Connectivity patterns of QST transcripts and a multivariate perspective on GEI
The positioning of traits within a co-expression network relies on their patterns of connectivity with other traits. Identifying where adaptive traits fall within the co-expression network can help us understand whether adaptive evolution proceeds through large pleiotropic effects or through smaller fine tuning of weakly pleiotropic traits (Jordanet al ., 2004; Des Marais et al ., 2017; Josephs et al ., 2017). Additionally, comparing co-expression networks across environments allows identification of modules targeted by selection and those that are un-altered by changing environmental conditions, which we investigated by calculating three measures of connectivity for each expression trait. These reflected intramodular connectivity (kWithin), connectivity to all transcripts disregarding module membership in the network (kTotal) and the difference between inter- and intra-modular connectivity (kDiff). Specifically, to evaluate garden-specific patterns of GEI (H2) from a multivariate perspective, we performed comparative analyses using 10,000 permutations with the WGCNA package (Langfelderet al ., 2011). We used two aggregate summary statistics to declare a module as being preserved across gardens. Zsummary represents a normalised value of various connectivity and density-based measures following a permutation test procedure, while medianRank simply scores each module based on the observed preservation statistic. Following Langfelder et al . (2011), we used both Zsummary andmedianRank for inference of module preservation. We declared modules as preserved across gardens if their Zsummary scores were higher than 10 andmedianRank scores were below 5.
To assess whether weakly pleiotropic traits dominate the architecture of adaptive evolution (H4), we implemented the network characterization approach developed by Mähler et al. (2017) to define the core of a module based on the top 10% of transcripts with the highest kTotal. We determined if the observed relationship betweenQST categories and connectivity was significantly different than expected by chance by performing 10,000 permutations. Our permuted sets were matched on bins of expression levels to be representative of the core transcripts since expression levels are often associated with connectivity in a co-expression network (Fig. S4).