Correlated evolution of life history and host range in the nonphotosynthetic parasitic flowering plants Orobanche and Phelipanche (Orobanchaceae)


Gerald M. Schneeweiss, Department of Biogeography and Botanical Garden, University of Vienna, Rennweg 14, A-1030 Vienna, Austria.
Tel.: +43-1-4277-54062; fax: +43-1-4277-9541; e-mail:


Most members of the nonphotosynthetic parasitic genera Orobanche and Phelipanche (Orobanchaceae) have narrow host ranges, and, as they grow on perennial hosts, are (at least potentially) perennial themselves. A few species, however, have wide host ranges and grow on annual hosts, and are thus (at least facultatively) annuals themselves. Among the latter are the weedy species, which include economically important pest taxa such as Orobanche crenata or Phelipanche aegyptiaca. Using a phylogenetically based maximum likelihood approach, which takes phylogenetic and branch length uncertainty into account, we can show that the life trait host range and life history evolve in a correlated fashion. This supports the hypothesis that parasite specialization is associated with predictable resources (i.e. long-lived hosts) and generalism with unpredictable ones (i.e. short-lived hosts), a pattern often found in animal parasites. The mechanisms and temporal sequence of the life trait changes and their interrelations remain speculative.


Among angiosperms, c. 4000 species (i.e. approximately 1% of known angiosperm diversity: Nickrent et al., 1997) are parasitic and as such rely partially or completely on a host plant, from which they obtain nutrients and water. Depending on the presence or absence of functional chloroplasts, and therefore photosynthetic activity, parasitic plants can be classified as hemi- or holoparasites (Musselman & Press, 1995). Parasitic plants are understood to play an important role in ecosystems, affecting, for instance, species diversity and nutrient cycling (Press & Phoenix, 2005; Bardgett et al., 2006).

Two important life traits of parasitic plants that potentially affect their performance at the ecosystem level are host range and life history. This also appears to be the case in Orobanche and Phelipanche (Orobanchaceae in the circumscription of Young et al., 1999), members of the largest group of holoparasitic plants. The majority of their species have narrow host ranges and grow exclusively on perennial host plants, thus being at least potentially perennial themselves. This trait combination has been proposed as the ancestral one in Orobanche and Phelipanche (Manen et al., 2004; G. M. Schneeweiss, unpublished) and agrees with the hypothesis that parasite specialization tends to be associated with predictable resources (Ward, 1992), i.e. long-lived hosts, as shown in animal parasites (Sorci et al., 1997; Sasal et al., 1999; Krasnov et al., 2006; Šimkováet al., 2006). These species are usually found in natural ecosystems, where they are often uncommon compared with their potential hosts, and they do not harm their hosts in an obvious way (Teryokhin, 1997). In contrast, a few Orobanche and Phelipanche species have wide host ranges and grow on annual hosts, and are thus (at least facultatively) annuals themselves. These species are usually found in anthropogenically disturbed ecosystems, where they are often frequent and can cause severe to lethal damage to their host plants, thus showing a weedy behaviour (Parker & Riches, 1993; Teryokhin, 1997). This is particularly pronounced in economically important pest species such as Orobanche crenata and Phelipanche aegyptiaca, which can cause major yield losses (Riches & Parker, 1995; Abu-Irmaileh, 1998; Musselman et al., 2001). It is therefore hypothesized that shifts in host range and life history play an important role in the evolution of weediness in Orobanche and Phelipanche.

As a basis for a better understanding of the evolution of weediness, this study investigates whether host range and life history evolve in a correlated fashion. Using a phylogenetically based maximum-likelihood procedure, which takes phylogenetic and branch length uncertainty into account, the association between shifts in host range and changes in life history is investigated.

Materials and methods

Phylogenetic analysis

Phylogenetic hypotheses based on DNA sequence data from the nuclear ribosomal region were established for members of the related genera Orobanche and Phelipanche (Schneeweiss et al., 2004; Carlón et al., 2005; Table S1). Although this marker is known to be potentially problematic (Bailey et al., 2003; Álvarez & Wendel, 2003), at present this data set is the one with the broadest taxon coverage available, including about one-third of the known species diversity (c. 60 of c. 170 species; Uhlich et al., 1995). Phylogenetic analyses were conducted in a Bayesian framework using the programme bayesphylogenies (available at: Apart from the usual advantages of a Bayesian inference, i.e. the incorporation of phylogenetic and branch length uncertainty (Huelsenbeck et al., 2001; Holder & Lewis, 2003), this programme additionally allows the detection of pattern-heterogeneity in sequence or character-state data without a priori assignment of partitions (Pagel & Meade, 2004). For both data sets, incorporation of more than one rate matrix did not significantly improve the likelihoods, whereas modelling rate heterogeneity across sites with a gamma distribution did (data not shown). For this reason, single DNA substitution models with six substitution rates and rate heterogeneity modelled with a gamma distribution were used (thus corresponding to the model suggested by the hierarchical likelihood ratio test implemented in modeltest 3.7: Posada & Crandall, 1998). All parameter values were estimated during the analyses. Trees from the posterior distribution were combined from two independent runs (3 × 106 generations each, sampling every 300th tree) after model parameters reached stationarity (burn-in of 1.5 × 105 generations, i.e. 500 trees each), resulting in a posterior set of 19 000 trees.

