Speaker
Description
A common goal in the theory of Chemical Reaction Networks (CRNs) is to design systems that reproduce or approximate a desired dynamics. This theory supports ongoing efforts in synthetic biology and molecular nanotechnology to emulate the functional molecular networks seen in nature. Here, we propose a molecular version of a recurrent artificial neural network, the RNCRN, which we prove is able to approximate arbitrary dynamics, and demonstrate that functionality for systems that exhibit multi-stability, oscillations and chaos. The neural net nature of the RNCRN means that it is particularly well suited to extrapolating a CRN, defined for all concentrations of the executive species, from a relatively small region of defined target behaviour. We show that this property allows the RNCRN to approximate exotic limit cycles that are not easily expressed as an attractor of a simple dynamical system. Similarly, the RNCRN can interpolate between two defined regions of qualitatively different target behaviour, automatically generating an appropriate bifurcation.
Bibliography
@article{dack2025recurrentneuralchemicalreaction,
title={Recurrent neural chemical reaction networks that approximate arbitrary dynamics},
author={Alexander Dack and Benjamin Qureshi and Thomas E. Ouldridge and Tomislav Plesa},
year={in press},
journal = {Cell Systems},
url={https://arxiv.org/abs/2406.03456},
}
@misc{dack2026recurrentneuralchemicalreaction,
title={Recurrent neural chemical reaction networks trained to switch dynamical behaviours through learned bifurcations},
author={Alexander Dack and Tomislav Plesa and Thomas E. Ouldridge},
year={2026},
eprint={2602.02374},
archivePrefix={arXiv},
primaryClass={q-bio.MN},
url={https://arxiv.org/abs/2602.02374},
}
