The ITS region was the most informative region followed by the matK region (Table 1). The trnLF and the nad1 regions resolved 10% or less of the clades with strong support defined as ≥ 90% bootstrap support. There were no strongly supported conflicts among any regions (Bayesian consensus tree with posterior probabilities and parsimony bootstrap consensus tree for the total evidence analysis and bootstrap consensus trees of individual regions are provided as Additional file 1: Figures S1-S7). The number of resolved clades (87%; Table 1) and the number of resolved clades supported by ≥ 90% bootstrap (62%) was highest in the total evidence analysis, which was also the only analysis that resolved all the major lineages. The Bayesian analysis of the total evidence matrix provided the same overall topology as parsimony analysis and all major clades were strongly supported by Bayesian analysis (Figure 2; Additional file 1: Figures S1 and S2).

¹PPI: potentially parsimony informative characters. ²CI: Consistency index. ³RI: Retention index. ⁴Percent of resolved clades in the Bootstrap consensus tree with > 50% BS (bootstrap support) and with ≥ 90% BS are proportions of the possible number of clades (# taxa −1).

The topology (Figure 2) of the total evidence analysis largely supports previous studies 42 56 . The African tribe Amaryllideae (100% BS; PP = 1.00) is sister to the remainder of the Amaryllidoideae (100% BS; PP = 1.00), and this is strongly supported by all analyses. The next major split resolves an American clade (66% BS; PP = 0.99) and a Eurasian clade (65% BS; PP = 1.00) as sisters (100% BS; PP = 1.00) and a clade (66% BS; PP = 1.00) with the African monogeneric tribe Cyrtantheae (100% BS; PP = 1.00), and tribe Haemantheae (100% BS; PP = 1.00) as sisters (100% BS; PP = 1.00), and the Australasian tribe Calostemmateae (100% BS; PP = 1.00) as sister to these.

In the ITS analysis ( Additional file 1: Figure S3), tribe Amaryllideae (100% BS) is sister to the remainder of Amaryllidoideae (61% BS). Within the remainder of Amaryllidoideae, tribe Calostemmateae (100% BS) is sister to a clade (54% BS) including tribes Cyrtantheae (100% BS), Haemantheae 95% BS) and the American and European Amaryllidoideae. Tribes Cyrtantheae and Haemantheae are sisters (96% BS). In the combined plastid analysis ( Additional file 1: Figure S6), Cyrtantheae is sister to the remainder of Amaryllidoideae except tribe Amaryllideae (75% BS). In both the matK ( Additional file 1: Figure S4 supporting online material) and the combined plastid analysis tribe Calostemmateae is sister to tribe Haemantheae (matK: 70% BS; plastid: 94% BS).

The low bootstrap support (65%) for the Eurasian clade in the total evidence analysis (Figure 2, Additional file 1: Figure S2) may be caused by uncertainty in the placement of the genus Lycoris. The remainder of the Eurasian clade is strongly supported in all analyses except trnLF and nad1, which are the two regions providing the least resolution and support in general ( Additional file 1: Figures S3-S7 in the supporting online material).

The relationship of individual types of compounds and biological activity could be assessed using the D metric developed to deal with discrete binary characters 57 . D is equal to 1 if the observed chemical component has a random distribution (i.e. no phylogenetic signal). D is equal to 0 if the component is distributed exactly as would be predicted under a Brownian motion model of gradual divergent evolution (i.e. strong phylogenetic signal). See Materials and Methods for details. Of the seven alkaloid types which are found in more than one species we found evidence for some phylogenetic signal of five types (Table 2a) 57 . With the exception of tazettine and galanthindole, D values were significantly different from that expected under a random distribution of the components across the phylogenetic tree (D = 1). Inhibition of acetylcholinesterase (AChE) and binding to the serotonin re-uptake transporter (SERT) were used as proxies for biological activity. Both measures of bioactivity, AChE activity and SERT activity also displayed significant phylogenetic signal (Table 2b). However, in none of the above cases was the phylogenetic signal sufficiently strong to be considered as indistinguishable from a Brownian motion model of evolution, where traits are strongly ‘clumped’ on the phylogeny (and D = 0) 57 .

D is equal to 1 if the observed chemical component has a random distribution (i.e. no phylogenetic signal). D is equal to 0 if the component is distributed exactly as would be predicted under a Brownian motion model of gradual divergent evolution (i.e. strong phylogenetic signal). P values represent the probability that the observed D value is equal to 1 or 0, respectively (P >0.05 indicates that the observed value is not significantly different from these values).

There was a statistically highly significant correlation between differences in chemical profile and phylogenetic distance, indicating that closely-related species tend to have more similar chemical profiles than more distantly-related species, although the effect was not strong (Mantel test: r = 0.085, p = 0.002). There was also statistically significant correlation between chemical profile and phylogenetic distance in the genus level comparison (Mantel test: r = 0.090, p = 0.024), although the effect was also weak.

