Speaker
Description
Bacteria interact in different ways, from competition to cooperation, depending on resource availability in the environment.
Identifying environments that foster these dynamics is both important and challenging, as the outcomes often depend on subtle alignment between the dynamic metabolic needs of species.
A common community approach to this challenge has been metabolic modeling.
Genome-scale metabolic models can identify interactions between species, but they are computationally expensive when applied to large communities.
In fact, metabolic modeling often cannot fully capture the discrete nature of species, as these models typically optimize the total biomass of the system rather than the behavior of individual organisms.
We propose a simpler and more tractable alternative, the Metabolic Bacterial Game (MetaBGame), in which bacteria strategically consume or produce environmental compounds according to their metabolic needs.
The framework employs multi-agent reinforcement learning and can operate under perfect or imperfect information, reflecting whether agents fully observe or only infer each other’s metabolic states.
Our results demonstrate that in MetaBGame we can scale and learn communities up to thousand agents in 34 hours. On top of that, our simple design allows easily identifying competitive and cooperative strategies. MetaBGame is implemented in JAX, allowing scaling across different hardwares, including CPUs and GPUs.
