Data collection

Occurrence data for subspecies of O. hupensis were assembled from Qian [27]. This national surveillance effort of schistosomiasis was carried out at the village level between the 1950s and 1980s across 12 Chinese provinces. In all, 5029 towns and villages reported presence of O. hupensis [27]. Rather than using centroids of infested counties, which reduces precision, we georeferenced individual villages using Google Maps. These points varied in terms of clumping, so we subsampled them to reduce sampling bias and spatial autocorrelation [28], as follows. First, we arranged infested provinces according to sample density (i.e. number of occurrence points divided by area of the province). The median served as the standard sampling effort, and all provinces presenting densities above that value were subsampled randomly to a lower density. In the end, we had 1996 occurrence points: 1402 O. h. hupensis, 470 O. h. robertsoni, 64 O. h. guangxiensis, and 60 O. h. tangi (S1 File).

Several approaches have been used to select environmental datasets for ecological niche modeling; the best environmental datasets would be ecological relevant to species in question [29]. At regional and continental scales, climatic factors have excellent predictive power in determining risk associated with disease transmission (e.g. schistosomiasis [25,26], West Nile virus [30]). Hence, we used subsets of the 19 bioclimatic variables developed by Hijmans et al. [31], chosen as follow. First, variables that combined temperature and precipitation (i.e. mean temperature of wettest quarter, mean temperature of driest quarter, precipitation of warmest quarter, precipitation of coldest quarter) were excluded because they display artificial discontinuities between adjacent grid cells in some areas [32]. The importance of each of the remaining 15 variables was assessed by a jackknife analysis of variable importance in Maxent ([33], see below for Maxent detail), and unimportant variables were discarded. Highly correlated variables were then removed in SDMtoolbox, a python-based GIS toolkit for spatial analysis [34]. Eight variables (S1 Table) that showed ecological relevance (regularized training gain >0.14) and low correlation with other variables (Pearson correlation <0.9) were chosen in the end. All variables were analyzed at a spatial resolution of 2.5 minute.

Climate space comparison

Climatic spaces occupied by the four subspecies were first compared along each environmental dimensions using violin plots, which combine the functions of boxplot and kernel density, providing a better indication of the shape of the data distribution. We used NicheA, a toolkit to create and visualized ecological niches in environmental spaces [35], to visualize climate niches occupied by each subspecies in reduced multiple environmental spaces: we displayed the first three principal components derived from the 8 bioclimatic layers, and plotted minimum volume ellipsoids (MVEs) around occupied conditions. We quantified niche overlap between pairs of subspecies using Schoener’s D [36]; this metric ranges from 0 (no overlap) to 1 (complete overlap), and was used to test niche identity and niche similarity between subspecies. Niche identity and similarity tests were performed to determine whether climate spaces occupied by the two subspecies were identical or exhibited significant difference, and whether these differences were caused by the environmental feature spaces [37]. Niche identity was tested by randomly allocating occurrence records within each pair 500 times, according to observed numbers of records, and comparing observed and simulated Schoener’s D estimates. In contrast, niche similarity was tested by shifting the centroid of the observed occurrence densities to a random location within the available environmental space 500 times, and comparing observed with the null distribution of simulated estimates of Schoener’s D [37]; climate variables measured at locations across the available backgrounds of subspecies were combined and projected onto the first two principal components using PCA_env package [37]. Smoothed densities of occurrences and available environments in each grid cell were calculated and compared among the four subspecies [37].

Background environments for climate niche comparisons and niche model calibration should include only areas that have been accessible to the populations under study [38]. We delimited this area by buffering a convex hull around known occurrences by 200 km (Fig 1) in SDMtoolbox [34]. This approach reflects a compromise between including all environments that have been accessible to the species, and still covering a broad-enough extent to minimize extrapolation and detect climatic differences between presence and background records [39].

