Determining the photosynthetic relatives of Rafflesiales has long presented a challenge owing to the extreme reduction and/or modification of morphological structures that have accompanied the evolution of this lineage 311. Molecular phylogenetic approaches, although providing great promise in resolving such questions, also come with their own set of challenges that includes losses of some genes, substitution rate increases in other genes, and horizontal gene transfer. Examples of the first process can be seen in chloroplast genes such as rbcL that are typically used to infer phylogenetic relationships among angiosperms but have not yet been amplified from any Rafflesiales and are presumed lost 5. Increased substitution rates in the normally conservative plastid rDNA has been demonstrated in these holoparasites 412. Similarly, accelerated rates in mitochondrial SSU rDNA, typically very conservative in many photosynthetic angiosperms, occur in Rafflesia and Cytinus 13. Despite these complications, molecular phylogenetic analyses of some holoparasite lineages with comparatively lower rates have been tractable. For example, the mitochondrial genes atp1 and matR were used, in combination with nuclear rDNA and chloroplast genes, to reliably place Hydnoraceae with Aristolochiaceae 11.

Long-branch attraction, a bias in certain phylogenetic inference methods in which similarity due to convergent or parallel changes produces an erroneous phylogenetic grouping of taxa 10, is often implicated as the reason for anomalous phylogenetic groupings 14. It has been suggested that some data sets with marked among-lineage rate heterogeneity cannot be applied to particular phylogenetic problems owing to hypothesized long-branch attraction artifacts 15. In their unconstrained parsimony analysis of several angiosperm SSU rDNA sequences, Barkman et al. 9 found that the branch leading to Rafflesia was several times longer than any other branch, and that this branch was attracted to the second-longest branch in the tree – the one between gymnosperms and angiosperms. For these reasons, they argued that nuclear SSU rDNA sequences are of limited utility for assessing the phylogenetic position of Rafflesia.

Barkman et al. 9 analyzed their SSU rDNA data using only parsimony, not model-based methods (e.g., ML or BI methods) that are less likely to be misled by long-branch attraction 16. Our ML analysis of the SSU rDNA data recovers a topology that closely matches the matR topology presented by Barkman et al. 9 in which Rafflesia is closely related to Malpighiales and Mitrastema is a member of Ericales (Figure 4). These results highlight the requirement to analyze SSU rDNA data with methods less biased by long-branch attraction than parsimony, as well as the advantage gained by independent confirmation of results obtained from a single gene.

Several authors have suggested that adding taxa can "break up" long branches and allow parsimony to recover the correct topology 171819. Our parsimony analysis of the 103- and 77-taxon SSU rDNA data sets, in which we included representatives of all genera of Rafflesiales (i.e., sequences that could potentially break the Rafflesia long branch), recovers a nearly monophyletic Rafflesiales containing all of the longest terminal branches in the tree (see additional data file 3). Based on our simulation study and MP analyses of data sets from which all but one parasite group was removed, we believe that this topology represents a case of long-branch attraction. These simulation results support the contention that the branches leading to the parasitic taxa are long enough to attract one another (Figure 5), a result in agreement with previous work 36.

Taxon sampling is not a cure-all for long-branch attraction problems 20. Even for the data sets simulated on the full 77-taxon tree, MP returned incorrect long-branch clades nearly 80% of the time. MP did nearly as poorly with data sets simulated on a 77-taxon tree as it did on data sets simulated on a 20-taxon tree. Evaluation of the ML tree for the SSU data (Figure 4) shows that increasing the number of taxa from 20 to 77 did not improve the result because the long parasite branches were not broken. Instead, shorter (nonparasite) branches were broken which did not help MP recover the true topology for the simulated data sets. MP analyses of the full 77-taxon SSU data set that included all parasite clades resulted in a worse estimate of the phylogeny than MP analyses of smaller data sets in which only single parasite clades were included. Thus, the frequently stated view that increased taxon sampling can help MP avoid long-branch attraction problems may only be true if the added taxa are not distantly related long-branch clades themselves.