For the purpose of the present study, we consider the total evidence approach to provide the best estimate of phylogeny and all major lineages are supported by both parsimony and Bayesian analyses. The present study has doubled previous sampling of Amaryllidoideae from 51 species 56 to 108 and from combined analysis of two DNA regions 42 56 to four DNA regions representing all three genomes. The only previous study resolving relationships among basal lineages was based on plastid ndhF sequences 42 and resolved Calostemmateae and Haemantheae as sisters and tribe Cyrtantheae as sister to these. However, in the present study (Figure 2), the African tribes Cyrtantheae and Haemantheae are strongly supported by both bootstrap and Bayesian posterior probabilities as sister clades. Tribe Calostemmateae is sister to these, although this was only weakly supported by the bootstrap, but strongly supported by Bayesian posterior probabilities. A sister group relationship of the two African tribes Cyrtantheae and Haemantheae and the Australasian tribe Calostemmateae as sister to these appears more convincing than the alternative based on biogeography. However, in terms of morphology there may be some room to question this relationship. The indehiscent capsule of Calostemmateae has more in common with the indehiscent baccate fruit of Haemantheae (resembling the unripe fruit of Clivia, Scadoxus, Haemanthus, and Cryptostephanus) than with the dehiscent capsule of Cyrtanthus.

Our approach to quantify overall correlation between chemical and phylogenetic diversity has previously been applied to show positive correlations between pheromone differences and nucleotide divergence in Bactrocera fruit flies 58 and phylogenetic correlation of cuticular hydrocarbon diversity in ants 59 . We have now shown the potential application of this approach to explore correlations between phylogenetic and chemical diversity of medicinal plants.

We found significant phylogenetic signal for five out of seven tested individual alkaloid types and for both AChE and SERT bioactivity proxies, although the signal was not strong enough to be indistinguishable from a Brownian model of evolution, where traits are strongly clumped on the phylogeny 57 . There was also a highly significant correlation between differences in chemical profiles and phylogenetic distance in both species and generic level analyses. Despite this, there was often still considerable difference in chemical make-up even between phylogenetically very close species (Figure 3) with the average number of alkaloid types differing between congeners being 2.26. Whereas members of some genera such as Crinum and Galanthus have generally similar chemical profiles, other genera such as Hippeastrum and Narcissus have striking diversity.

An explanation for the moderate correlation found could be either methodological artefacts or underlying ecological or genetic differences 38 . We minimized methodological artefacts by using the same plant accessions for both phylogenetic, chemical, and bioactivity studies, and by analysing our data with consistent methods. Chemical profiles were based on types deducted from hypothetical pathways and could be an oversimplification of the chemical diversity contained by over 500 individual alkaloid structures known from the subfamily 45 47 .

The strength of correlation could be dependent on taxonomic scale. Whereas alkaloids derived from norbelladine and its derivatives are almost exclusively restricted to the subfamily Amaryllidoideae 45 , and alkaloids with AChE activity appear to be phylogenetically constrained within Narcissus 18 , the considerable variation at the species and genus level found in this study corresponds well with within species variation of alkaloid profiles in for example Galanthus 51 60 .

Evaluation of extensive historical drug data, marine natural products, medicinal plants and bioactive natural products suggests that drugs are derived mostly from pre-existing drug-productive families that tend to be clustered rather than randomly scattered in the phylogenetic tree of life 61 . Zhu et al. 61 further suggest that efforts to identify new potential drugs can therefore be concentrated on exploring a number of drug-productive clusters. However, based on our results, such a strong presumed correlation between phylogeny and bioactivity appears to be an oversimplification, at least at the taxonomic scale tested in our study. Based on our data for the medicinally important plant subfamily Amaryllidoideae, it appears that phylogeny can predict chemical diversity and bioactivity, but considerable caution must be emphasized. We also suggest that phylogenetic correlation of chemical traits of interest may need to be assessed for a particular phylogenetic framework before it is used for prediction of occurrence in un-investigated taxa.

A predictive approach could enable deduction of biosynthetic pathways, defence against herbivores, more efficient selection of plants for the development of traditional medicine and lead discovery as well as inform conservation priorities as outlined in the introduction. A plethora of data on phylogenetic relationships, chemical constituents and bioactivity are available through public databases (e.g., GenBank) and in the literature. Systematic in silico data mining could enable more efficient use of predictive approaches to speed up all of the above applications 22 61 62 63 64 65 .