Ecological niche modeling

To forecast climate change effects, we used fMaxent (fine-tuned Maxent, see below) and ensemble approaches [11,12] to calibrate models under present conditions, which were then transferred onto climate conditions for 2050 and 2080. Maxent is the most commonly used method in ENM, and it can fit arbitrarily complex models to explain relationships between environmental variables and occurrence data (version 3.3.3k; [40]). However, because an excessively complex model will be extremely specific to input data and perform poorly when extrapolating, Warren and Seifert proposed using a sample-size-adjusted Akaike information criterion (AICc) as criteria with which to address overfitting; this approach does not control model fit directly, but rather uses AICc to choose appropriate settings [7]. We used the “ENMeval” package [9] to fine-tune Maxent models by seeking the minimum value of AICc among candidate models. ENMeval provides an automated way to execute Maxent models across a user-specified range of regularization multiplier (RM) values and features combinations (FC). We set the RM range to 0.5–6.0 with increments of 0.5, and used 6 FCs, to cover a broad range of model settings. The block method was used to partition occurrence data into four bins, 3 of which were used for training and the remaining one for testing (bin combination, BC), which is desirable for studies involving model transferring [9]. In all, 2160 models (12 RMs × 6 FCs× 6 BCs × 5 occurrence groups) were generated for the four subspecies and for O. hupensis as a whole.

Ensemble models are used commonly in forecasting climate change effects, seeking to generate a consensus estimate that reduces individual model uncertainty by reflecting the central tendency of multiple models [11,12,41]. Here, outputs from six modelling algorithms, including generalized additive models (GAM), generalized boosted models (GBM), generalized linear models (GLM), random forests (RF), genetic algorithms (GARP), and the fMaxent model described above were included in ensembles. Individual GAM, GBM, GLM, and RF models were developed using BIOMOD2 [42], as implemented in R [43]; GARP models were developed in desktopGARP [44]. Details of implementation of each algorithm are provided in the supporting information (S2 Table). Model ensembles typically use a weighted averaging approach, in which models are weighted according to their interpolative performance (e.g. [30]). However, a recent assessment pointed to the challenge of balancing model interpolative accuracy against transferability [29,45]. Therefore, rather than using weighted averages, we used the PCA (median) method to identify the “central tendency” of individual model predictions [8,11]. The PCA measures, for each model, its ability to follow the general trend of predictions of the six models. This method calculates the median of the four individual models that had higher factor values among the six models [8,11,12].

The occurrence data used to fit the niche models were split randomly into two datasets, for calibration (70% of points) and interpolation evaluation (30% of points). Performance of individual and consensus models was evaluated via a partial ROC (receiver operating characteristic) approach [46]. Comparing to traditional AUC (area under the ROC curve), which was criticized because present data are more reliable than absence data in model evaluation [46], the partial ROC approach takes the quality of occurrence points into account and weights more on omission error [46]). Here, AUC calculations were limited to ROC spaces over which models actually made predictions, and only omission errors <5% were considered (i.e. E = 5%; [46]).

Final model runs incorporating all point data were used for visualizations and risk assessments. A modified least training presence threshold based on E = 5% was applied to fMaxent model predictions for O. hupensis and O. h. hupensis to generate binary predictions. We did not generate threshold predictions in ensemble future projection because such predictions are not applicable and hard to interpret (i.e. individual models for generating consensus models were different in present and future predictions).

Future climate scenarios

Future climate variables were downloaded from WorldClim [31], the Consultative Group on International Agricultural Research (CGIAR), and the research program on Climate Change, Agriculture and Food Security (CCAFS). To reduce uncertainty regarding future climate conditions (S1 Fig), rather than using the 13 original global climate models (GCMs, S3 Table) from the IPCC 5th Assessment, the PCA (median) protocol was also used to generate consensus “climate models” among the 13 GCMs for each climate dimensions for 2050–2060 and 2080–2090 (S4 Table).

