Pair-wise seed dispersal interaction data is scarce for Tijuca National Park. Therefore, to build seed dispersal networks, we first gathered plant and animal species occurrence data in Tijuca (ICMBio 2008) and excluded all plant species that were not dispersed by animals based on our natural history knowledge. Our network includes 124 frugivorous species, including the four reintroduced species, and their 472 animal-dispersed plant species. Then, to estimate the pairwise interactions between plants and seed dispersers, we used the Atlantic-Frugivory dataset (Bello et al. 2017), which compiles pairwise seed dispersal interactions across the Atlantic forest. Because the yellow-footed tortoise is the only seed disperser in our dataset that had no interaction recorded in the Atlantic-Frugivory dataset, we made an extensive literature review on Chelonoidis spp. fruit consumption to infer its diet (see details in Supplementary Material, Tables S1 and S2). Finally, we assembled seed trait data by compiling information available in the Atlantic-Frugivory dataset and conducting a literature search in Google Scholar for ‘“Plant species” (seed OR fruit) (diameter OR length)’ and complemented with publications that we knew of but did not appear in our search. Thus, data about species’ interactions was gathered from the literature a posteriori to the reintroductions.

Interaction networks

We used four different models to infer pairwise interactions in our network. Therefore, we built four types of networks based on different assumptions about which interactions would occur in Tijuca.These four different approaches to network estimation allow us to investigate how sensitive our estimates of network patterns were to different assumptions when assigning interactions, and if they were consistent regardless of how networks were estimated. Thus, these network models take in account the uncertainty about interactions occurrence, reducing the potential biases of each assumption and verifying if our results hold independently of the model used.

For all networks, we assumed that interactions recorded elsewhere in the Atlantic Forest would also occur in Tijuca if the same plant and animal species were present in Tijuca. However, because some interactions might be absent in the literature records, we added interactions based on the following assumptions:

Phylogenetic model: in this network there are two sets of nodes and each node depicts one species. We assumed that, if an animal species was recorded interacting with a plant of a given genus, this animal would interact with all plant species of the same genus found in Tijuca;

Genus-level model: in this network, there are two sets of nodes. Animal species are depicted as nodes, whereas each node in the second set depicts a plant genus. Thus, plant taxa were collapsed to the genus level, which would reduce the effects of under-sampled interactions within genus, but leading to a description of network patterns at a different taxonomic level;

Probabilistic phylogenetic model: we assume that there is a probability p that a given animal species i - that interact with a plant species j from the genus k – would also interact with another plant l from the same genus k. We assume that p is fixed to all potential pairwise interactions and we assume p = 0.5. Thus, in this network, if one interaction between an animal species i and plant genusj occurred, the same animal iended up interacting with a fraction of plant species of genusj as well. We built one hundred pairs (with and without reintroduced species) of these networks to estimate the mean effects of reintroduced species on network structure (see below the network structural patterns studied).

Seed size model: we used plant species’ seed length and seed diameter data to construct a network. We first used empirical data on species interactions to record the maximum and minimum seed length and diameter of fruits consumed by each animal species, with no taxonomic constraints. We then assumed that each animal species interacts with all plant species in the network with seed length and seed diameter values in between the maximum and the minimum observed for plants. We assumed that if one of the seed dimensionswas not within the interval formed by the minimum and maximum value of seeds dispersed by the animal, then the animal did not interact with the plant, i.e., the interaction is forbidden by this seed dimension (Allesina et al, Jordano et al. 2003).

This network model strongly minimizes the effects of sub-sampling since understudied species with an extremely low number of recorded interactions are connected with a wide range of plants in the network, given that they can consume fruits/seeds of broadly different sizes. Having said that, this model may overestimate the number of interactions and is limited to a subset of plant species (n = 249 species) in our database with known values of seed diameter and length. In this network 9animal species end up having no interaction partners because either they interacted only with plants with unknown values of seed size or there were no plant species that had seed length and diameter simultaneously compatible with the interacting range of the animal species. We ran network analyses with and without these non-interacting species but there were no significantly differences in results (S. Table4).

Several networks were built to take in account the uncertainty about interactions occurrence, to reduce the potential biases of each assumption and to check if our results hold independently of the model used.