Speaker
Description
Inferring nonlinear interaction structures in complex microbial communities remains a central challenge in systems biology. Generalized Lotka-Volterra (GLV) models are a foundational framework for modeling interacting dynamical systems and are widely used in ecology. However, parameter estimation in GLV models is often hindered by the exploration of numerically unstable regions of parameter space, where certain parameter combinations generate rapidly diverging trajectories. This leads to numerical instability during integration, deteriorating the performance of optimization algorithms, particularly when broad parameter bounds are considered.
We propose a robust parameter estimation strategy that combines biologically informed initial bounds with an iterative re-optimization scheme. By restricting early searches to dynamically stable regions, the method avoids premature exploration of divergent trajectories.The feasible domain is then progressively adapted through successive re-optimization steps, enabling the search to gradually move toward regions providing high-quality fits.
We validate the approach on several case studies of increasing complexity. Notably, our strategy enables reliable calibration of a 12-species GLV model of the human gut microbiome, which was previously very challenging due to blow-up instabilities. We also illustrate the interplay between experimental design, practical identifiability and mechanistic insights from the inferred networks.
