Data set

Here we used a dataset of 44 sites of seasonal tropical forests in central-eastern Brazil (Fig. 1; Table S1 [suppl.]), with 22 semideciduous forest sites and 22 deciduous forests sites. Semideciduous forests are associated with the Atlantic Forest domain and between 50 and 70% of the forest canopy loses its leaves in the dry season (Neves et al. 2017). These sites are under a Köppen Cwa climate (subtropical with dry winters and rainy summers), with average annual rainfall between 1400 and 1500 mm and average monthly temperature between 19 and 20 °C. Deciduous forests are affiliated with the Caatinga domain (a nucleus of seasonally dry tropical forests - SDTF) in which more than 70% of the species lose their leaves in the rainy season (Pennington et al., 2009), and that is located in the southern Caatinga or in islands of fertile soil in Cerrado domain (Fig. 1). These sites are located under a Koppen Aw/As climate transition (tropical dry winter), with average annual rainfall between 750 and 1000 mm and average monthly temperature between 22 to 24.6 ° C (Maia et al., 2020).

Each site was sampled with a varied number of permanent sampling units (5 to 128 per site; total of 1077) with varying dimensions (400 m² in most cases, but a few with 225 and 300 m²) depending on local terrain features and fragment size. Total sampled area was 40.2 ha: 23.68 ha in semideciduous forests and 16.52 ha in deciduous forests. Within each sampling unit, we measured and identified to the species level all trees with a diameter at breast height (DBH; 1.30 m above the ground) ≥ 5 cm.

Plant identification followed APG IV (APG, 2016) and name standardization followed REFLORA (Flora do Brasil, 2020), using the flora package (Carvalho, 2016) implemented in the environment R v. 3.6.1 (2020). Forest inventory data are stored in the ForestPlots.net system (https://www.forestplots.net; codes presents in Table S1 [suppl.]) and are available upon request. These data were then used to generate two matrices of species occurrence on the sites, one for each vegetation type (semideciduous and deciduous), which were used in the analyses described below. The final matrices had 664 species occurring in 22 semideciduous forests and 433 species occurring in 22 deciduous forests.

 

 

Figure 1.  Location of the 44 sites of tropical seasonal forests used in this study in Brazil, South America and in relation to the main biogeographic regions in Brazil.

 

 

Association rule analysis

We explored patterns of species coexistence in each vegetation type (semideciduous and deciduous forest) using association rule analysis (ARA, Agrawal et al., 1993; Silverstein et al., 1998; Ferrarini & Tomaselli 2010; Rossi et al. 2014; Leote et al., 2020). ARA is a data mining tool that identifies associations between categorical observations (in this case, species names) in large data sets. It proposes relationship rules in the template: “if species X, then species Y”. Thus, based on a set of events (often called “transactions”) and the presence of elements in these events, it is possible to relate the existing categories and associate their coexistence with a probability of occurrence and importance (Ferrarini & Tomaselli, 2010; Rossi et al., 2014; Zumel et al., 2019; Leote et al., 2020). In the case of our data, the events (or transactions) are the sites of collection and the set of elements are all tree species present in each one.

Each rule has two parts: the rule left side (RLS) is a unique category level of reference (e.g., one species) that has an occurrence frequency in seasonal forest sites, which is called support; and the rule right side (RRS) that is composed by one or more category levels and that is related to the RLS in a frequency called confidence. Confidence values are always based on the rule support value and inform how frequent is the rule in the dataset (Ferrarini & Tomaselli, 2010; Rossi et al., 2014; Zumel et al., 2019; Leote et al., 2020). Thus, if for a given rule the RLS species occurs in half of the sites (support of 0.5) and the whole rule has confidence of 1.0, the RRS species are associated with the RLS species in all its occurrences, that is, they coexist in half of the sites. The rules obtained present three other measures: lift, coverage and Fisher’s exact test p-value. Lift is a measure of the rule’s importance and is obtained through the relation between the observed rule confidence and the previously expected confidence, in which lift > 1 indicate the rule is more frequent than expected by initial constraints and lift equal to 1 indicates that rule’s species are independent from each other and that the rule is not real. That is, more the value of lift, greater are the chances of having the sp 2 if the site already has the sp1. In addition, in cases where the lift has values less than 1, there is a negative relationship between the species, in which the occurrence of the first can negatively affect the occurrence of the second. Coverage is a measure of the rule’s frequency that evaluate the support of the species present in the RLS. Fisher’s exact test p-value is the rule’s significance obtained through Fisher’s exact test of contingency for small samples, and evaluate whether the rule is more significantly frequent than expected by chance (Hahsler, 2006). The full process of rules’ obtaining by applying constraints is summarized in Fig. 2.