Analysis of character state evolution

Life trait evolution was investigated using bayesmultistate (Pagel et al., 2004; available at:, which uses a maximum-likelihood approach to character state reconstruction (Pagel, 1997, 1999). The model used here for reconstructing character states is one of a gradual mode of trait evolution, and therefore longer branches will contribute more to character state changes than shorter branches. Using a scaling parameter (denoted κ) would allow stretching or compression of phylogenetic branches, and κ = 0 would render trait evolution independent of branch length (Pagel, 1994), showing a punctual mode of trait evolution. Although the use of this scaling parameter is suggested by results from initial analyses with discrete 4.0 (available at:; data not shown), at the time of conducting the analyses its use was not yet implemented in bayesmultistate. To avoid entrapment in local optima, 20 trials of likelihood estimation per tree were conducted (command mltries = 20). Prior to analyses, out-group taxa and identical sequences of the same taxon were pruned from the trees. For Phelipanche, however, one out-group taxon (Aphyllon uniflorum = Orobanche uniflora) had to be maintained because of the uncertain relationships of the main lineages in Phelipanche. To investigate the influence of the inclusion of A. uniflorum on the results, it was coded: (i) as a perennial with a wide host range (the actual states for this species), and (ii) as ambiguous for both traits. Life history and host range were coded as binary characters based on data obtained from the literature (Beck-Mannagetta, 1930; Novopokrovskij & Tzvelev, 1958) and personal observations (Table S1): (at least potentially) perennial vs. (at least facultatively) annual; narrow host range (i.e. attacking members of one plant family only) vs. wide host range. In some cases, unambiguous coding of traits was not possible, and these were coded as ambiguous (character state 01). In many cases, the actual lifespan of the parasite is not known, so the lifespan of the host was used instead. Hence, it was not the trait ‘life history of the parasite: annual vs. perennial’ that was investigated, but rather the trait ‘the necessity for the parasite to complete its own life cycle, determined by the life span of the host, within: 1 year vs. two or more years’. The relatively coarse coding scheme for host range at the level of host families was necessitated by the lack of sufficiently detailed knowledge for many species, although a more refined coding scheme might be desirable for some species (e.g. Phelipanche purpurea on Achillea spp. and Artemisia vulgaris vs. Phelipanche bohemica on Artemisia campestris, all hosts belonging to Asteraceae).

The following analyses were conducted on two data sets for each group, one including all taxa (henceforth called data set ‘all’), the second including only taxa with unambiguous character state assignments (henceforth called data set ‘unambiguous’; Table S1). To test for correlated character evolution, the likelihood of a model in which each trait evolves independently (with a total of four parameters) is compared with the likelihood of a model where traits evolve in a correlated fashion (with a total of eight parameters; Pagel, 1997). Statistical significance of the observed likelihood ratios was assessed as follows: using those trees that resulted in the smallest and the largest observed likelihood ratios, a null distribution of the likelihood ratio was generated using the Markov chain simulation procedure as implemented in discrete 4.0, employing 1000 replicates (simulation 1 and simulation 2). As this program does not allow ambiguous character state assignments, only trees from the data sets ‘unambiguous’ were used for the simulation. The observed likelihood ratio is considered significant: (i) if the smallest observed likelihood ratio is larger than the upper 95% or more of the simulated likelihood ratio (method I; this method is expected to have low type I error rate); and (ii) if the smallest value of the central 95% of the observed likelihood ratio values is greater than the largest simulated likelihood ratio (method II). The null distribution of the likelihood ratios only applies to the observed likelihood ratios from data sets ‘unambiguous’. As data sets ‘all’ result in smaller observed likelihood ratios (see Results), the null hypothesis of no difference in likelihood ratios is less likely to be rejected, rendering this test for data sets ‘all’ conservative. Using a similar approach to Dunn et al. (2005), the two hypotheses of correlated and noncorrelated evolution were also compared via a likelihood ratio test with four degrees of freedom, using the arithmetic mean and median of the log likelihoods of all trees in the posterior distribution.


