Fruits were collected at Ilha do Cardoso (25°03′ S and 48°53′ W), a pristine land-bridge subtropical island, located in the south of São Paulo state, Brazil. The study site is a 15 100 ha protected island composed exclusively of Atlantic rainforest, including lowland and highland tropical rainforest, mangroves, dune vegetation and restinga forest (Barros et al. 1991). The climate at Ilha do Cardoso is generally warm and wet throughout the year, with a mean annual rainfall of roughly 3000 mm and average ambient temperature of 23.8 °C (Funari, Struffaldi-De-Vuono & Salum 1987; Castro, Galetti & Morellato 2007).
We collected quantitative information on fleshy fruit traits for 105 vertebrate-dispersed species belonging to 44 families (see Appendix S1 in Supporting Information). The sample was dominated by trees (59 spp.), followed by shrubs (29 spp.), vines (eight spp.), herbs (five spp.), hemiparasites (two spp.) and bromeliads (two spp.), and represents c. 50% of the known vertebrate-dispersed plants in the area (E. Cazetta unpublished data). Fruits from these species are mainly dispersed by birds (74 spp.), followed by mixed agents (birds and primates as well as other mammals; 15 spp.) and primates and other mammals (16 spp.) (Appendix S1), which is a common pattern found in vertebrate-dispersed plants in the Atlantic forest (Almeida-Neto et al. 2008).
Fruit phenotypic traits
For each species, we collected c. 30 ripe fruits (range 8–50) from at least three individuals per species (for some species, we obtained fruits from up to five individuals). From these samples, we obtained data on fruit morphology, chemical composition and colour. Due to logistic constraints, sample sizes varied depending on the type of measurement: morphological variables were measured in all species (n = 105 spp.), whereas fruit colour and chemical composition were obtained from smaller subsets of the total (n = 58 spp. and 78 spp., respectively). Therefore, we report the total number of species or the degrees of freedom of each analysis.
Morphological traits were measured on 30 fruits per species. For each fruit, we quantified the following variables: length and diameter of the fruit and its seeds, fruit fresh mass, pulp fresh and dry mass, seed mass and number of seeds per fruit. From these measurements, we also calculated the relative yield of fresh fruits (dry mass of the pulp/fresh mass of whole fruit) and the pulp to seed ratio (dry mass of pulp/mass of the seeds).
Colour measurements and contrast calculation
To estimate fruit colour, we measured the reflectance spectra of 15 fruits per species, as well as 10 leaves and 10 accessory structures (nongreen pedicels, capsules) if present. We performed all measurements with an Ocean Optics USB2000 spectrometer and a Top Sensor System Deuterium-Halogen DH-2000 (both Ocean Optics, Duiven, the Netherlands) as a standardized light source. Fruit reflectance was measured as the proportion of a standard white reference tile (Top Sensor Systems WS-2). We calculated three variables that characterize reflectance spectra: total brightness, chroma and hue (Endler 1990). Brightness is the intensity of reflected light from a given surface and is measured as the cumulative sum of the light intensity reflected (R) between 300 and 700 nm. Variation in brightness is perceived on the scale of white to black. Humans (and presumably other primates as well) perceive chroma and hue as chromatic aspects of visual information. Chroma (or saturation) describes the colours’ similarity to a neutral grey; colours that contain no or very little grey are deeply saturated, whereas a grey with a little tint of colour has a low saturation (see Kelber, Vorobyev & Osorio 2003 for a thorough review). Chroma is calculated as reflectance variation between 300 and 700 nm ((Rmax−Rmin)/Raverage). In contrast, hue defines colour differences that are related to human colour categories, for example red, yellow and green. While there is no evidence that animals other than primates perceive hue and chroma as humans do, some experiments on chickens suggest that they can categorize colours in a way similar to humans (Jones, Osorio & Baddeley 2001). Hue (or spectral shape) is measured as the wavelength of maximal reflection (λ(Rmax)).
