Caterpillars on a phytochemical landscape: The case of alfalfa and the Melissa blue butterfly

Abstract Modern metabolomic approaches that generate more comprehensive phytochemical profiles than were previously available are providing new opportunities for understanding plant‐animal interactions. Specifically, we can characterize the phytochemical landscape by asking how a larger number of individual compounds affect herbivores and how compounds covary among plants. Here we use the recent colonization of alfalfa (Medicago sativa) by the Melissa blue butterfly (Lycaeides melissa) to investigate the effects of indivdiual compounds and suites of covarying phytochemicals on caterpillar performance. We find that survival, development time, and adult weight are all associated with variation in nutrition and toxicity, including biomolecules associated with plant cell function as well as putative anti‐herbivore action. The plant‐insect interface is complex, with clusters of covarying compounds in many cases encompassing divergent effects on different aspects of caterpillar performance. Individual compounds with the strongest associations are largely specialized metabolites, including alkaloids, phenolic glycosides, and saponins. The saponins are represented in our data by more than 25 individual compounds with beneficial and detrimental effects on L. melissa caterpillars, which highlights the value of metabolomic data as opposed to approaches that rely on total concentrations within broad defensive classes.

TABLE S1 Details on individual compounds (numbered along the left column), including module numbers (see Figure 2), chemical class assignments, ridge regression beta coefficients (for caterpillar survival, development time and adult weight as in Figure 3), and mass spectra results: m/z, retention time (RT ave.) and relative mass defect (RMD). Compounds are organized by module number followed by compound number (an arbitrary designation). Cpd

TABLE S2
Compounds and coefficients from lasso regression allowing for possible effects of all individual compounds (as in Figure 3) as well as all pairwise interactions among compounds. Listed here are interactions selected by lasso regression, using binomial and Gaussian regressions for survival and adult weight, respectively (with units on those scales, as in Table S1). Results for development time are not shown here as none of the potential interactions for development time were selected by lasso regression.

FIGURE A1
Pairwise correlations between phytochemical modules (represented by eigenvectors m1, m2, etc.) and plant traits specific leaf area (SLA), protein and leaf toughness. Values shown are Pearson product-moment correlation coefficients.

FIGURE A2a
Pairwise correlations (as in Figure A1), but among compounds within modules m1 to m4. The bottom row in each graph shows correlations among individual compounds and the eigenvector used in other analyses for the given module.

FIGURE A2b
Pairwise correlations (as in Figure A1), but among compounds within modules m5 to m10. The bottom row in each graph shows correlations among individual compounds and the eigenvector used in other analyses for the given module.

FIGURE A2c
Pairwise correlations (as in Figure A1), but among compounds within modules m11 to m14. The bottom row in each graph shows correlations among individual compounds and the eigenvector used in other analyses for the given module.

FIGURE A3
Visualization of variation among plants (bars) in phytochemical modules. Each bar is the average of z-scores for compounds comprising a given module. Colors correspond to modules as in Figure 1. The order of plants (along the x-axis) is arbitrary.

FIGURE A4
Illustration of correlational structure among compounds, as in Figure 2, but here color coded by compound class identities instead of by module assignment.

FIGURE A5
Results from cross-validation (a, c, e) and resampling analysis (b, d, f) of Bayesian regressions reported in Table 1. In the observed vs predicted plots (a, c, e), each dot represents the observed and predicted performance of caterpillars associated with individual plants, which was the level at which cross-validation was conducted (i.e., one plant was left out of each iteration). In the resampling analyses (b, d, f), "modules" were constructed based on randomly assembled collections of compounds selected to match the structure of the initial analyses. For example, the survival regressions included modules 2, 3, 9, 10, and 11 (Table 1). Because m2 includes 21 compounds, for one iteration of the resampling analysis 21 compounds were randomly selected, from which the first eigenvector was taken for use in the analysis (and same for the other modules). From each of 1000 randomly-created sets of modules, R 2 values were retained and summarized in plots (b, d, f), along with dashed lines representing the R 2 values of empirical collections of compounds and associated eigenvectors. In all cases, empirical models outperformed all but a tiny fraction of simulated models. The fraction of the simulated models with R 2 values greater than the empirical models are as follows: 0.007 (b), 0.027 (d), and 0.023 (f).

FIGURE A6
Additional details on interactions by compound class. Top panel shows the number of times that compounds of a certain class appear in pairwise interactions (as in appendix Table S2) as a function of the total numbers of those compounds in different classes ("Count in data"). Classes that are represented by a greater number of distinct compounds in the total dataset are, in general, more likely to be observed in pairwise interactions: the dotted line shows the expected frequency of observation given random sampling from the total pool of compounds. Saponins and alkaloids are overrepresented in interactions, appearing in 44 and 41% (respectively) more pairwise interactions than expected by chance. Bottom panel shows average interaction coefficients for all interactions involving compounds of particular classes: most pairwise interactions are relatively small in magnitude, with the exception of interactions involving phenolic glycosides, that tend to be strongly negative. For both panels, interactions are considered from analyses of caterpillar survival and weight; no interactions were detected for development time (Supporting Information Table S2).

FIGURE A7
Compound specific effects vs coefficients of variation for individual compounds. The compound specific effects based on ridge regression are the same as those listed in Table S1.