The majority of Orobanche and Phelipanche species have narrow host range and perennial life history (Table S1). A few species show wide host range and annual life history (e.g. Orobanche cernua var. cumana, O. crenata, Phelipanche aegyptiaca, P. mutelii; Fig. 1). In Orobanche, a few species occurring in the same group as some taxa with trait combination wide host range/annual life history (Orobanche minor, O. crenata) show wide host range but perennial life history (Orobanche owerinii, Orobanche pubescens, Orobanche transcaucasica; Fig. 1). The latter trait combination is as yet unknown from Phelipanche species, but present in the out-group species A. uniflorum.

Figure 1.

 Phylogenetic relationships of Phelipanche (a) and Orobanche (b) inferred from a Bayesian analysis of nuclear ITS-data (50% majority-rule consensus tree). Taxon names are given with the respective GenBank accession numbers (see Table S1). Numbers at nodes are posterior probability values ≥ 0.5. An asterisk and a plus indicate the clades, for which the posterior distributions of character state probabilities are shown in Figs S1 and S2.

The obtained ranges of likelihood scores and the corresponding likelihood ratios of the uncorrelated and correlated models of trait evolution of different data sets of Orobanche and Phelipanche are summarized in Table 1. Although different coding schemes of the out-group taxon for Phelipanche affect the likelihood scores, the obtained likelihood ratios are very similar (Table 1). In Phelipanche, all methods used for assessing significance of the differences in likelihood scores favour the model of correlated character evolution, irrespective of the coding scheme used for the out-group Aphyllon (Table 2; Fig. 2a). Significance in the Orobanche data set is generally lower (i.e. P-values are higher) than that in the Phelipanche data set, but still most methods favour the model of correlated character evolution. The exception is method I, which does not reject the simpler model of uncorrelated evolution, because here more than 5% of simulated likelihood ratios exceeded the smallest observed value (Table 2, Fig. 2b). This method, however, is expected to be very conservative with a high type II error rate.

Table 1.   Ranges of likelihood scores (−ln) and likelihood ratios (LR) of the uncorrelated and correlated models of trait evolution of different data sets of Orobanche and Phelipanche, given as (maximum) 95% credible interval (minimum); arithmetic mean/median.
  1. Data sets ‘all’ and ‘unambiguous’ refer to data sets including all accessions and only accessions with unambiguous character state assignment respectively. For Phelipanche, ‘original’ and ‘alternative’ refer to coding the out-group Aphyllon uniflorum as having the observed character states (wide host range, perennial life history [1, 0]) or ambiguous character states (01, 01).

ln uncorrelated
(44.403) 40.162– 31.878 (28.572); 36.100/36.104(44.171) 39.701– 31.526 (28.024); 35.614/35.619(22.252) 18.778– 17.633 (15.389); 18.608/18.739(19.791) 17.549– 16.297 (13.861); 17.380/17.546(17.888) 17.323– 16.109 (13.698); 17.144/17.283(16.632) 16.013– 14.697 (12.178); 15.842/16.010
ln correlated
(40.537) 34.548– 27.784 (24.824); 31.026/30.978(37.185) 33.697– 27.235 (24.394); 30.330/30.289(15.139) 13.700– 10.846 (9.071); 12.352/12.401(13.367) 11.647– 9.305 (8.040); 10.645/10.712(12.961) 11.268– 8.575 (6.964); 9.807/9.754(11.431) 9.515– 7.214 (5.804); 8.346/8.325
(8.533) 6.783– 3.328 (−0.180); 5.073/5.079(8.613) 6.935– 3.608 (2.434); 5.283/5.283(9.133) 7.353– 5.040 (3.463); 6.256/6.254(9.482) 7.600– 5.898 (4.264); 6.735/6.735(8.464) 8.021– 6.009 (4.421); 7.338/7.447(8.493) 8.039– 6.495 (4.662); 7.496/7.582
Table 2.   Significance (P-values) of the observed likelihood ratios inferred from (i) a likelihood ratio test with four degrees of freedom using arithmetic mean and median of the log-likelihoods of all trees in the posterior distribution and (ii) comparison with a null distribution generated by Markov chain simulation.
  1. aThis is based on the second smallest observed likelihood ratio of 1.699. The smallest value of −0.18 would result in P-values of 0.999 for simulation 1 and 0.998 for simulation 2.