The visual system of our two most important seed disperser groups, birds and primates, differs with birds being superior in discriminating colour differences (Håstad & Odeen 2008). We selected the three variables that characterize colour properties that primates can extract from visual stimuli and that are also likely to be perceived by birds because they have superior visual abilities. To account specifically for avian vision, we calculated additionally chromatic contrasts between the mean reflectance of fruits and leaves or secondary structure according to eye models of avian vision (Vorobyev & Osorio 1998). The model first calculated the quantum catch of each class of single cones presented in a tetrachromatic visual system (LWS, MWS, SWS and UVS), based on analytical approximation of cone visual pigments and oil droplet spectra. The quantum catches are used to find relative contrasts against fruits and background. Contrasts were characterized in units of ‘just noticeable differences’ (jnd), following the model proposed by Vorobyev & Osorio (1998). One jnd is the threshold of discrimination; higher values indicate discrimination of two colours. For example, we found on Ilha do Cardoso that increasing fruits’ contrasts also increases fruit detection by birds up to a threshold of 40 jnds (Cazetta, Schaefer & Galetti 2009). We opted to employ this model because most dispersers in Ilha do Cardoso are birds, and our previous study showed that birds attend primarily to chromatic contrasts to locate fruits in this community (Cazetta, Schaefer & Galetti 2009).
To estimate fruit chemical composition, the pulp or aril from fruits was frozen at −18 °C following collection, until the moment of analysis. For each fruit sample, we ran three replicates and used the mean of the three values for statistical analyses. The following variables were estimated from chemical analyses: protein composition, lipid content, total soluble sugar content and secondary compounds (phenols).
Proteins were determined with the Kjeldahl method, which consists in estimating total nitrogen in the fruit as a surrogate for total protein (protein composition was calculated by multiplying total nitrogen by 6.25) (Jeffery et al. 1989). This conversion factor may overestimate the total amount of protein found in fruits, but it does not alter conclusions on the covariance between protein contents and fruit visual traits. Lipid content was determined employing the standard macro-gravimetric method (Bligh & Dyer 1959). Total sugar content (glucose + fructose + sucrose) was estimated with gas chromatography–mass spectrometry (GC–MS), because the interference of coloured material in fruit pulps or arils prevented the use of more standard colorimetric methods (e.g. the Anthrone method; see the study by Pooter & Villar 1997).
To determine the content of secondary compounds (phenols), we employed the extraction method proposed by Price & Butler (1977), employing butanol and methanol extracts (see details in the study by Schaefer, Schmidt & Winkler 2003). Although not all phenolic compounds have similar extraction probabilities, this index of phenolic compounds correlated negatively in our community with the consumption of fruits by birds and fruit pathogens alike (Cazetta, Schaefer & Galetti 2008), demonstrating that this index is biologically meaningful. We focused on phenols because they commonly occur in fleshy fruits, and they are the most widespread secondary compounds in ripe fruits (Buren 1970; Herrera 1982; Cipollini & Stiles 1992; Foley, McLean & Cork 1995). The contents of these compounds were analysed with photometric measurements (for details, see the study by Cazetta, Schaefer & Galetti 2008).
We reconstructed the phylogeny of plant species from Ilha do Cardoso with the Phylomatic software, employing the Phylomatic conservative supertree (Webb & Donoghue 2005). We analysed the phylogeny of fruit traits to evaluate to what extent our results on trait covariances are influenced by phylogeny. It was not our principal aim to study the evolution of fruit traits given that our sample did not comprise an adequate and balanced sample within the phylogeny. Because estimates of divergence times were not available, we tested four different arbitrary branch lengths: constant branch lengths (all branch lengths = 1), Grafen (1989), the variant Nee’s (Purvis 1995) and Pagel (1992). Statistical adequacy of the different branch lengths were tested by plotting the absolute values of standardized phylogenetic independent contrasts against their standard deviations, and branch lengths were considered statistically adequate if there was no correlation between these values. In essence, this means that there is no consistent bias as a function of divergence times during the calculation of contrasts (i.e. contrasts are uniformly distributed across long and short branches); hence, the standardization procedure can be considered adequate for statistical purposes (Garland, Harvey & Ives 1992). These analyses were performed with PDTREE (available in the PDAP package; Garland et al. 1993), employing log10-transformed phenotypic values to improve normality, except for sugar content which was square-root-transformed. Because we were interested in performing analyses employing a single phylogeny, we selected Nee’s arbitrary branch lengths transformed with Grafen’s rho of 0.5 (see Results), which was statistically adequate for all studied phenotypic variables.
