Parts | Days | Selection | Search | Updates | Downloads | Help
BP: Fachverband Biologische Physik
BP 1: Systems and Network Biophysics
BP 1.10: Talk
Monday, March 18, 2024, 12:30–12:45, H 0112
Inference of dynamical networks in biology with recurrent neural networks — •Pablo Rojas1, Marie Kempkes1, Claudia Arbeitman1,2, and Martin Garcia1 — 1Theoretical Physics, University of Kassel, Kassel, Germany — 2CONICET, Argentina
The inference of networks in dynamical systems is crucial to the mechanistic understanding of complex systems. Biological systems often rely on the emergent behaviour resulting from the interaction of multiple units. Inferring the existence of links between nodes of a network from measured time series is an inverse problem that becomes more complex under the presence of non-equilibrium, strong nonlinearity, noise, delays and large number of interacting nodes - frequent conditions in experimental time series. In this work, we present a method that uses recurrent neural networks to learn the underlying connections in a dynamical system from its multivariate time series. A key aspect of the method is that it does not assume a mathematical model, i.e. equations, defining the dynamics of the network. Thus, the method is model-free, which makes it applicable to a broader range of systems. We apply this method to a range of biological systems under far-from-ideal conditions to evaluate its performance. We sketch a comparison against other neural network architectures to showcase its advantages.
Keywords: network inference; biological networks; recurrent neural networks; reservoir computing; time series analysis