Mechanistic microbial ecosystem model inference: A new method to analyze time series data

Modern molecular observational technologies (e.g. sequencing, probes) can produce high-resolution time series of microbial concentrations. The dynamics in those data hold important clues about the ecology and interactions of the various system components (phytoplankton, DOM, bacteria). These datasets are typically analyzed using empirical methods (e.g. correlation networks), but those provide limited mechanistic insight (e.g. a network may show interactions between two viruses) and are of limited use for making predictions. Mechanistic approaches do not suffer from these limitations, but are generally not applied to big data sets, maybe because of the large number of processes and components.

Here we introduce mechanistic microbial ecosystem model inference (MMEMI), where the model structure, including components, processes and parameters are optimized within mechanistic constraints to match observed dynamics. The model allows for any number of components that interact via a several processes, including exudation, photosynthesis, grazing, heterotrophy and others. The optimization problem is challenging because of the many components, non-linear interactions and feedbacks, which is handled using a method that mimics natural speciation or diversification (e.g. stepwise increase in the number of phytoplankton species).

The MMEMI method is applied to a four-year, daily resolution marine microbial time series, including observations for 20 phytoplankton and 40 bacteria types. The resulting mechanistic model includes additional 30 POM and DOM types. The following results are discussed: (a) Model calibration (years 1-3) and validation (year 4); (b) Predicted interactions, including specific phytoplankton > DOM > bacteria linkages, and comparison to auxiliary data (e.g. functional gene contents); (c) Predicted mass fluxes between ecological compartments and to the deep ocean; and (d) Functioning under a warmer future climate.