Relevant publications that I have authored or co-authored during my PhD.
Sofia Sartore, Daniel Nagel, Georg Diez, Gerhard Stock
The Journal of Physical Chemistry Letters (2026)
Compressing complex, high-dimensional data into a simpler summary is one of the most common operations in data analysis — but it almost always loses information, and sometimes entire structures in the data disappear entirely. In this paper we show that for dynamical and stochastic systems, fast, random fluctuations blur the boundaries between distinct states before any compression happens. The fix is straightforward: apply a smoothing filter to the high-dimensional data first, separating the slow and meaningful trends from the fast noise — a classical signal-processing idea applied in an unusual place. On a protein folding dataset, this simple preprocessing step recovered ten times more distinct, well-separated states compared to the standard approach.
Georg Diez
PhD Thesis, University of Freiburg, Faculty of Mathematics and Physics
Proteins are molecular machines that constantly shift shape — and these shape changes determine how they function. Capturing this motion produces enormous, high-dimensional datasets that are hard to interpret directly. In my PhD thesis, I designed machine learning methods that encode physical knowledge to extract clear, interpretable models of protein behavior.
Georg Diez, Nele Dethloff, Gerhard Stock
The Journal of Chemical Physics, 163, 124105 (2025)
Most machine learning methods for compressing high-dimensional data treat observations as independent, ignoring the order in which they were collected. For time-series data, this throws away valuable information: not just what the system looks like at each moment, but how quickly it is changing. In this paper, we address this problem by embedding a probabilistic model of temporal structure directly into our neural network architecture, so the compressed representation is aware of dynamics instead of just snapshots. As a proof of concept, we first show on an analytical example that we are able to perfectly separate two geometrically indistinguishable states that differ only in their dynamics. Applied to a protein that switches between two known shapes, the method succeeds in identifying a third, previously unknown state with a distinct biological function — one that every conventional approach had missed entirely.
Daniel Nagel, Georg Diez, Gerhard Stock
The Journal of Chemical Physics, 161, 054108 (2024)
The standard way to measure whether two variables are related only works well for simple, linear relationships. Real-world data is often messier: variables can be multidimensional and their dependencies nonlinear. Mutual information is a more general, model-free alternative, but it comes with a catch: without normalization, its values have no natural scale — they are hard to interpret and impossible to compare across datasets. This paper introduces a nonparametric estimator based on local k-nearest-neighbor statistics that fixes this by normalizing the result in a principled way, so comparisons are meaningful regardless of the data's shape or dimensionality.
Georg Diez, Daniel Nagel, Gerhard Stock
Journal of Chemical Theory and Computation, 18(8), 5079–5088 (2022)
When a dataset has thousands of variables, most of them capture noise rather than the phenomenon of interest. Before any modeling can begin, you need to know which variables actually matter — a problem known as feature selection. Here, we introduce MoSAIC, a method that solves this automatically: it measures pairwise correlations across all variables, then uses a community-detection algorithm to group the ones that genuinely move together, discarding the rest. The result is a clean, interpretable subset of variables that reflects real structure in the data and allows you to focus on the processes that really matter. The method has a single tunable parameter, scales to large datasets, and is released as an open-source Python package with a scikit-learn–compatible API.
Matthias Post, Benjamin Lickert, Georg Diez, Steffen Wolf, Gerhard Stock
Journal of Molecular Biology, 434(17), 167679 (2022)
Proteins regulate their activity by transmitting signals across their structure — a small local change at one site triggers a large rearrangement far away, much like a mechanical switch. Using large-scale computer simulations and data-driven correlation analysis, we mapped exactly how this signal travels and identified which internal components participate. The key finding is that the switch is all-or-nothing: an entire network of contacts must change simultaneously, driven by random thermal fluctuations. This provides a concrete picture of how rare, large-scale events emerge from local noise — a pattern that appears across many complex systems beyond biology.