  2. Data sets ‘all’ and ‘unambiguous’ refer to data sets including all accessions and only accessions with unambiguous character state assignment respectively. For Phelipanche, ‘original’ and ‘alternative’ refer to coding the out-group Aphyllon uniflorum as having the observed character states (wide host range, perennial life history [1, 0]) or ambiguous character states (01, 01). Simulation 1 and 2 refer to the null distribution of the likelihood ratio obtained from the trees with the smallest and largest observed likelihood ratios respectively (see Materials and methods for details). P-values other than those from the likelihood ratio test are calculated using the two methods (I and II) described in the text.

Arithmetic mean0.0380.0320.0140.0090.0050.005
Simulation 1
 Method I0.173a0.0700.0030.00100
 Method II0.0200.0150000
Simulation 2
 Method I0.191a0.0750.003000
 Method II0.0160.0090000
Figure 2.

 Distributions of observed (solid and dotted lines) and simulated (dashed lines: thick for simulation 1 and thin for simulation 2) likelihood ratios between models of uncorrelated and correlated character state evolution obtained from 19 000 data points, using data sets ‘all’ and ‘unambiguous’ (thick and thin lines respectively) in (a) Phelipanche (using the original or alternative coding of the out-group: solid and dotted lines respectively) and (b) Orobanche. See text for details on coding and data sets definition.

As part of the analysis of character state evolution employed in this study, rates of character change and ancestral character states are estimated. Unfortunately, the values for the rates of character change vary tremendously (e.g. from 0 to > 5000 for the change from annual life history/narrow host range to annual life history/wide host range in Orobanche), indicating that the likelihood surface for these parameters is flat, and that a broad range of values essentially returns the same likelihood. Consequently, the posterior distributions of character state probabilities at a given node are very broad (sometimes ranging from 0 to 1). This is particularly pronounced at nodes subtending groups with different character state combinations, as in O. minor and its relatives (Fig. S1), but it also affects nodes subtending groups with identical character state combinations, as in Orobanche teucrii and Orobanche caryophyllacea (Fig. S2). For instance, a probability > 0.05 of a trait combination wide host range/annual life history in the ancestor of O. teucrii and O. caryophyllacea, which a priori would be considered very unlikely because both species have narrow host range/perennial life history, still has a posterior probability ≫ 0.05 (Fig. 2d) and can thus not be rejected with any confidence. For this reason, neither the inferred rates of character changes nor the ancestral character states of certain nodes are discussed further.


Evolution of host range and life history

The majority of Orobanche and Phelipanche species have a narrow host range and grow on perennial hosts (Beck-Mannagetta, 1930; Uhlich et al., 1995; Teryokhin, 1997). An association between parasite specialization and resource predictability (host longevity) is expected (Ward, 1992). This has been repeatedly shown for animal parasites, e.g. fleas or monogenean flatworms (Krasnov et al., 2006; Šimkováet al., 2006), and is suggested here for parasitic plants. It is likely that host recognition, being the first phase of host–parasite interaction, is important for this distribution of life traits. Host recognition comprises several steps, which are all equally important for the successful establishment of a connection to the host, such as germination and haustorium formation (Bouwmeester et al., 2003). Each step invokes chemical signalling between the host and the parasite. For example, seed germination and haustorium development occur only after stimulation by chemical signals released by the host (xenognosins; Atsatt et al., 1978; Keyes et al., 2001; Yoder, 2001; Bouwmeester et al., 2003). Each step can determine host specificity and thus host range, e.g. a plant might produce germination stimulants and hence induce germination of the parasite, but may be resistant in later stages of the parasite's lifecycle (e.g. Chittapur et al., 2001). For obligate parasites, which entirely depend on their hosts, it is crucial to specifically respond to signals from the correct hosts to maximize the chances of a successful establishment of a connection. One would therefore expect most Orobanche and Phelipanche species to be very specific and grow only on a few host species to which they are well adapted, and thus to have narrow host ranges. Although annual plants are suitable hosts due, for example, to weaker lignification of host tissue allowing easier penetration, or to decreased defence mechanisms compared with perennials (Feeney, 1976; Coley et al., 1985), they impose severe time constraints on the parasite's development. These time constraints might explain the prevalence of perennial hosts.

From the available data it is not possible to infer the temporal sequence of character state transitions. Using a parsimony based method of character state reconstruction, it has been suggested that the ancestral character states are narrow host range and perennial life history (Manen et al., 2004; G. M. Schneeweiss, unpublished). This fits the scenario outlined above, which suggests that holoparasitic plants are expected to be well adapted to a few host species only. However, the maximum-likelihood method employed here reveals that the ancestral character states are totally ambiguous, that is, each of the four trait combinations in the model of correlated evolution has a probability of 0.25 (data not shown).

