HAT reaction barriers is crafted using quantum mechanical (QM) calculations. Then, a graph neural network (GNN) is trained on this data set to predict HAT reaction barriers. The inputs to the GNN are structures sampled from MD simulations, requiring no prior optimization. Thus, the model offers a significant speed-up compared to traditional methods, like reactive force fields, or direct QM calculations. Furthermore, the KIMMDY 2.0 software package is presented, which can perform kinetic Monte Carlo driven reactive MD simulations in a flexible and user-friendly manner. It supports HAT and homolysis reactions, and can be easily extended to any reaction for which rates are available. For HAT, they are obtained using the afore- mentioned GNN. However, this approach is not limited to HAT. Through the use of machine learning models, reaction rates for arbitrary reactions become accessible, only requiring a limited amount of QM calculations during training. In KIMMDY, these models enable combining the accuracy and flexibility of QM calculations with the efficiency of MD.