Ó³»­´«Ã½

Abstract:

Microbial community dynamics are typically modeled using either generalized Lotka-Volterra (gLV) models, valued for their interpretability, or machine learning approaches, favored for their predictive performance. Consumer-resource models offer an attractive mechanistic alternative but have been largely confined to theoretical studies. This leaves a significant gap: no existing framework combines mechanistic interpretability with strong predictive performance in a data-driven setting. To bridge this gap, we previously developed the neural species mediator (NSM), which embeds a consumer-resource model within a neural network to achieve both improved accuracy and interpretability. The NSM model has demonstrated strong predictive performance and the ability to recover meaningful species–metabolite interactions from synthetic microbial community data. However, limitations remain: the model relies on phenomenological assumptions, requires absolute abundance measurements, and cannot directly infer species–metabolite interactions when metabolite dynamics are unobserved. In this talk, I will present our recent efforts to address these limitations by incorporating physical conservation constraints into the NSM and expanding its applicability to a broader range of experimental settings. Our key advances include replacing fitted parameters with physically motivated constraints, enabling direct inference of metabolic interactions from community time-series data alone, and extending the framework to compositional data. Throughout, I will highlight how embedding conservation constraints in hybrid models can expand the range of experimental settings in which mechanistic, interpretable models of microbial community dynamics can be applied.

Biography:

Pak Lun Kevin Cheung is a PhD candidate in Electrical and Computer Engineering at the University of Wisconsin–Madison, currently working with the Venturelli Lab at Duke University. Kevin has collaborated with Dr. Ophelia Venturelli since his undergraduate studies at the University of Wisconsin–Madison and began his PhD under the mentorship of Dr. Paul Milenkovic, where he built a foundation in nonlinear systems and optimal control. He now brings together these disciplines in his current research, focusing on computational and mathematical approaches for understanding and designing microbial ecosystems. He applies tools from dynamical systems, Bayesian optimization, and machine learning to develop predictive, interpretable models and to enable efficient design-learn-test cycles for shaping community structure and function.