There is some phylogenetic structure to this data, 8 of the 14 climate parameters demonstrate phylogenetic conservancy using the randomisation test of the quantitative convergence index. Figure 7 shows the mean values of the climatic parameter Annual Temperature Range, plotted on the phylogeny of Compton et al. 25 . It is evident that the clade comprising C. repandum, C. balearicum and C. creticum (Subgenus Psilanthum) all share low temperature ranges, whilst the clade C. parviflorum-C. elegans share higher ranges. In contrast, other climate parameters appear similar for much of the genus, all but 5 species share a dry season with the driest month providing less than 1 mm of rainfall per day. However the phylogenetic structure in the data is not uniform across the tree; for example, comparing C. elegans with its wide-ranging sister taxon, C. coum, shows that they differ for most precipitation values. Overall, the climate variable mean precipitation in the warmest month has the best fit for this phylogeny (QVI = 0.27). This may, in part, be due to the high level of similarity of this value amongst most Cyclamen, also the two highest rainfall values (C. purpurascens and C. colchicum) are sister taxa.

The phylogeny in figure 7 shows ancestral state reconstruction for annual temperature range. The lineage leading to subgenus Psilanthum shares the low range exhibited by its constituent species (fig 7 node 2). Table 2 shows the ancestral reconstruction for the means of all the climatic parameters. The ancestral Cyclamen lineage (node 1) has the Mediterranean climatic characteristic of dry summers and wet winters, demonstrated by mean daily precipitation 5 times lower in the warmest month than in the coldest. There is also some frost tolerance exhibited, averaging more than 1 month with minimum temperature below freezing.

If we examine the ancestral model with reference to today's climate (fig 8), an area of central Greece and Western Turkey is selected which is close to the current centre of diversity highlighted in figure 5. Other internal nodes on the phylogeny each show differing present-day realisations of model reconstructions. For example the lineage leading to the clade comprising C. hederifolium – C. colchicum demonstrates the broad tolerance of its wider-ranging constituent species. However, there is a predominance of Turkish areas for the majority of these models.

We can examine which of the extant species of Cyclamen is closest in climatic preference to our model reconstruction. It is not possible to use the full interpredictivity measure to determine which of the present day Cyclamen is most like the ancestral lineage. However, in the area selected by the model you can find the species C. hederifolium, C. repandum, C. coum, C. graecum and C. intaminatum. No extant species has a distribution similar to this area (Kappa < 0.02 for all species). However, a crude measure of model similarity based on the sum of squares of the differences of means for each climate layer suggests C. balearicum is overall the most similar to our ancestral reconstruction.

As well as comparing the constituent climatic parameters, we can compare directly the resulting BIOCLIM models for each species. Figure 9 displays a measure of model similarity for each pair of Cyclamen species. More than two thirds of all comparisons show zero prediction of other species distributions, which suggests that most Cyclamen are climatically isolated. It is also clear that there are several climatically wide-ranging species (C. coum, C. hederifolium, C. cilicium, C. graecum, C. persicum, C. repandum) which account for the majority (69%) of all positive predictions. The bioclimatic envelope for C. coum shows some overlap with the envelopes of all other Cyclamen except C. somalense.

We can combine sets of these predictions together to produce measures of group similarity, this permits an examination of climatic preferences within a clade. Table 3 shows the mean model similarity amongst the subgenera recognised by Grey-Wilson 26 (Tab. 3a) and selected clades from the Cyclamen phylogeny (Tab. 3b). Notice that subgenus Corticata, subgenus Persicum and the clade above node 13 (C. parviflorum-C. libanoticum) all show similarity levels of zero. Several groups (subgenus Psilanthum and the clade above node 16) demonstrate mean similarity levels which are almost double that within Cyclamen as a whole (Tab. 3c). However, if we test whether these similarity values are above that of a random group of Cyclamen of a similar size we find that the group means are not significantly different from the random groups. On closer examination we find that the higher similarity values are driven by the presence of a single wide-ranging species within the groups.