The fMaxent and ensemble models based on present predictions were applied to these future conditions. Future climate models applied to the intermediate scenario of representative concentration pathways of 4.5 (i.e. “RCP45”; [47]) in which future anthropogenic greenhouse gas emissions were estimated to peak around 2040. This scenario was chosen because it represents the middle range of available four scenarios, and as such is considered more realistic than models based on extremely high or extremely conservative scenarios [47]. Climatic similarities between present and future in 2050 and 2080 were assessed using mobility-oriented parity (MOP) metrics, a correction and simplification of multivariate environmental similarity surfaces [39].

Spatial comparison

Different degrees of overlap were observed in the eight climate dimensions among the four subspecies (S2 Fig). Oncomelania hupensis hupensis and O. h. robertsoni occupied similar temperature and precipitation dimensions in terms of annual mean temperature (bio1), mean diurnal temperature range (bio2), and annual precipitation (bio12), but not isothermality (bio3), temperature seasonality (bio4), mean temperature of warmest quarter (bio10), or precipitation of driest month (bio14); O. h. guangxiensis and O. h. tangi occupied similar temperature and precipitation regimes in terms of annual mean temperature (bio1), temperature seasonality (bio4), mean temperature of warmest quarter (bio10), annual precipitation (bio12), and precipitation of driest month (bio14), but not mean diurnal temperature range (bio2) or isothermality (bio3) (S2 Fig). The four subspecies showed diverse responses to precipitation seasonality (bio15).

Minimum volume ellipsoids occupied by the subspecies overlapped broadly (Fig 2). The size of the MVEs corresponded roughly to the geographic range extent of each subspecies (Figs 1 and 2), with O. h. hupensis and O. h. robertsoni occupying larger volumes than O. h. tangi and O. h. guangxiensis. Niche overlaps between pairs of subspecies also corresponded roughly to their genetic distances estimated by 16S sequence (Fig 2 and S3 Fig): i.e. the close relationship between O. h. hupensis and O. h. tangi coincided with the highest climatic niche overlap (D = 0.215) among all pairs (S3 Fig). Similar patterns were observed between O. h. tangi and O. h. guangxiensis (D = 0.147), but to a lesser extent (Fig 2 and S3 Fig). The null hypothesis of niche identity was rejected in all pairwise comparisons (S3 Fig). However, in analyses of niche similarity, the null hypothesis could not be rejected, except for O. h. robertsoni versus O. h. tangi (S3 Fig). Results of niche identity and similarity thus suggest that, although the four subspecies occupy unique climate spaces, the nonequivalence of niche spaces derives from a background effect, and not from biological differences.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Minimum volume ellipsoids occupied by the four subspecies of Oncomelania hupensis.

The inset shows the phylogenetic relationships of the four subspecies based on 16S sequences [18]. Gray dots denote the background conditions available to O. hupensis across mainland China.

https://doi.org/10.1371/journal.pntd.0006021.g002

Model performance

Individual model performances varied across model algorithms and subspecies in interpolation validations (Fig 3). The machine learning methods (i.e. fMaxent, GBM, RF) generally showed better discriminant ability than regression models (i.e. GAM, GLM); GARP showed unstable performance (Fig 3). Similar to the machine learning models, consensus models showed good discriminant ability for the individual subspecies and for O. hupensis as a whole (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distributions of model evaluation results (AUC ratios) of individual and consensus models in interpolation validations.

Models were generated for the four subspecies separately and for Oncomelania hupensis as a whole.

https://doi.org/10.1371/journal.pntd.0006021.g003

Using all of the occurrence data, parameters of AICc-selected models (i.e. fMaxent) differed from default settings (S4 Fig). Based on block partitions of occurrence data, mean AUCtest values of fMaxent models were 0.79, 0.79, 0.80, 0.86, and 0.94 for O. hupensis (as a whole), O. h. guangxiensis, O. h. hupensis, O. h. robertsoni and O. h. tangi, respectively, with fMaxent models of O. h. hupensis (AUCdiff = 0.09) and O. h. tangi (AUCdiff = 0.02) showing less overfitting than the other three (S4 Fig). In species-wide consensus models, the first principal component explained 46.7–72.2% of individual model variation (Table 1). Consensus models were discriminated by the first axis of the PCA, and each individual model was selected in consensus model processing (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Variance projection of individual models on the first component obtained from a principal component analysis (PCA) of the six individual predictions.