A possible mechanism involved in changes of host ranges is a heritable polymorphism in response to xenognosins, as is known for the hemiparasitic Orobanchaceae Triphysaria pusilla (Jamison & Yoder, 2001). This can result in the formation of intraspecific host races, as is suggested from glasshouse experiments for Orobanche and Phelipanche (Musselman et al., 1981; Boulet et al., 2001). These studies do not directly focus on this question, however, and further experiments incorporating more Orobanche strains are required. Other scenarios, such as changes in the xenognosin receptor(s) or changes in the required threshold of xenognosins, concerning both duration of exposure (Chang & Lynn, 1986; Smith et al., 1990) and quantity of exudates, are currently not testable because of insufficient data.

The reasons for the association between wide host range and annual life history are elusive. One hypothesis is that with an increasing number of potential host species the probability of attacking annual hosts simply increases as well. If correct, a positive correlation between number of attacked host species and number of annual host species is expected, which could be tested if sufficiently detailed and correct data were available. A second hypothesis is based on the proposal that species with a wide host range have a higher chance of finding a suitable host than species with a narrow host range, and will therefore generally have larger populations. It may be that in the transitional phase from perennial to annual hosts a considerable number of parasites, being not yet perfectly adapted, are not able to finish their life cycle on annual hosts. This loss of offspring is expected to be less critical in larger populations. Although demographic data of species with contrasting trait combinations would allow us to test this hypothesis, such data, at least for natural habitats, are lacking. In man-made or anthropogenically disturbed habitats, species with wide host range and annual life history often occur in large numbers, but it is unclear if this is because of the life traits or because of specific conditions of the habitats. Host unpredictability because of low host abundance has been suggested to favour parasite generalists (Norton & Carpenter, 1998; Tripet et al., 2002), and host longevity (annual life history) may play a similar role (Šimkováet al., 2006). Facultative autogamy is known for several weedy Orobanche and Phelipanche species (Musselman et al., 1981), and this pollination mode may play an important role in the fixation of mutations advantageous for growing on annual hosts. Comparative studies using closely related species with contrasting life traits, such as those in the clade comprising O. minor and its relatives (Fig. 1b), would clearly help to address such questions.

Evolution of weediness

Of the approximately 170 species of the genera Orobanche and Phelipanche, few are weedy species (e.g. Orobanche picridis, Phelipanche mutelii, P. nana) and only five have become major pests (Orobanche crenata, O. cumana, O. minor, Phelipanche aegyptiaca, P. ramosa; Parker & Riches, 1993; Riches & Parker, 1995; Teryokhin, 1997). These taxa share wide host ranges and annual life histories, whereas the majority of Orobanche and Phelipanche species have narrow host ranges and perennial life histories (Table S1). This suggests that these traits play an important role in the evolution of weediness in Orobanche and Phelipanche. It is obvious that annual life history is an obligate prerequisite for a parasite growing on annual host plants, which usually dominate in anthropogenically disturbed habitats such as agricultural fields. The role of wide host range in virulent Orobanche races is less clear, although it might indirectly facilitate the evolution of annual life history.

Possible limitations

Several factors warrant caution in the interpretation of the above results. These include the incompleteness of sampling, which can affect the distribution of character states along the phylogeny and thus bias the estimation of rate parameters. For example, unsampled species might have trait combinations not encountered in the current data set, or species belonging to well-defined clades within their respective genera might show trait combinations differing from their close relatives. This situation is complicated by the taxonomic uncertainty of the species, especially in less explored regions such as Near and Central Asia, where major components of the species diversity of both Orobanche and Phelipanche reside (Beck-Mannagetta, 1890, 1930; Teryokhin et al., 1993; Uhlich et al., 1995).

Several potential biases stemming from the method used have been recently addressed in the context of ecological specialization of insects (Nosil & Mooers, 2005; Stireman, 2005). These include, for example, higher transition rates to the more common state. Currently implemented methods of character state reconstruction assume that the rate of character change is the same in each group over the whole phylogeny. The actual distribution of character states in both Orobanche and Phelipanche would, however, suggest the presence of rate heterogeneity. For instance, in Orobanche, species with annual life history are confined to two clades (O. cernua s.l., O. minor and related species), which significantly also include very closely related taxa with perennial life history. The situation is similar to that of a molecular clock in models of sequence evolution, which has been found to be violated in most of the real data sets (Bromham & Penny, 2003; Renner, 2005). It is expected that in analogy to local molecular clock or relaxed molecular clock approaches (reviewed in Renner, 2005; Welch & Bromham, 2005), future models of character evolution will allow variation in the rate of character change.


I thank Mark Pagel for help with the program discrete. This study was supported financially by the Austrian Science Fund (FWF P14352-Bio).