Abstract
We present MBPert, a generic computational framework for inferring species interactions and predicting dynamics in time-evolving ecosystems from perturbation and time-series data. In this work, we contextualize the framework in microbial ecosystem modeling by coupling a modified generalized Lotka-Volterra formulation with machine learning optimization. Unlike traditional methods that rely on gradient matching, MBPert leverages numerical solutions of differential equations and iterative parameter estimation to robustly capture microbial dynamics. The framework is assessed within the context of two experimental scenarios: (i) paired before-and-after measurements under targeted perturbations, and (ii) longitudinal time-series data with time-dependent perturbations. Extensive simulation studies, benchmarking on standardized MTIST datasets, and application to Clostridium difficile infection in mice and repeated antibiotic perturbations of human gut micribiota, demonstrate that MBPert accurately recapitulates species interactions and predicts system dynamics. Our results highlight MBPert as a powerful and flexible tool for mechanistic insight into microbiota ecology, with broad potential applicability to other complex dynamical systems.