Deciphering microbial community dynamics using cross-sectional data-informed NeuralODE

Understanding the ecological mechanisms of host-associated microbial ecosystems typically relies on either cross-sectional or time-series data. Cross-sectional analyses are limited in their ability to assess intervention effects, whereas time-series models require dense and informative sampling that is often impractical. Here, we present an enhanced Neural Ordinary Differential Equations (NeuralODE) framework that, for the first time, integrates cross-sectional data into the dynamic modeling of sparse and weakly informative temporal data. We develop two instantiations of this framework, tailored to relative and absolute abundances, and introduce a dynamic keystoneness metric to quantify species importance over time. Across simulated and real-data benchmarks, incorporating cross-sectional data improved performance over competing methods, particularly in data-scarce settings. Moreover, biological validation demonstrated that the framework recovers experimentally supported interactions and prioritizes identified influential species. Together, these results establish our method as a reliable framework for mechanistic modeling of microbial ecosystems, offering new insights into their dynamic behavior.