To investigate whether fruit morphology, chemistry and colour of the plant species in Ilha do Cardoso show significant phylogenetic signal, and whether this signal differs between these suites of characters, we employed the randomization procedure of the matlab program Physig.M developed by Blomberg, Garland & Ives (2003). This method consists in shuffling tip values across the real phylogeny and obtaining the distribution of variance in contrasts under this null model. If the variance of the real data set is lower than 5% of the values obtained from the randomizations (1000 in our study), we rejected the null hypothesis of no signal with P < 0.05. We also compared the k-statistic obtained in each analysis, which describes the amount of phylogenetic signal in the continuous-valued traits and allows for comparisons across phylogenetic trees and phenotypic traits (Blomberg, Garland & Ives 2003). A value of 1 indicates exactly the amount expected under Brownian motion evolution along the specified phylogeny (topology and branch lengths). Values smaller than 1 indicate less phenotypic resemblance between relatives than expected and larger than 1 indicate stronger than expected phylogenetic clumping of trait values.
We performed Pearson’s correlations to test for pairwise associations between visual cues and estimates of fruit quality and defence to document the covariance among phenotypic fruit traits. Importantly, animals may combine a set of indicators rather than base their decisions during foraging on pairwise correlations only. Multivariate analyses may thus be more adequate. Therefore, we analysed the relation between macronutrients and phenols employing a model comparison approach. In this context, we were primarily interested in determining the association between fruit morphology, colour and nutrients, and we evaluated whether visual cues such as colour and morphology are good indicators of fruit quality, as judged by its chemical composition. While regular correlation and regression analyses are adequate to test whether dispersers can rely on visual cues to select better fruits, phylogenetically informed tests are generally more adequate to address if traits at different levels of organization have coevolved in response to selection.
Subsequently, we used the matlab program RegressionV2.M (Lavin et al. 2008) to perform linear statistical models via both ordinary least squares (OLS) (i.e. nonphylogenetic) and phylogenetic generalized least square regressions (PGLS) (Garland & Ives 2000; Garland, Bennett & Rezende 2005). Ordinary least square regressions assume that the unexplained residual variation is independent among species, while PGLS assumes that residual variation among species is correlated, with the correlation given by the degree of phylogenetic relatedness between species. We used both approaches to understand whether relationships among fruit traits are evolutionary conserved (and thus better explained by PGLS models) or evolutionary labile (and thus better explained by OLS models).
Regression models included chemical compounds (lipids, sugar, proteins and phenols) as the dependent variables and fruit morphology and colour as independent ones. Because morphological traits were highly correlated, we first performed a diagnostic principal component analysis and selected the variables that explained most of the variance in our data based on the score values (data not shown). Therefore, the morphological variables we selected were fruit diameter, seed number and relative yield. We included these variables even though the relative yield of fruits or the number of seeds is not involved in signalling. However, fruit consumers may learn that some fruits are relatively unrewarding and thereby use colour or other external stimuli (fruit size and shape) to avoid these fruits. The colour variables included in the model were brightness, hue, chroma and fruits’ chromatic contrasts as seen by birds.
We compared models including morphology or colour as independent variables employing the Akaike Information Criterion (AICc, which corrects for small sample sizes; Turkheimer, Hinz & Cunningham 2003), using the smaller-is-better formulation. When comparing a series of models, the one with the lowest AICc is considered to be the best. We also calculated the Akaike weights (wi) that are particularly intuitive because they may be expressed in terms of percentage and can be interpreted as ‘weight of evidence’ in favour of the model with the lowest AICc when compared with all alternative models (Turkheimer, Hinz & Cunningham 2003). Comparisons between models allowed us to determine (i) which visual cues can be employed as accurate indicators of fruit quality and (ii) whether the inclusion of phylogenetic information significantly improves the fitness of the tested regression models. We did not employ Bonferroni corrections in this analysis because this approach can be very conservative in the context of model comparison (Burnham & Anderson 2002). Instead, we opted to report relations that were significant with a P < 0.05 weighed by the evidence provided by different models. That is, if a given relationship is statistically significant in more than one model, the total weight of the evidence favouring a significant relationship corresponds to the Akaike weights wi summed across models.