Inferring species interactions from observational data is one of the most controversial issues in community ecology. Part of the problem is that an interaction between one pair of species can ripple through an ecological network and produce surprising indirect consequences. For example, a species could indirectly harm its mutualists by attracting their natural enemies. In this talk, I show how methods from statistical physics and machine learning can enable ecologists to make quantitative predictions about the landscape-level consequences of positive and negative species interactions. The proposed models, called Markov networks or Markov random fields, provide a rigorous and well-specified method for predicting the types of communities that would arise from any possible matrix of pairwise species interactions. Just as importantly, Markov networks’ parameters can be estimated from observational data by maximum likelihood. This represents a major advance over previous methods, based on null models, which do not provide estimates of species’ pairwise interaction strengths.
For this presentation, I generated pairwise species interaction matrices with known structure and used them to produce simulated communities of positively- and negatively-interacting species. I then evaluated several methods for detecting the sign and strength of species’ pairwise interactions. In all cases, existing methods for detecting species interactions from co-occurrence data (based on null models or covariance patterns) had unacceptably high error rates. Treating the simulated communities as Markov networks and estimating species’ interaction terms by maximum likelihood worked much better, even when the number of species was relatively large and the simulated communities were also affected by exogenous factors such as environmental heterogeneity. Topic Selection: The role of biotic interactions in structuring species distributions: synthesizing across ecological disciplines and spatial scales in the face of climate change