Can co-occurrence networks predict plant-plant interactions in a semi-arid gypsum community?

Biotic interactions are important drivers shaping communities and their net effect can be difficult to assess when many species are involved or the study system is not well-known. Co-occurrence data are increasingly used to infer biotic interactions, but their accuracy remains poorly studied. We hypothesize that the predictions of biotic interactions based on co-occurrence networks are scale-dependent, being more accurate when co-occurrence is sampled at the spatial scale at which biotic interactions occur. We studied a plant community in a semi-arid gypsum environment, where plant-plant interactions result in a clumped distribution of the vegetation, which allows a straightforward identification of the scale at which interactions occur (i.e. vegetation patches). We used Bayesian network inferences to detect co-occurrence patterns at two spatial scales ( patches of about 0.05m(2) and plots of 2.25m(2)), and measured plant-plant interactions based on the effects of plant species on the establishment and recruitment of their neigbours. The co-occurrence network inferred at the patch scale did not reflect the plant-plant interactions observed, and the networks inferred at two different spatial scales showed contrasted topologies. Negative spatial associations between species predominated in the network inferred at the patch scale, partially due to a low species richness in vegetation patches. Meanwhile, positive relationships predominated at the plot scale. Relationships at the plot scale could reflect the final outcome of multi-specific biotic interactions, but without providing accurate information about pairwise interactions. Our results suggest that co-occurrence networks can be useful to generate hypothesis about the mechanisms structuring communities, but caution is needed in the interpretation of co-occurrence patterns in terms of biotic interactions, even when they are measured at apparently appropriate spatial scales.