Speaker
Description
Rare transitions between metastable states in stochastic gene regulatory networks are difficult to resolve in practice, as direct stochastic simulations require large amounts of data to capture these events reliably. In this talk, we present a neural network–based approach (ISOKANN) to learn low-dimensional reaction coordinates that describe the slow dynamics of such systems \cite{SikorskiCapturing, yousefian2025exploring,sikorski2024learning}.
These coordinates provide both a coarse-grained, dynamical view of the system and a basis for adaptive sampling. Starting from unbiased simulations, computational effort is iteratively focused on transition regions by resampling along the learned coordinate. This substantially reduces the amount of simulation time required while improving the resolution of transition pathways. At the same time, the learned representation captures the dominant slow processes in a dynamically consistent way.
We illustrate the approach on prototypical stochastic gene regulatory models and discuss how combining representation learning with adaptive sampling enables an efficient analysis of rare event dynamics. The emphasis is on intuition and practical insights rather than technical details.
Bibliography
@incollection{SikorskiCapturing,
url = {https://doi.org/10.1515/9783111376776-004},
title = {Capturing the macroscopic behavior of molecular dynamics with membership functions},
booktitle = {Mathematical Optimization for Machine Learning: Proceedings of the MATH+ Thematic Einstein Semester 2023},
author = {Alexander Sikorski and Robert Julian Rabben and Surahit Chewle and Marcus Weber},
publisher = {De Gruyter},
address = {Berlin, Boston},
pages = {41--58},
doi = {doi:10.1515/9783111376776-004},
isbn = {9783111376776},
year = {2025},
lastchecked = {2026-03-23}
}
@article{sikorski2024learning,
title={Learning Koopman eigenfunctions of stochastic diffusions with optimal importance sampling and ISOKANN},
author={Sikorski, Alexander and Ribera Borrell, Enric and Weber, Marcus},
journal={Journal of Mathematical Physics},
volume={65},
number={1},
year={2024},
publisher={AIP Publishing}
}
@inproceedings{yousefian2025exploring,
title={Exploring Metastable Dynamics of Gene Regulatory Networks with {ISOKANN}},
author={Yousefian, Maryam and Donati, Luca and Sikorski, Alexander and Weber, Marcus and R{\"o}blitz, Susanna},
booktitle={International Conference on Computational Methods in Systems Biology},
pages={126--149},
year={2025},
organization={Springer}
}
