Speakers
Description
Dynamical digital twins stand out as powerful mechanistic tools to advance biomedical challenges. However, their design involves the integration of numerous multimodal time-resolved datasets calling for a change of practice, from manual design to automated discovery. The development of data-driven model learning approaches is constrained by the realities of biological data: scarcity and heterogeneity of time-resolved datasets available across multiple conditions, partial observability, and large candidate network spaces (1).This minisymposium aims to provide a coherent overview of modern strategies for inferring reaction networks' structure and dynamics under these constraints, clarifying how modeling assumptions and prior knowledge affect identifiability and robustness. Across four talks, we will cover a broad range of modeling formalisms, including Bayesian dynamical-system learning (2), sparse identification for tochastic delay differential equations (3), and equation discovery in complex systems with chaotic regimes (4). Critically, we will emphasize lessons from real-data applications, and discuss benchmarking practices needed to make automated digital-twin inference reliable and reproducible.
