Published at Sep 7, 2026
Author: Marvin van Aalst
From September 7-11, Ljubljana will host the 22nd conference of the International Study Group for Systems Biology (ISGSB 2026), bringing together researchers working at the interface of theory, model, and experiment. Dr. Marvin van Aalst will represent CPBL with a talk titled “Making sense of time series with universal differential equations.”
Mechanistic models based on ordinary differential equations (ODEs) are a cornerstone of systems biology, but their predictive power is limited by our ability to specify every model term a priori. When biological mechanisms are partially unknown or too complex to parameterize from first principles, the gap between model and data grows, and classical fitting approaches cannot close it. Universal differential equations (UDEs) offer a principled way to bridge this gap: a UDE replaces unknown or uncertain model terms with trainable function approximators, typically neural networks, embedded directly within an ODE system. The mechanistic scaffold is preserved, the unknown parts are learned from time series data, and the result remains interpretable in biological terms.
The talk demonstrates this approach across two use cases from plant systems biology. In a model of sugar metabolism, UDEs recover intracellular metabolic fluxes directly from concentration time series, without requiring explicit rate laws — the learned flux functions reproduce known regulatory features and generalize across experimental perturbations, providing a data-driven complement to traditional kinetic modeling. The second case applies UDEs to identifying the parameters governing pulse amplitude modulation (PAM) fluorometry protocols. PAM measurements are widely used to characterize photosynthetic state, but the relationship between instrument protocol and observed signal depends on parameters that are difficult to determine independently. By embedding the unknown protocol response as a learnable term within a mechanistic photosynthesis model, these parameters can be recovered directly from experimental fluorescence traces.
Together, the two examples illustrate a general workflow: start with a mechanistic ODE model, identify the poorly constrained terms, replace them with neural network surrogates, and train on time series data. The result retains the interpretability of a mechanistic model while gaining the flexibility to learn where prior knowledge is insufficient — a practical entry point into hybrid modeling for systems biologists already working with ODE-based models and time series data.
We wish Dr. van Aalst an inspiring conference and look forward to hearing about the discussions in Ljubljana!