Speaker
Description
The automated discovery of dynamical digital twins from time series data, known as model learning, is a central challenge in systems biology, particularly in the presence of noise, partial observability, and limited data availability.
We recently conducted a comprehensive review of available methods for data-driven discovery of dynamical systems, and identified 117 algorithms based on either symbolic or sparse regression, which we evaluated across eight key biological and methodological challenges \cite{metayer2026data}.
We then propose a novel method based on sparse Bayesian inference that jointly estimates the structure and parameters of dynamical models while incorporating biological prior knowledge. The approach is specifically designed to integrate multi-condition data, as commonly encountered in biological experiments.
The method was first validated on simulated systems, where it exhibits robust performances in recovering interactions under multiple noise levels. It was then applied to circadian gene expression data from human lung cells to automatically infer the dynamical interactions between the clock and the innate immune receptor NLRP3. The ambition is to provide interpretable representations of the underlying biological processes and enable the exploration of system perturbations.
Overall, this work contributes to the development of automated pipelines for constructing biologically grounded digital twins, with potential applications in personalized medicine.
Bibliography
@article{metayer2026data,
title={Data-driven discovery of digital twins in biomedical research},
author={M{\'e}tayer, Cl{\'e}mence and Ballesta, Annabelle and Martinelli, Julien},
journal={Briefings in Bioinformatics},
volume={27},
number={1},
pages={bbaf722},
year={2026},
publisher={Oxford University Press}
}
