Speakers
Description
Mechanistic differential equation models are widely used in biology and medicine, but have limited ability to handle incomplete system knowledge. Hybrid modelling approaches (also termed “Scientific Machine Learning") aim to address this by combining mechanistic differential equations with data-driven machine learning components. This minisymposium focuses on a conceptually simple, yet powerful, such approach: universal differential equations (UDEs). UDEs embed neural networks (or other universal function approximators) directly into differential equations and can be fitted using familiar parameter estimation workflows. By learning from data, the embedded network can capture unknown or highly complex dynamics that are difficult to represent with purely mechanistic models. In their simplest form, UDEs extend classical parameter estimation from fitting unknown parameters to learning unknown functions, such as protein production rates as functions of a transcription factor concentration. In more complex settings, they can represent substantial components of unknown system dynamics.
This minisymposium highlights the rapidly developing field of UDEs and their applications in biology. Following a brief introductory overview, it will feature presentations spanning systems biology, epidemiology, and ecology. The program also covers emerging software tools for training UDEs, as well as methodological advances including identifiability analysis and effective model-training strategies.
