Plant–insect associations have been a significant component of terrestrial ecology for over 400 million years. Exploring these interactions in the fossil record, through novel perspectives, provides a window into understanding evolutionary and ecological forces that shaped these interactions. For the past several decades, researchers have documented, described and categorized fossil evidence of these interactions. Drawing on powerful tools from network science, we propose here a bipartite network representation of fossilized plants and their herbivore-induced leaf damage to understand late Paleozoic plant–insect interactions at the local community level. We focus on four assemblages from north-central Texas, but the methods used in this work are general and can be applied to any well-preserved fossil flora. Network analysis can address key questions in the evolution of insect herbivory that often would be difficult to summarize using standard herbivory metrics.

     Network Construction from Data.  We constructed weighted bipartite networks using plant–insect associations of four floras from the latest Pennsylvanian to middle Permian of north-central Texas, USA (Maccracken and Labandeira, 2020; Schachat et al., 2014; Schachat et al., 2015; Xu et al., 2017), where network nodes were either plant taxa (at the most highly or less resolved ranks) or herbivory damage types (DTs) (Labandeira et al., 2007). In such a network, edges represent interactions between the two nodes they connect, and are present only between nodes of the two different classes and not between those of the same class. Therefore, an edge can only be present between a DT and a plant taxon in our networks, and each edge denotes the presence of a particular DT in the plant taxon to which it is connected. The width (or weight) of an edge represents the number of occurrences of a given DT in a plant taxon. The width of each node denotes their respective number of occurrences. Each network is normalized according to its total number of DT occurrences and leaf specimens for each node-class.

     Network Analysis and Metrics.  After constructing the weighted bipartite networks, we used the packages bipartite (Dormann et al., 2008; 2009), vegan (Oksanen et al., 2013) and igraph (Csardi and Nepusz, 2006) to calculate and display various network statistics associated with the network structure at each locality. In this analysis, we suggest network measures used for bipartite networks in the context of fossil herbivory networks (Table 1). As detailed below, we categorize the measures into five groups, depending upon what they measure: centrality, co-occurrence, nestedness, robustness, and specialization, as well as the level of study. The level of study indicates whether the property being quantified pertains to a single node, a class of nodes (such as all DTs or all plants in our case), or of the entire network.

     Comparison with Modern Plant-Herbivore Interaction Networks.  In order to compare our reconstructed networks (Appendices S1, S2) with their modern counterparts, data was collected for 41 host plant–herbivore interaction networks, representing a variety of habitats and environmental gradients (Michalska-Smith and Allesina, 2019; Appendices S3–S8). Although the host plant class of nodes is same for both the fossil as well as current networks, the herbivory class is represented by DTs in the former, and by insect taxa in the latter. The same metrics were calculated that were explored for the DT networks of the four Paleozoic floras at whole network level and at each node class level. A principal component analysis (PCA) was performed of the network metrics obtained at different network levels (whole network and node-class levels) using the R package factoextra (Kassambara and Mundt, 2020) to document similarities among the localities and within different node-classes for the set of modern and fossil networks.

Use of the datasets are self-explanatory in the structure of the datasets.