The bioclimatic niche models, when examined within a future climate scenario for the 2050s, give an estimate of where each species' preferred climate will be. Figure 10 combines these predictions to repeat the Cyclamen diversity map (fig 5) for the future climate scenario. The models with unrestricted dispersal demonstrate a northward shift in the area of climatic preference, which is particularly evident for the Maxent models (fig 10b). The unrestricted dispersal scenario is highly unlikely without substantial human intervention, a more likely scenario is the restricted dispersal hypothesis demonstrated in Figure 11. Here we see major reductions in suitable climate suggested by both algorithms, particularly in central Italy, the former Yugoslavia, Sicily, Southern Turkey and the far eastern extent. Overall, Greece & Turkey still show the most diversity, but the extent is greatly reduced. When we consider the models for each species in turn (Table 4), we find that the area of climatic suitability for every Cyclamen has reduced for both modelling algorithms, the majority by more than 60%. Many of these species are considered to be at high risk of extinction due to climate change.

Many Cyclamen share a preference for seasonal Mediterranean climate characteristics. The ancestral Cyclamen was probably well suited to this environment, and may have had a similar climatic preference to the extant species C. balearicum. A question remains on the timing of the origin of Cyclamen and how this compares with estimates of the emergence of the Mediterranean climate zone.

Although, for some ancestral model reconstructions, no present-day area was within the suitable envelope, this does not invalidate the model, as past climates could exhibit climate types not seen in the present day 35 . However, failure to predict any suitable areas of the relevant palaeoclimate could demonstrate the invalidity of the models. Yesson and Culham 23 have demonstrated the plausibility of this general approach to estimates for ancestral areas based on reconstruction of climate preferences for ancestral lineages in other taxa.

There is not a phylogenetic pattern of inherited range size for Cyclamen. Several sister pairs of species are seen to have drastically different ranges, all but two species pairs on the phylogeny differ in their total area of distribution by more than 90%. This fits the pattern of schizo-endemic distribution of C. balearicum and C. repandum found by Thompson et al. 36 , whereby a single wide ranging species is assumed to be progenitor to a narrow ranging isolated sister species. This is supported by the lack of overlap in bioclimatic niche models. This suggests a combination of geographic, edaphic and climatic constraints are limiting the maximum area of distribution of some species more than others.

Examining the extinction risk from a phylogenetic perspective (fig 12a, b) we see clearly that within each major lineage there is a pattern of contrasting extinction and survival predictions. BIOCLIM and Maxent predict contrasting patterns of individual species threat. When examined from an overall phylogenetic viewpoint, while many individual species are at high risk, each major lineage is seen to contain at least one species with a reasonable chance of survival. This pattern lowers the overall risk to phylogenetic diversity, and presents the risk in a more favourable light 37 38 .

Whilst many of the constituent climatic variables demonstrate phylogenetic conservancy, the resulting models which combine the individual variables into bioclimatic envelopes show little overlap between adjacent nodes. It may be the case that measures for comparing these envelopes are inadequate when the majority of species are climatically isolated from one-another. The only point of comparison of bioclimatic niche models that is currently used is to compare the areas that they select when overlaid into present day data 19 39 . This is also true for the comparison measures used in model validation such as Cohen's Kappa 40 . This is not a direct measure of model similarity. Furthermore such measures may 'hide' real differences for climate types not evident in the present day.

Some studies have found patterns of risk associated with a particular type of climate 18 41 . Thuiller et al. 41 found species from Mediterranean climates to be at lower risk of extinction than species from other climates, whereas Thuiller et al. 18 found species in Mediterranean regions to be the most sensitive to the changing climate, with species in mountainous regions under great threat but other species having the potential to expand. For Cyclamen there is no particular climate-type which is more at risk than another. There is little overlap between any of the bioclimatic niche models of the high-risk Cyclamen species, many of these species occur in differing climates and geographies. For example the highest risk group (BIOCLIM models) includes the Mediterranean-type climate of C. rohlfsianum; the continental climate of the mountain dwelling C. colchicum; the island endemic C. cyprium; and the African C. somalense which is barely 10° north of the equator. However, for the Maxent models the wide-ranging, more generalist species are consistently assigned a high risk.

