A roadmap towards predicting species interaction networks (across space and time)

Abstract

Networks of species interactions underpin numerous ecosystem processes, but comprehensively sampling these interactions is difficult. Interactions intrinsically vary across space and time, and given the number of species that compose ecological communities, it can be tough to distinguish between a true negative (where two species never interact) from a false negative (where two species have not been observed interacting even though they actually do). Assessing the likelihood of interactions between species is an imperative for several fields of ecology. This means that to predict interactions between species-and to describe the structure, variation, and change of the ecological networks they form-we need to rely on modelling tools. Here, we provide a proof-of-concept, where we show how a simple neural network model makes accurate predictions about species interactions given limited data. We then assess the challenges and opportunities associated with improving interaction predictions, and provide a conceptual roadmap forward towards predictive models of ecological networks that is explicitly spatial and temporal. We conclude with a brief primer on the relevant methods and tools needed to start building these models, which we hope will guide this research programme forward. This article is part of the theme issue 'Infectious disease macroecology: parasite diversity and dynamics across the globe'.

Keywords: biogeography; deep learning; ecological forecasting; ecological networks; machine learning.

Figure 1.

Proof-of-concept: an empirical metaweb (from Hadfield et al. [46]), i.e. a list of co-occurrences within a species pool, is converted into latent features using probabilistic PCA, then used to train a deep neural network to predict species interactions. Panels (a) and (b) represent, respectively, the receiver-operating-characteristic curve and the precision-recall curve, with the best classifier (according to Youden’s J) represented by a black dot. The expected performance of a neutral ‘random-guessing’ classifier is shown with a dashed line. Panel (c) shows the imputed using t-distributed stochastic neighbour embedding (tSNE), and the colours of nodes are the cluster to which they are assigned based on a k-means clustering of the tSNE output. Empirical interactions are shown in purple, and imputed interactions in grey.

Figure 2.

A conceptual roadmap highlighting key areas for the prediction of ecological networks. Starting with the input of data from multiple sources, followed by a modelling framework for ecological networks and the landscape, which are then ultimately combined to allow for the prediction of spatially explicit networks.

Figure 3.

The nested nature of developing predictive and forecasting models, showcases the forward problem and how this relies on a hierarchical structure of the modelling process. The choice of a specific modelling technique and framework, as well as the data retained to be part of this model, proceeds directly from our assumptions about which ecological mechanisms are important in shaping both extant and future data.