Predicting seed-dispersal interactions across the Afrotropical Ficus-frugivore network using machine-learning

Symposia title
Inferring Seed Dispersal Interactions Using Machine Learning

Abstract title
Predicting seed-dispersal interactions across the Afrotropical Ficus-frugivore network using machine-learning

Background and objectives
Climate change caused by natural and anthropogenic sources impacts the survival and range of species, potentially causing network-rewiring due to loss and gain of new interactions. Species networks are complex, and researchers rarely have an exhaustive list of all possible interactions in a network, let alone in the novel networks that will be emerging. Artificial intelligence, specifically machine learning classifiers have previously been successful in inferring species interactions using phenotypic traits.

Methods
We developed a machine learning approach based on observed interactions, species traits, and phylogeny to infer potential interactions in a continental scale seed-dispersal meta-network. Specifically, we focused on seed-dispersal interactions between frugivorous mammals and birds that consume and disperse Figs (Ficus) in the Afrotropics. From 920 published studies, we built a database of 5201 empirical interactions between 110 fig species and 604 frugivore species, and collected phylogenies and >100 functional traits of both figs and frugivores.

Results
Our Random Forest model trained on observed interactions, species traits and phylogeny was able to classify interactions with high fidelity. Using 3100 total mammal and bird species and the 112 fig species in Africa, we predicted interactions between currently co-occurring species, and additional interactions between species that could potentially overlap under future range redistributions. We found distinct spatial patterns in the number and nature of inferred interactions and that using only observed interactions underestimates connectance and robustness towards extinction of local networks.

Conclusions
Predicting trophic interactions using machine learning is a novel way of investigating the sensitivity of species communities at a continental scale and how functional traits drive species interactions. It will provide insights into how species range shifts driven by global change drive extinction velocity of both plants and frugivores and may lead to potential network rewiring where interactions are lost, and new ones might be gained.