These broad classifications of risk can hide some noteworthy detail. For example, although C. balearicum is designated Endangered (>50% area loss), it is noted that the BIOCLIM future area selection gives no suitable area for this species within the Balearics, and less than 5% of its present day French distribution is considered climatically suitable, indeed most of the predicted suitable climate occurs in Northern France. It is interesting to note that the Balearics are still suitable for other Cyclamen species. It is reassuring to note that the French populations of C. balearicum have a far higher genetic diversity than the Balearic populations, but the latter demonstrate the higher ecological diversity 31 .

There is a link between current range size and extinction risk due to climate change. For the BIOCLIM models every species with a current distribution of under 35,000 km² is considered the highest risk with the exception of C. pseudibericum. In contrast no species with a range size above 35,000 km² loses all area. Range size and proportion of area lost is strongly negatively correlated for the BIOCLIM (Pearson correlation = -0.816, P < 0.001). In contrast the Maxent models show weak positive correlation of proportion loss and range size (Pearson correlation = 0.332, P = 0.142). The changes in range size show a distinct scattering across the cladograms whether modelled with BIOCLIM or Maxent (fig 12).

There are of course other factors than climate limiting species distributions. At the fine scale factors such as soil, aspect, and surrounding vegetation will have a role in determining the establishment of individuals given an otherwise suitable climate. Such fine-scale limiting factors are of secondary importance when modelling species distribution at a global or regional scale 42 . These criticisms of species distribution modelling and the extinction risk measures based on them are discussed at length in other places 6 19 41 42 43 44 45 46 47 and we refer the reader to these references.

The BIOCLIM models are restrictive in their definitions of the edge of a suitable climate, predictions for all areas are essentially a binary presence/absence and do not allow any climatic parameter to be outside the observed range 47 . For example, the BIOCLIM models predict no Cyclamen in the UK, but we know that three species are naturalised there. Alternative modelling algorithms such as Maxent have the advantage of producing probability measures rather than binary presence/absence, which provide more realistic edges to ranges 48 . However, for the UK naturalised Cyclamen, the Maxent models also fail to select any of their naturalised distribution within the selected threshold.

Both modelling algorithms predict large range loss for all Cyclamen species, and the predicted future diversity maps are broadly similar (fig 11). However, predictions for individual species differ dramatically. The Maxent models show the highest area loss, yet the BIOCLIM models suggest many more species at the highest extinction risk. This kind of algorithm dependent difference has been observed in other studies 44 49 .

For the BIOCLIM predictions, it is evident that range size is a key factor in determining extinction risk due to climate change. Those species at high risk tend to have smaller distributions and smaller climatically suitable areas. This is in general agreement with others who have found a relationship between niche breadth and projected range loss 18 41 , (it is trivial to note that for Cyclamen the widest-ranging species have the broadest niches). However, the relationship between range size and range loss reported by Thuiller et al. 41 is a bell shaped curve with the largest ranges suffering substantial losses, although it is possible that no Cyclamen would fall into the category of truly wide-ranging. In contrast the Maxent models demonstrate the opposite pattern (albeit weakly), with widest-ranging species losing proportionally more area. It should be noted, however, that bioclimatic niche models may be less accurate for wide-ranging generalist species 4 50 .

Thuiller et al. 51 considered selecting the best algorithm for each species in turn based on validation scores. Such a policy, in this case, would combine the highest area loss of both methods producing high overall area loss and high numbers of species with 100% projected area loss. However it is to be noted that present day validation may not reflect model validity for the future 46 51 .

