COS 130-10
Predicitng community composition from pairwise interactions

Thursday, August 13, 2015: 4:40 PM
344, Baltimore Convention Center
Jonathan Friedman, Physics, MIT, Cambridge, MA
Logan M. Higgins, Microbiology, MIT, Cambridge, MA
Jeff Gore, Physics, MIT, Cambridge, MA

The ability to predict the structure of complex, multispecies communities is crucial for understanding the impact of species extinction and invasion on natural communities, as well as for engineering novel, synthetic communities. This seems a daunting task, as community structure is dictated by a multitude of biotic and abiotic interactions. Often, the details of these interactions are unknown, rendering mechanistic modeling unfeasible. In such cases, phenomenological models, such as the classical generalized Lotka–Volterra (gLV) model, are employed. While a lot of our intuition comes from such models, their predictive power has rarely been tested experimentally. To test the predictive power of such phenomenological models, we used a laboratory-based community comprised of up to 8 heterotrophic soil bacteria as a model system. We measured the growth dynamics of each species in isolation, and the outcome of competition between all species pairs. We then used these measurements to predict the steady-state composition of complex communities composed of more than 2 species.


Pairwise competitions resulted in a diverse set of outcomes, including coexistence, exclusion, and bistability, and displayed evidence for both interference and facilitation. The performance of species in isolation (growth rate and carrying capacity) correlated with their competitive performance, but there were numerous pairs for which the competitive outcome was not in line with the individual species` performance. The majority of pair outcomes (>85%) could be captured by the gLV framework, and the composition of multispecies communities could be well predicted for communities composed solely of such pairs. Our results demonstrate the predictive ability and utility of simple phenomenology, but also its limitations: accurate predictions are possible without any mechanistic details, but only in cases that fit within the modeling framework. Additionally, the model parameters are valid only for the environment in which they were measured, and results are hard to extrapolate to novel environments.