However, a methodological framework still needs to be developed. In the present study, we have suggested an approach for testing correlations between phylogenetic and chemical diversity and biological activity using experimental data generated for this purpose. One method for subsequent identification of specific nodes in phylogenies with high bio-screening potential has been proposed by Saslis-Lagoudakis et al. 22 using tools from community ecology. In this approach, a matrix of ethnomedicinal use was composed and used to identify nodes in a phylogeny of Pterocarpus (Fabaceae), which have more medicinal taxa related to a specific category of use than expected by chance. This approach could be useful for identifying alternative resources or substitute taxa in cases where supply of a medicinal plant or natural product of interest is limited or where species in use are subject to conservation concerns 22 . However, for the purpose of increasing the chance of making truly new discoveries such as new compounds and/or new activity profiles, it may be more relevant to identify clades that possess activity of interest and at the same time do not correspond to well known compounds with well known activity profiles 66 .

Other methods for predictive in silico data mining may be combined with a phylogenetic selection approach, e.g., exploration of natural product chemical space as developed by Backlund and co-workers 65 67 68 . Another computerized geospatial tracking tool linking bioactive and phylogenetic diversity has been developed for microorganisms 63 . The concept of virtual parallel screening developed for natural products by Rollinger 64 , which simultaneously enables fast identification of potential targets, insight into a putative molecular mechanism and estimation of a bioactivity profile, could allow for optimal selection of relevant targets.

Specimens were collected in their natural habitat or obtained from botanical gardens or specialist nurseries. Sampling included 108 (over 10%) of circa 850 species in Amaryllidaceae subfamily Amaryllidoideae 69 with Agapanthus campanulatus L. used as outgroup ( Additional file 2). Sampling represents 43 of circa 60 genera and all currently recognized tribes except Griffineae Ravenna 41 42 56 . Samples from tribes Galantheae and Haemantheae were partly retrieved from previous studies 21 51 . The same accessions of plant material were used for both molecular, chemical, and bioactivity analysis to minimize effects of intraspecific and ecological variation.

DNA was extracted using the Qiagen DNeasy kit (Qiagen, Copenhagen, Denmark) from 20 mg of dried leaf fragments. Amplification and sequencing of the nuclear encoded ITS and plastid encoded matK and trnLF regions followed Larsen et al. 51 . Amplification and sequencing of the mitochondrial nad1 region followed Cuenca et al. 70 . Primers used for amplification and sequencing are listed in Additional file 3. Both strands were sequenced for each region for all taxa whenever possible. Sequences were edited and assembled using Sequencer 4.8^TM software (Gene Codes, Ann Arbor, MI, USA). All sequences are deposited in GenBank and accession numbers JX464256- JX464610 are listed in Additional file 2. Sequences were aligned using default options in MUSCLE 71 as implemented in the software SeaView 72 .

Phylogenetic analyses were conducted using both parsimony and Bayesian inference. Most parsimonious trees (MP) were obtained with PAUP v. 4.0b10 73 using 1,000 replicates of random taxon addition sequence and TBR branch swapping saving multiple trees. All characters were included in the analyses and gaps were treated as missing data. We analysed the four regions separately to identify strongly supported phylogenetic conflicts among the regions, prior to performing a combined total evidence analysis. By using total evidence the explanatory and descriptive power of the data is maximized 74 . Bootstrap analyses 75 of the four individual datasets and the combined dataset were carried out using 1,000 replicates. Bayesian analysis of the combined dataset was performed with MrBayes 3.1.2 76 . We first selected the best fitting model (GTR + I + G; Parameters: lset NST = 6 RATES = gamma) of molecular evolution using the Akaike criterion (AIC) in Modeltest v. 3.8 77 . The analysis was performed with 1,000,000 generations on four Monte Carlo Markov chains. The average standard deviation of the split frequencies was 0.01 after 232,000 generations and < 0.005 after 1 million generations corresponding to an effective sample size of 115 using the software Tracer v. 1.5.0 78 . The first 2,500 (25%) trees of low posterior probability were deleted and all remaining trees were imported into PAUP. A majority rule consensus tree was produced showing the posterior probabilities (PP) of all observed bi-partitions. We also performed a partitioned analysis allowing different models for the three genomes. However, in consideration of the limited information present in our plastid and mitochondrial datasets (Table 1), partition rich strategies are not always the best ones and in some cases less complex strategies have performed better 79 80 . Although the Bayesian MCMC approach is good at handling complex models, there is a risk of over-parameterization, which can result in problems with convergence and excessive variance in parameter estimates 81 .

