Making Food Webs Out Of (Almost) Nothing At All
Food web reconstruction through phylogenetic transfer of low-rank network representation (2022) Strydom and Bouskila et al., Methods in Ecology and Evolution, https://doi.org/10.1111/2041-210X.13835
Understanding food webs (and more generally how different species interact) is important in helping us to understand ecological processes, but sampling (observing) interactions in the field is pretty challenging. Observing a parrot? Simple. Observing a possum? No problem. Observing a parrot evicting a possum from a tree-hollow? Rarer.
This means that data on species interactions is sparse. But we do have data for some regions, and things like computers and fancy maths (think machine learning) at our disposal. Which leads to the question: can we learn something from the places for which we do have interaction data and ‘transplant’ this knowledge and create an interaction network for a region with no data at all?
The focus here is to try and use predictive methods to help and at least give us a idea of who might potentially be eating who and use this to construct a metaweb (a full list of potential interactions) for a region that has plenty of species data, but no species interaction data.
What We Did
The primary goal was to use a transfer learning framework to predict the Canadian mammalian metaweb by using knowledge gained from European mammals. Transfer learning is a machine learning methodology that uses the knowledge gained from solving one problem and applies it to a related (destination) problem. You need a medium for transfer though, something to interpret how to translate knowledge from one system to the next. Here, we used phylogeny – how the different species are related genetically.
Specifically we used the known European network to learn about how species are interacting – in this case what determines who is eating who. Then we need to transfer this knowledge to the Canadian species pool by looking at the mammalian phylogeny. In this case we are assuming that species that are closely related are most likely going to be eating (or be eaten) in the same way. For example we know what the European lynx eats and can then infer that the Canadian lynx is likely to eat similar prey. This is summarised in the figure below – note in this case ‘latent generality traits’ refer to traits which make a species a predator and ‘latent vulnerability traits refer to traits that make a species prey.
What We Found
Our focus here was on methodology – if our predictions here were good, this technique could shed so much light on other ecosystems – so we were more concerned with how good our predictions were. In order to test our predictions we used some pairwise interaction datasets for regions within Canada to see our predictions matched up. Importantly these datasets were not used in the actual prediction framework and (for this modelling exercise) we had no prior known interaction data for Canadian species. Despite Europe and Canada sharing only 4% of their species we we are able to correctly predict 91% of the known interactions – which is pretty neat!
These results are so rewarding, but how far can we push the transfer learning framework before we ‘break’ it? In the example in the manuscript we were using two closely related regions. Europe is very similar to Canada in terms of both environment as well as species – at least in terms of sister taxa (thinkEuropean vs. Canadian lynx). So what happens when we look at areas that are a bit more different, like Australia? They have very different species (marsupials vs placental mammals) and environments and the model performance might not be up to scratch.
The other fun thing to think about is how to move from potential to realised interactions. As I said earlier we are predicting interactions that could potentially occur within Canada, but this doesn’t mean that they will actually happen in a specific point/sub-region within Canada. For example wolves will probably eat small rodents, but if there are enough larger prey about they will probably ‘ignore’ mice in favour of a more filling meal. So the next step is to ask ourselves how to refine our predictions enough to reflect these nuances.
Being able to have at least an approximation of what a food web may look for areas for which we previously had none is a big help if we want to start asking questions about more global patterns or behaviours of food webs. This transfer learning framework could be just such a tool to give us (what appears to be) reliable predictions from which to work with in our quest to fill in the global interaction map.
Tanya Strydom is a PhD candidate at the Université de Montréal, mostly focusing on how we can use machine learning and artificial intelligence in ecology. Current research interests include (but are not limited to) predicting ecological networks, the role species traits and scale in ecological networks, general computer (and maths) geekiness, and a (seemingly) ever growing list of side projects. Tweets (sometimes related to actual science) can be found @TanyaS_08.