We are taking an evolutionary approach to predicting ecological interactions, e.g. food webs. We seek to answer questions such as: given a list of organisms in a community, what are the possible interactions among them? How might interactions change with addition or deletion of nodes or changes in link strengths? We have compiled the largest existing meta-database of predator-prey interactions and, for the first time for many older datasets, mapped nodes to the most appropriate scientific names. This allows us to use weighted taxonomic and phylogenetic distances or N-dimensional trait space to predict trophic relationships where no other data is available. Using our Food Web Constructor, we find that a simple heuristic using taxonomic distance recovers almost 50% of known trophic links in food webs, on average, significantly better than a simple database search. Not surprisingly, larger food webs - typically more recent -- are more difficult to predict. We can demonstrate that including older food webs in the database improves the precision of predictions (55% of predicted links are correct compared to 39% when older webs are not used), even though they are usually smaller and less well-resolved taxonomically than more modern studies.
Future work will involve using databases of organismal and environmental characteristics, taking explicit advantage of machine-learning techniques to classify interactions, adapting phylogenetic algorithms (e.g. ancestral node reconstruction) for estimating missing values, and using optimality criteria such as maximum likelihood or parsimony or graph-based measures to constrain whole network solutions. Do these methods produce biologically realistic networks with similar graph-theoretic indices and compartments? These methods can produce hypotheses for networks to be used in more extensive systems models, available now only for well-studied ecosystems such as the Everglades. They can also be used to explore host-parasite or symbiotic interactions, as well as interactions among non-organismal entities. They are alternatives to graph-theoretic approaches and to fluid flow approaches. Public access to the online tools is available at http://spire.umbc.edu. This work is part of the Spire project (Semantic Prototypes in Research Ecoinformatics, PI Tim Finin).