Alumni-Preis 2025

Betreuer: Prof. Dr. Gerhard Stock

Abstract of thesis:

Molecular dynamics simulations are an important method to understand fundamental processes of protein dynamics by providing a detailed spatiotemporal description in terms of Cartesian coordinates of each atom. This high-dimensional data poses challenges in extracting the relevant information. Therefore, dimensionality reduction methods are used to transform the essential information of a system into a low-dimensional representation that captures the most relevant dynamics. The resulting free energy landscapes allow for an easy understanding of the underlying physics as it helps to find pathways, state models, or transition states.

In this thesis, we investigate autoencoder methods for dimensionality reduction that go beyond well-established linear methods such as principal component analysis. These methods can recognize non-linear and, in parts, temporal relationships and furthermore enable generative modeling. To this end, we study the dynamics of the protein models Ala2, Aib9, and T4 lysozyme. Their different size and dynamical behavior place different demands on the architecture and mathematical formulation of the models. We find that these variations of the model architecture focus on different aspects of the free energy landscape of the systems. Moreover, we discuss various issues of the training process including the instability and the large number of parameters of the models. We show that the use of graph-based networks is advantageous for processing molecular dynamics data in autoencoders because a graph can be seen as a natural representation of a protein. Since the model knows about the protein structure when using a graph, this results in a better estimate of the free energy landscape and adds robustness to the model. Additionally, we follow a new direction by applying a model that incorporates temporal dependencies into the dimensionality reduction.