The main advantage of BIOCLIM in the context of Phyloclimatic modelling is its independent treatment of climatic variables. As far as we are aware, this is the only method that permits independent optimisation of environmental characteristics on a phylogeny following established phylogenetic reconstruction methods. These optimised variables can then be directly converted into a BIOCLIM bioclimatic niche models. Reconstruction of ancestral states on a phylogeny requires a distance metric to directly compare the output models of different species. There are many problems associated with using the more complex (and often more precise) models currently being developed. For example, many discard non-informative input parameters, meaning models are not directly comparable amongst species as different species models will be built with different subsets of variables. Other models are based on the amalgamation of multiple model outputs, again making direct comparison of models difficult.

The 21 species recognised by Compton et al. 25 were used for this study. Their phylogenetic study of Cyclamen is used to define relationships within Cyclamen for this paper. Their study produced phylogenies based upon nuclear (ITS) and plastid (trnL-F) DNA, as well as morphological data. The phylogenetic trees have strong topological similarity, differing mainly in degree of resolution. One of the two maximum parsimony trees based on the combined ITS + trnL-F dataset is used throughout this analysis. The alternative topology differs only in the position of C. graecum relative to C. rohlfsianum within series Persicum, our chosen topology agrees with findings of Clennett 52 with regards to this clade.

Locality data were collected for each of the 21 Cyclamen species. These data came from four sources. Firstly from Cyclamen Society collecting trips, which has made many trips over the period 1987–2004 covering Greece and the Greek islands, Israel, Sardinia and Turkey, providing 2,315 observations for 12 species 28 . Secondly, data from the Global Biodiversity Information Facility 53 , provides 232 observations for 6 species from 6 institutions covering Austria, Croatia, Germany, Greece and the Greek islands, Israel, and Italy (a full list of institutions in provided in the appendix [see Additional file 1]). Thirdly, data from specimen labels at the University of Reading Herbarium (RNG) containing 61 observations for 12 species covering the Balearics, Algeria, Austria, the Caucasus, Corsica, Croatia, Cyprus, Greece and the Greek islands, Israel, Italy and Turkey.

Points were excluded if they fell in an ocean or sea. Exclusions were minimal, comprising fewer than 3% of the total observations, primarily being records from near coastlines with poor geographic resolution (fig 3 shows an example of poor geographic resolution). Species with fewer than 10 observations were excluded from analyses, following Graham et al. 20 . Data from the UK National Biodiversity Network for C. hederifolium, C. repandum and C. coum were excluded from the analysis as these represent records outside the native range of these species 34 . Exclusions reduced the number of analysable Cyclamen down to 13 species.

The fourth source for distribution data was the detailed distribution maps of Grey-Wilson 26 . This alternative approach allowed analysis of all Cyclamen species. These maps were digitally captured and pseudo distribution points were produced by taking the centroid of every overlapping quarter degree square. Quarter degrees were chosen as this is the resolution of the climate data used in this study. This produced distribution data for all 21 Cyclamen species. Due to its restricted distribution C. libanoticum has only 5 pseudo-localities following this methodology, and is the only species that falls below the 10 observations recommended by Graham et al. 20 , however reasonable models were still produced for this species, and this species is included in all analyses.

Cyclamen range size was estimated by summing the area of each quarter degree grid square that overlapped with the digitised distribution maps.

Present day observed climate data were taken from an observed climatology dataset (known as CRU CL1) from the Climate Research Unit 32 54 . Future data were taken from the Hadley centre general circulation model (HadCM3) from the IPCC website 33 . The period 2040–2069 hereafter referred to as 2050s was chosen as the future time-frame. This allows sufficient time for the effects of climate change to be detectable, whilst minimising the overall timeframe to reduce uncertainties 55 56 There are many different future climate scenarios based on different predictions of socio-economic factors, we have chosen the A2 scenario representing the mid-range of climate change severity 56 . In order to produce future climate datasets which are directly comparable with the observed data, we applied modelled changes in climate to the present day observed data. The modelled changes were calculated by subtracting the 2050s prediction from directly comparable modelled present day climate following Peterson et al. 24 and IPCC recommendations 57 .