A requirement of the molecular phylogenetic approach to inferring evolutionary histories of organisms is vertical transmission of genetic material from parent to offspring. In contrast, horizontal gene transfer (HGT) describes the movement of genetic material between organisms of no direct ancestor-descendant relationship. Although the frequency of HGT is currently not well understood among prokaryotic and eukaryotic organisms, it is clear that HGT can compromise accurate inference of genealogical history. In plants, lateral movement of genetic material has been documented for mobile genetic elements such as introns 3334353637 but only recently has convincing evidence emerged documenting HGT of mitochondrial genes 3839. Genes of the mitochondrion are extensively used to infer evolutionary relationships in plants 404142, thus highlighting the importance of characterizing the frequency of HGT across genes and taxa.

Incongruence among gene trees derived from different data sets can derive from a number of factors such as technical causes (insufficient data, gene choice, sequencing error, taxon sampling and identification), gene/genome-level processes, and organism-level processes (e.g., hybridization/introgression, lineage sorting, and HGT) 43. HGT has only recently been recognized as a potentially important force in the evolution of plant mitochondrial genomes and detecting HGT is highly dependent upon the presence of multiple gene data sets with robust taxon sampling 3839.

We believe that incongruence between the the mitochondrial and the nuclear gene trees (Table 1) stem not just from long-branch attraction artifacts but also from cases of HGT. The placement of Apodanthes and Pilostyles on the atp1 tree as sister to Pisum (a legume, the family of hosts for Pilostyles) represents a likely case of HGT. The atp1 data conflict with those from matR that associates Apodanthaceae with Cucurbitales. Moreover, we infer that the SSU rDNA tree better represents the organismal phylogeny because it seems less likely that nuclear genes would be influenced by HGT 4445. The main rationale for this is that nuclear rDNA cistrons are repeated hundreds to thousands of times in tandem arrays at nucleolar organizing regions of the chromosomes. Although it can be envisioned that concerted evolution could homogenize all rDNAs in the parasite with a form obtained via HGT, the probability of this happening is small given the vastly different number of starting copies.

In their study of Rafflesiaceae s. str. and Mitrastemonaceae, Barkman et al. 9 discounted HGT as a possible explanation for their results because they state the phenomenon is rare and the overall topology of the matR tree closely matched results from other molecular phylogenetic investigations of angiosperms. The present study confirms that HGT is not implicated for the two lineages studied by Barkman as well as Cytinaceae, but this process could be invoked for Apodanthaceae. More recent work by these authors 46 identified several cases of HGT from host to parasite for atp1. These included Dalea to Pilostyles, Tetrastigma to Rafflesia, and Lithocarpus to Mitrastema. In addition, HGT of another mitochondrial gene, nad1, has been reported for Rafflesia and Sapria, both of which occur on the same clade as their hosts (Tetrastigma) on a gene tree 20. These examples demonstrating the presence of host genes in parasitic plants provide the most compelling evidence for HGT. This form of transfer is intuitively logical given the intimate contact between cells of the two organisms via the endophytic haustorium. However, parasitic plants exist in complex ecosystems where they are in physical contact with many other organisms (bacteria, fungi, phytophagous and pollinating animals, etc.) that could potentially affect HGT. That such nonhost HGT may also be occurring is evidenced by the presence of an apparent cucurbitalean matR gene in Pilostyles and Apodanthes. Moreover, present-day hosts of parasitic angiosperms do not represent the only conduit for HGT if host choice has shifted through time as the parasite lineage evolves. For example, Barkman et al. 9 state that Mitrastema only parasitizes Fagales (e.g., Lithocarpus and Castanopsis, both Fagaceae) but this parasite has also been recorded from Aquifoliaceae, Asteraceae, Elaeocarpaceae, Juglandaceae, and Myrtaceae 47. Host latitude for this species would be broader if rare hosts and hosts of parasite ancestors were fully known, thus expanding the taxonomic spectrum of potential HGT sources.