Percentage of variance that PC1 explained, together with the factor loadings of individual models, are shown. Bold values are the four models for which the PC1s were the highest for each combination of species x time, which were selected for the “PCA(median)” consensus. Individual models were generated for the four subspecies separately and for Oncomelania hupensis as a whole.

https://doi.org/10.1371/journal.pntd.0006021.t001

Variation among individual model predictions spatially was observed in both present and future (2050 and 2080; S5 Fig). Some areas identified as suitable nonetheless corresponded to environments beyond the climate envelope of the calibration area at present, thus involving non-analog climate conditions (S6 Fig). For example, fMaxent identified disjunct suitable areas around Beijing in northern China (Fig 4), but these areas involved model transfer into novel climate conditions (S6 Fig), making their interpretation uncertain and unwise. Within the distribution of each subspecies (Fig 1), projection of present ENMs onto future climate datasets generally involved little extrapolation (MOP metrics; S6 Fig). Transferring present-day models onto future climate scenarios, fMaxent models were more conservative than ensemble models (Figs 4 and 5): the western part of the predicted distribution based on consensus models was cleaner than predictions based on fMaxent, and the consensus method did not make the isolated predictions in the fMaxent model (Fig 4). Both fMaxent and consensus approaches identified a pattern of range expansion and suitability increase in O. h. hupensis (Figs 4 and 5). In O. h. robertsoni, both models identified an eastward shift, whereas in O. h. guangxiensis, a northward shift was indicated (Figs 4 and 5). In O. h. tangi, the two models showed contrasting predictions (Figs 4 and 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Predicted potential distributional changes for Oncomelania hupensis as a whole.

The fine-tuned Maxent (fMaxent) and consensus models were calibrated on the accessible area and transferred onto future climate conditions in 2050–2060 (top) and 2080–2090 (bottom). White points indicate the surveillance sites, hatched areas indicate the administrative endemic areas.

https://doi.org/10.1371/journal.pntd.0006021.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Potential distributional changes for the subspecies Oncomelania hupensis hupensis and Oncomelania hupensis robertsoni.

The fine-tuned Maxent (left side) and consensus models (right side) were calibrated on the accessible areas and transferred onto future climate in 2050–2060 (up) and 2080–2090 (bottom). White points indicate the surveillance sites, hatched areas indicate the administrative endemic areas.

https://doi.org/10.1371/journal.pntd.0006021.g005

SNWD and surveillance sites

Binary predictions were based on fMaxent outputs, as thresholding future predictions from ensembles is difficult. Overlapping the Central Route and Eastern Route of SNWD with the binary future predictions for O. hupensis and O. h. hupensis, the southern Central Route coincides with suitable areas for O. h. hupensis in 2050–2060, and its suitable areas will expand northward along the southern Eastern Route by 2080–2090 (Fig 6). All of these areas are beyond the reach of present surveillance sites for schistosomiasis monitoring. Because a northward expansion of O. hupensis may occur considering future climate warming, these potential expansion areas need to be better covered by future surveillance efforts. Future surveillance efforts should also consider potential re-emergence of O. h. guangxiensis and O. h. tangi, as some areas of increasing suitability were noted for these two subspecies as well (Fig 5), although present intervention efforts have brought the snails to near extinction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Future transmission potential of schistosomiasis projected on the Central Route and Eastern Route of the South-to-North-Water-Diversion project.

Areas of agreement of fMaxent binary predictions of Oncomelania hupensis (as a whole) and O. h. hupensis are overlaid on the routes.

https://doi.org/10.1371/journal.pntd.0006021.g006