Based on the occurrence data (presence/absence) of species in sites of each vegetation type (analysis were run separately for semideciduous and deciduous forests), we used the apriori algorithm (Borgelt & Kruse, 2002) from the arules package (Hahsler et al., 2020) in R v. 3.6.1 (2020) to obtain all pairwise association rules that met the following criteria: support equal to or greater than 0.63, which corresponds to the occurrence of the species in RLS in more than 60% of the sites (at least 14 sites of each vegetation type); and confidence of 0.8 for the rule. These criteria indicate that species on RRS must be related to the species on RLS in at least 80% of its occurrences. We focused on pairwise species associations aiming to reveal direct associations between them and analyze their similarities more closely. From the rules obtained, we selected those with significant Fisher’s exact p-value (<0.05), with coverage equal to or greater than 0.63 and lift greater than or equal to 1.1 (more than 10 % frequent than expected). By establishing these criteria, we were able to select strong combinations composed of species of wide occurrence, as well as to select associations of high importance that characterize the vegetation types.

Based on the selected rules, we explored the similarities between the coexisting species in each vegetation type. For that, we evaluated the compatibility between families of the coexisting species in selected rules of each vegetation type, as well as the similarities between them regarding two functional characteristics: potential size (DBH, cm), calculated as the 95th percentile of all DBH measurements of that species in the data set; and wood density (WD, g/cm³), extracted from a global database (Wood Density database; Zanne, 2009) that uses the average species WD value or the genus or family average when species-level data is unavailable. We chose these functional traits because they hint on species ecological behavior: wood density is associated with growth rate, mechanical resistance and hydraulic efficiency (Chave et al., 2009); while potential size is a proxy of resistance to canopy light and ability to reach advanced successional stages in forests (Falster & Westoby, 2005). With the functional data of the species in selected association rules (Table S2 [suppl.]), we thus quantified the difference between the functional traits of the species in each rule (left, RLS and right, RRS), evaluating the pattern of functional similarity between associated species. Graphics were made with the packages arulesViz (Hahsler, 2019) and ggplot (Wickham, 2016) for R program (R Core Team, 2020).

 

Figure 2.  Theoretical framework of the process of obtaining association rules between species using association rules analyzes (ARA). The scheme shows that from the set of species occurrence in the study areas, two limitations are initially proposed (support and confidence) for obtaining the initial rules. This set of initial rules then goes through a second selection process through the constraints Fisher’s p-value, Lift and Coverage, to finally obtain the final rules.

 

ResultsTop

Association rule analysis (ARA) with the established criteria resulted in 238 (out of 928 rules) significant and frequent pairwise species associations in semideciduous forests, and in 11 (out of 62) significant and frequent pairwise species associations in deciduous forests (Tables S3 and S4 [suppl.]). In the semideciduous forests, the pairwise associations are formed by a group of 33 species in 28 genera and 18 families. In the deciduous forests, the pairwise associations are formed by a group of 8 species in 8 genera and 5 families. In semideciduous forests, maximum support was 0.68 and maximum coverage was 0.68; maximum lift was 1.47 for the association between Dendropanax cuneatus (DC.) Decne. & Planch. (Araliaceae) and Annona dolabripetala Raddi (Annonaceae) (Table S3 [suppl.]; Fig 3 - a). In deciduous forests, maximum support was 0.68 and maximum coverage was 0.73; maximum lift was 1.21 for the association between Ptilochaeta bahiensis Turcz. (Malpighiaceae) and Cenostigma pluviosa (DC.) L.P. Queiroz (Fabaceae) (Table S4[suppl.]; Fig 3 - b). In both semideciduous and deciduous forests, some species pairs were mutually related, with occurrence dependence found in both directions; i.e., the association was consistent when swapping the species in the left and right sides.

Most of the pairs are formed by taxonomically unrelated species, at least as far as the family level (Table S2 [suppl.]). In the semideciduous forests, only 5.8% of all pairwise associations were between confamilial species. In the deciduous forests, Goniorrhachis marginata Taub. and Machaerium acutifolium Vogel. formed the only confamilial species pair (both belong to Fabaceae) (~ 9 % do total; Table S2 [suppl.]). In the rules selected for semideciduous forests, there are associations between species with different functional characteristics (Fig S1 [suppl.]), but most associations occur between species with similar potential size (DBH) and wood density (Fig S1 – a [suppl.]). In deciduous forests the trend seems to be the same, but the low number of rules selected prevents an accurate assessment (Fig S1 – b [suppl.]). In general, association rules tend to be formed by taxonomically unrelated species that are similar but not functionally equal.

 

Figure 3.  Representation of the first 20 association rules between pairs of species according to the lift values for semideciduous forests (a) and of the 11 selected association rules between pairs of species for deciduous forests (b). For deciduous forests, since only 11 rules emerged in this vegetation type, the figure represents all of them. Each circle represents an association rule and the arrows point to its participating species. Circle sizes are directly related to rule support values (the larger the circle, the higher the support), while circle colors are related to the rule lift value (the stronger the color, the greater the lift value). The arrow points the direction of co-occurrence between species, considering the reference species of the association rule. The situation in which two species have two arrows between them, each pointed in one direction, points out that co-occurrence occurs in both directions. In other words, species X associates with Y and Y associates with X.