Behind the paper: why we cannot tell different types of species interaction networks apart

April 22, 2024 PLOS Biology Behind the paper Biology Ecology PLOS Biology

Image credit

In this ‘behind the paper’ post, Chris Brimacombe discusses why it is so difficult to study the structure of species interaction networks.

My name is Chris Brimacombe and I am PhD candidate at the University of Toronto in the department of Ecology and Evolutionary Biology, with Dr. Marie-Josée Fortin. Generally, I work on the application of network theory to ecological data, but I am also interested in many different ecological questions, which can generally be tackled using common mathematical approaches, e.g., calculus and linear algebra. While I would like to take all the credit for our recent paper, it was genuinely a team effort, with an especially large contribution by Dr. Korryn Bodner. As a whole, all authors of the paper (including Matthew Michalska-Smith, Timothée Poisot, and Marie-Josée Fortin) have a genuine deep interest in understanding what the main factors are that determine the structure of species interaction networks. You may ask, “How hard can that be to figure out?” Well, given the difficulty of simply identifying a single species with all of their interactions in nature, you can imagine how difficult it is to put all of this information together for many species in order to construct a network.

When most ecologists talk about creating species interaction networks, we mean representing a group of species in a single object where species are nodes and their interactions are edges. For instance, in plant–pollinator networks — as the name suggests — each plant and pollinator are modelled as separate nodes, and whenever a pollinator touches a plant during some period in which an observer is watching, the interaction is recorded as an edge between these organisms’ nodes. Of course, there are many more types of species interaction networks, but some of the most common are seed-dispersal, plant–ant, plant–herbivore, and host–parasite. A keen reader may recognize that these networks are ‘bipartite’: where interactions occur only between species of opposite sets (e.g., no plant–plant or pollinator–pollinator interactions in a plant–pollinator bipartite network).

Toy example of how a plant–pollinator bipartite network is constructed. Here, pollinators (i.e., butterfly and bee) are represented as orange nodes, plants (i.e., white flower and purple flower) are represented as green nodes, and edges between nodes are plant–pollinator interactions (i.e., whenever an observer sees a pollinator touching a flowering plant). Arrows point to the nodes that each respective organism represents.

For many reasons, knowing what a species interaction network looks like is important, perhaps most of all since these networks can characterize the health of modelled ecosystems. Given that a species’ interactions are often lost prior to their extinction, knowing what their interactions are has some obvious benefits.

Due to the practical constraints of sampling species and their interactions in nature to create species interaction networks, freely available networks from previous studies are often used to test new hypotheses about species interaction networks. These networks are available, since researchers, in their generosity, have put them up on online repositories after they have built them and finished their own investigations. However, while all of these networks represent species interaction networks, we must always remember that they have been constructed by researchers in many locations across the globe, and reflect observations captured within specific sampling periods.

Originally, the idea for this paper was to use those freely available networks from online repositories to identify different types of species interaction networks using only their structure. This reflects our underlying beliefs that the same type of species interaction networks should all share a somewhat similar structure that makes them unique and identifiable. After all, if biological/ecological properties influence how organisms interact, surely plant–pollinator networks should look more similar to each other than host–parasite networks, and vice versa. The benefit of this being true would mean that just by seeing the structure of a network, we would be able to identify what type of species interaction network it was, without any information about the species or interactions that were embedded in it.

Toy examples of both a successful (left) and unsuccessful (right) attempt at identifying/separating plant–pollinator bipartite networks from host–parasite bipartite networks using only their structure. Each network has been quantified via multiple different structural measurements, and projected as a multi-dimensional object, as defined by their multiple structural measures, down into two dimensional space.

As a researcher, I am drawn to problems that are easy to state. In this vein, identifying species interaction networks based on their structure seems so simple, almost as if it was a homework problem in an undergraduate course. How could I possibly not be equipped to answer this; after all, I have been in university a long time … I must be able to solve this problem, right?

However, as I am sure the reader can predict, no matter the method that we tried to apply in order to identify different types of species interaction networks, we always failed. Although this result can now be summed up in a single sentence, we took an unreasonably large amount of time over the course of a few years to show it. While failing to identify different types of species interaction networks is not what we wanted, we reasoned that these findings made sense, given what we can reason about the data. Namely, since these networks are created by many different researchers using many different sampling strategies as well as network construction methods, and species themselves are exposed to many different environmental/biological drivers, of course we were unable to find defining structural features of networks from the same system. In short, the network data are just too messy. See below for one example of the type of data differences we found between plant–pollinator networks.

Two plant–pollinator bipartite networks each from a unique publication source, obtained from online repositories, and their corresponding matrix representations. We highlight in red distinct node differences between each network, defined by the researchers who built them: (left) plants are only those from a single genus Psychotria, and (right) pollinators include unidentified species. Each yellow square in the matrix corresponds to a plant–pollinator interaction between the organisms named in the respective row (plants) and column (pollinators).

The most surprising (but potentially unsurprising) result we found was that networks from the same publication were much more structurally similar to each other than networks from different publications. In many instances, one publication will contain multiple networks (all from the same system, e.g., plant–pollinator) which often occurs when researchers investigate how species interaction networks change across environments or time. Publication being a significant determinant of network structure was particularly reassuring for us, in that we felt that this provided evidence that we were approaching the problem in a way that made sense. Since networks from the same publication are generally built in parsimonious ways and under parsimonious conditions, it is only logical that networks from the same publication would appear structurally similar. Unfortunately, we cannot determine what the precise causes of these publication effects are — are they due to similar environmental/biological drivers, or sampling effects, or network construction methods? In reality it could be any or all of them given that each of these factors confound network structure within a publication in cryptic ways.

Apart from the results, the most important lesson I learned during this study is to look at the data carefully. As stated earlier, there is a prevalent view in the ecological literature that species interaction networks are structured according to the system they represent. While this may be true in theory, we were unable to show this in the empirical species interaction networks available to us. I think it becomes more obvious why we would be destined to fail from the outset when we tried to show this once one takes a look at the network data. Specifically, networks are built so differently from one another (at least those from different publications) that they have structural distinctness when compared to almost all other networks, implying a lack of possible structural cohesion between those from the same system.

The next step for our team is to investigate if food webs show the same sort of relationship we found with the networks used in this study. In contrast to the bipartite species interaction networks used here, food webs do not have such stringent limitations on how modelled species can interact, and interactions can occur between any organisms represented in the web/network. Given that our proposition is only as good as that data that supports it, for us, testing if food webs from the same publication are also structurally similar to each other is the next step needed to refine our understanding of the structure of species interaction networks. Or who knows … maybe this time food webs from the same biological system (e.g., terrestrial and aquatic) will be most structurally similar.

About the author

Chris Brimacombe is a PhD candidate at the University of Toronto. He previously undertook undergraduate research at Western University in the department of Applied Mathematics, and the University of Ottawa in the department of Biology.

Leave a Reply Cancel reply