Both the present and future datasets were processed into biologically meaningful climate profile parameters 7 . Standard deviation of mean temperature, mean daily precipitation in wettest month, mean temperature in warmest month, mean temperature, standard deviation of mean precipitation, number of months with minimum temperature above freezing, mean temperature in coolest month, mean daily precipitation in warmest month, annual temperature range, mean daily precipitation in driest month, lowest temperature in coolest month, mean daily precipitation in coolest month, mean daily precipitation, highest temperature in warmest month. These are similar to Busby 7 and were processed using the "Climate Data Processor" plug-in developed by Tim Sutton as part of the Quantum GIS project 58 .

The mean, standard deviation, minimum and maximum values were collected for all climatic parameters, for all species. Each value was independently optimised as a continuous character across the chosen phylogeny, using square change parsimony optimisation implemented in Mesquite 59 , following the methodology of Graham et al. 20 . A table showing each of these optimised values for every node on the tree is provided in the appendix [see Additional file 2]. These values define a BIOCLIM bioclimatic niche model 3 for each node on the tree. Models were created and manipulated using the BIOCLIM algorithm of the openModeller software package version 0.3.4 60 .

Additional models were constructed using the maximum entropy method (Maxent) for modelling species geographic distributions 9 . This new method has been shown to be very effective for species distribution modelling 4 . Models were built using the same data as the BIOCLIM models using Maxent version 2.3.0 61 . Thresholds were calculated for models to maximise the Kappa statistic. These thresholds were used to determine presence/absence for present day and future projections.

The quantitative convergence index (QVI) 62 was calculated for each optimised climate parameter. This measure is analogous to 1 minus the retention index for discrete characters. A randomisation test of this value was performed 63 . A random shuffle of the terminal node values was optimised on the chosen topology, this was repeated 100 times for each character. The observed QVI was compared with the distribution of the randomisation replicates to test if the observed value is different from a random placement of data on the fixed phylogeny. If the observed QVI was outside the 95% confidence interval of the randomisation replicates then it was considered significantly different from random 63 .

Only the variables demonstrating phylogenetic conservancy, with reference to the QVI, were used in the construction of ancestral bioclimatic niche models. This reduced the number of variables used in this stage of the modelling from 14 to 8.

Similarity of bioclimatic niche models was assessed using interpredictivity calculations 19 39 . This involves the overlaying of locality records of one species (Species A) into the predicted area of another (Species B). Similarity is measured as the percentage of points of A falling within the prediction area of the model of B, and vice versa 19 . Note this is not a symmetric measure, for example, if the envelope for A completely encompasses B then we could have B⊂A but A⊄B, therefore both directions of similarity were performed. Similarity was also assessed using Cohen's Kappa statistic 40 64 . Mean similarity for a group of species was assessed by averaging similarity values of each pair of species within the group.

Summary 'hotspot' maps were produced by creating binary presence/absence maps for each species based on their bioclimatic niche model output, then summing the number of species predicted in each quarter degree cell. This was accomplished using software accessible through the openModeller project 60 .

Present day bioclimatic niche models were examined within the 2050s climate scenario to produce a future distribution estimate based on climate suitability. Species with zero area predicted as climatically suitable are selected as high extinction risk 65 . When predicting future distributions and extinction risk it is necessary to take dispersal rates into account 2 66 . Plant migration rates have been estimated at 20–40 km over a 100 year timescale 67 , more recently even lower rates have been suggested 68 , it is highly unlikely that the ant-dispersed Cyclamen will exceed these dispersal rates. Our analysis uses a fifty year timescale and a geographic resolution of approx 25 × 25 km squares, so we employ a model permitting dispersal only within cells immediately adjacent to the current distribution. Therefore species which have future predictions which do not overlap with this the slightly expanded present distribution are also considered at risk, though this is regarded as lower risk than the case where there is no suitable area. Risk categories are assigned according to IUCN classifications based on area loss over 505 years 16 69 , namely: Extinction 100% loss, Critical >80%, Endangered >50%, Vulnerable >30%