Alkaloids were extracted from 300 mg dried bulb scales using 0.1% H₂SO₄ and clean-up on ion-exchange solid phase columns as described by Larsen et al. 51 . All extracts were concentrated under vacuum until dryness and re-dissolved to a standard concentration of 5 mg ml^-1 in MeOH. Alkaloid profiles were obtained by gas chromatography–mass spectrometry (GC-MS) as described by Larsen et al. 51 using a method developed by Berkov et al. 82 . Alkaloids were identified to type by comparison with the NIST 08 Mass Spectral Search Program, version 2.0 (NIST, Gaithersburg, Maryland) and with published spectral data. Alkaloid structures were scored to one of eighteen types (Figure 1) proposed by Jin 45 46 based on hypothetical biosynthetic pathways 51 . Only nine of the eighteen alkaloid types were recorded in the present study (Figure 1). Each type show characteristic fragmentation patterns in the MS-spectra 83 . In most of the cases, the database proposals with highest similarity could therefore be used to score the candidate structure indirectly to one of the types. Candidate structures were excluded from the profile if they could not be scored unambiguously to types.

AChE inhibition and SERT affinity of the standardized alkaloid extracts were tested using published methods 21 . AChE activity was conducted using isolated acetylcholinesterase (Electrophorus electricus, Sigma, Germany) and SERT activity using homogenates of whole rat brains except cerebellum. Galanthamine and fluoxetine hydrochloride were used as positive standards in the AChE and SERT assays, respectively. Data were analysed with the software package GraFit 5 (Erithacus Software Ltd.). Activity values are means of three individual determinations each performed in triplicate. In an initial screening, AChE inhibition was defined as minimum 50% inhibition at a concentration of 1.0 μg ml^-1. Subsequently IC₅₀ values were determined for all extracts deemed active according the initial screening. IC₅₀ values < 50 μg ml^-1 was considered active for the analysis. SERT activity was defined as more than 85% binding of extract at 5 mg ml^-1. Subsequently IC₅₀ values were determined for all extracts deemed active according the initial screening. IC₅₀ values < 50 μg ml^-1 was considered active for the analysis. These activity levels were designed to reflect the observed level of activity in the present study and do not necessarily reflect levels of pharmacological relevance, but within the range of proposed ecological relevance 84 . SERT activity data were not determined for eight species of Narcissus and these samples were pruned from the phylogenetic trees in the correlation tests.

We assessed the relationship of phylogeny to chemical diversity and biological activity by calculating the phylogenetic signal present for individual alkaloid types and types of biological activity. Each alkaloid type was coded as being either present (1) or absent (0) for each species. Likewise, for biological activity, species were scored for the presence or absence of AChE inhibition or SERT binding activity. Two of the alkaloid traits (belladine and cherylline) are found only in one species each, rendering calculation of phylogenetic signal for these traits meaningless, so they were not included in this part of the analysis.

To quantify phylogenetic signal we used the recently developed D metric 57 , specifically developed to deal with discrete binary traits. D is calculated as follows:

D = s obs − s ¯ b s ¯ r − s ¯ b

where s _obs is the observed number of changes in the binary trait (here, a chemical component) across the ultrametric phylogeny, s ¯ r is the mean number of changes generated from 1000 random permutations of the species values at the tips of the phylogeny, and s ¯ b is the mean number of changes generated from 1000 simulations of the evolution for the character by a Brownian motion model of evolution with likelihood of change being specified as that which produces the same number of tip species with each character state as the observed pattern. A D value of 1 ( s obs = s ¯ r ) indicates that the trait has evolved in a way that cannot be distinguished from a random manner (i.e., no phylogenetic signal), whilst a D value of 0 ( s obs = s ¯ b ) indicates that the trait has evolved in a phylogenetically highly correlated manner. Estimation of whether D differs significantly from 1 or 0 is achieved by evaluating where the observed number of changes (s _obs ) fits within the distribution of the 1000 generated s _r and s _b values respectively. Thus if 95% or more values of s _r are greater than s _obs then P (D = 1) ≤ 0.05 and the trait is significantly more phylogenetically structured than random expectation. Calculation of D was carried out using the packages caper 85 and ape 86 in the R v2.14.0 framework 87 .

We also quantified the relationship between overall chemical profile and phylogeny following an approach used to study the evolution of pheromone chemical diversity 57 58 . We constructed pairwise matrices; one of ultrametric phylogenetic distances (summed branch lengths) between species, and the other of chemical difference calculated as the binary squared Euclidean distance (i.e., the total number of alkaloid types that are absent in one taxon but present in another and vice versa). In addition to using all included species as terminal taxa, we also pruned the phylogenetic tree to genera and compared the resulting distance matrix to summed chemical profiles for each genus. In the case of polyphyletic genera both clades were retained. The correlation between phylogenetic distance and chemical difference was calculated using Mantel tests, with rows and columns of the distance matrix being randomly perturbed and the correlation coefficient recalculated 999 times to generate a null frequency distribution. These tests were performed using the program GenAlEx 88 .