Specialization predicted from environmental niche models
To link abiotic environmental variation with host specialization (Q3), we identified areas with optimal conditions for specialization using environmental niche models (ENMs). ENMs are a widely used tool for characterizing suitable habitat for a species based on correlation between environment and observed occurrences of an organism (Elithet al. 2010). An implicit assumption is that organisms should have higher fitness in locations with higher habitat suitability predicted by ENMs (Nagaraju et al. 2013; Wittmann et al.2016).
We constructed ENMs based on all parasite occurrences (all-occurrence model) and subsets of occurrences associated with a particular host. ENMs were built using MaxEnt (Phillips et al. 2006), a machine learning based method for modeling distributions based on presence-only data. Models were built using ENMeval (Muscarella et al.2014) to test a variety of tuning parameters. Eight environmental variables were considered, including several associated with differential parasitism on host species (soil clay content; Wilson-Jones 1955; Mohamed et al. 2001), host exudation and production of hormones required for parasite germination (soil nitrogen and phosphorus; Yoneyama et al. 2013), or parasitism severity (annual rainfall; Wilson-Jones 1955). Bioclimatic and topographic variables (annual rainfall, mean temperature of the wettest quarter, isothermality, potential evapotranspiration, and topographic wetness index) were obtained from CHELSA (Karger et al. 2017) and ENVIREM (Title & Bemmels 2018) datasets. Soil variables (clay content, nitrogen, and phosphorus) were obtained from SoilGRIDs250m or AfSoilGrids250m (Hengl et al. 2015, 2017). For AfSoilGrids and SoilGrids250m products, we down-sampled to 1-km resolution using bilinear interpolation with the ‘resample’ function of the rasterpackage (Hijmans 2020). To characterize the background of the study, we randomly sampled 10,000 points from within a 500 km radius of all occurrences; the same background points were used to build each of the host-specific models and the all-occurrence model. We previously found that habitat suitability values from an ENM of S. hermonthicawere associated with a large effect resistance allele in sorghum landraces, supporting both the accuracy of the model and some degree of local host adaptation to S. hermonthica (Bellis et al.2020).
Here, we hypothesized that ENM contrasts, which we define as the difference in habitat suitability between two different ENMs, might also predict variation in performance on different hosts. We created ENM contrasts by subtracting the logistic-transformed ENM output for parasite occurrence on a focal host species from the all host species-occurrence model for each grid cell. Higher potential for specialization on a given host is indicated by more negative values in ENM contrasts (lower bound of -1) and lower potential for specialization is indicated by more positive values (upper bound of 1). Values close to zero can be interpreted as locations where generalist parasites are likely. We calculated the average value within 50 km of S. hermonthica locations from empirical studies as a measure of predicted specialization by ENMs, but similar values were observed at a range of distances (Fig. S2).
We investigated the factors most important for predicting overall occurrence in each model using permutation importance. However, environmental factors most important for determining habitat suitability may differ across space; to identify variables that most influence model prediction at each location, we created limiting factor maps (Elithet al. 2010) with rmaxent (Baumgartner & Wilson nd), fitting models based on the optimal regularization multipliers and feature classes parameters determined by ENMeval (millet: betamultiplier = 3.5, noproduct, nothreshold; maize: betamultiplier = 0.5, noproduct, nothreshold, nohinge; sorghum: betamultiplier = 3.5, nothreshold). After masking to a 200 km radius of any known S. hermonthica occurrence, niche overlap statistics were calculated inENMTools version 0.2 (Warren et al. 2010), in environmental space with the ‘env.overlap’ function or in geographic space with the ‘calc.niche.overlap’ function (Warren et al.2019).