Contamination of parasite DNA with DNA from the host plant is an issue that must be given serious attention. Indeed, two sequences shown on the matR tree (Figure 1), Tetrastigma2 and Julbernardia are hosts for Rafflesia tuan-mudae and Berlinianche, respectively. These sequences were obtained by PCR amplification and sequencing from what was originally thought to be pure parasite genomic DNA. Sequences of the host (obtained from separate samples) were found to be identical to these "parasite" sequences, strongly suggesting contamination. In the case of Rafflesia, the DNA was obtained from a bud still attached to the host vine, both of which had been sectioned longitudinally. Disruption of these tissues likely resulted in transfer of host sap to the bud region where the tissue was sampled. Other samples of R. tuan-mudae from the same population, obtained as floral bracts with no host tissue, resulted in matR sequences that were similar to the other two Rafflesia species.

For Berlinianche, whose flowers are much smaller than those of Rafflesia (5 mm in diameter), extreme care (using a stereo microscope) was exercised to remove floral parts devoid of any host tissue. Despite this, the matR sequence obtained from the first sample was that of the host, Julbernardia. Later, silica gel dried samples of other populations of the parasite were extracted, again using extreme care in avoiding host contamination. PCR products were obtained using several mitochondrial matR primers, but none were found to be homologous to this gene following BLAST searches. This result shows that host DNA was not present in this sample in sufficient amounts to amplify and that the parasite matR gene, if present, is highly divergent at the priming sites used.

For all three Apodanthaceae genera, the conical style in female flowers is papillate and heavily secretory 31. This sticky surface tends to capture a variety of environmental debris, likely including extraneous pollen, fungal spores, and host tissues that have been disrupted upon collecting. Obtaining a proper nuclear SSU rDNA sequence for Pilostyles was extremely difficult. Despite PCR products of the correct sizes using a variety of primer combinations, the sequences obtained from genomic DNA derived from flowers were deemed contaminants following BLAST searches that showed them to be most similar to monocots, fungi, etc. Only when sequences from two accessions of Pilostyles (Texas and California) both were most similar to Malvales was this considered good evidence for their true phylogenetic affiliation. Retrospectively, it is likely that the sticky flowers had accumulated wind-dispersed pollen (e.g., grasses) and that this DNA, despite being in low concentration, had less divergent priming sites than the parasite target DNA, allowing PCR to preferentially amplify the contaminant DNA.

Given the accumulating molecular evidence for HGT from host to parasitic plant, it is worthwhile to consider potential mechanisms, along with their constraints, that may suggest further research. Relatively little information exists on the structure of the endophyte of Rafflesiales. Ultrastructural studies have been conducted on two species of Pilostyles: P. hamiltonii 48 and P. thurberi 49. These authors conflict, however, as to whether there exists symplastic continuity between host and parasite via plasmodesmata; the former indicated that such connections are the major path of nutrient uptake by the parasite whereas the latter rejected this idea. Despite this controversy, heteroplastic plasmodesmatal connections have been documented in another parasitic plant, Cuscuta 50 and indeed such connections can even form in heterografts between distantly related plant taxa 51. Given this, we assume that host-parasite plasmodesmatal connections exist in the endophytes of Rafflesiales. Transmission electron micrographs of Pilostyles suggest that intact, mature mitochondria are too large to pass through heteroplastic plasmodesmata, however, mitochondrial genomes or portions of the genome are certainly small enough for transmission. Once inside the parasite cell, there are various fates for the host gene. It could become incorporated into the parasite mitochondrial genome, and then either replace the parasite copy or exist as a duplicate, or the host gene could reside in the parasite nuclear genome. For the latter case, the gene would likely become a pseudogene given the requirement of mitochondrial-specific patterns of RNA editing. Two forms of atp1 are present in the primitive angiosperm Amborella trichopoda 38, one of which is derived from a HGT event from a eudicot. It is not known whether both forms of the gene exist in a single mitochondrial genome, in different mitochondrial genomes within the cell (i.e., heteroplasmy), or if one is nuclear and the other mitochondrial. Future work to address these questions would involve sequencing flanking regions of purported horizontally transferred genes to determine their subcellular location. Additionally, cDNA sequences obtained from matR mRNA would be useful to determine whether the gene is expressed and whether mitochondrial-specific RNA editing patterns are present.

DNA was extracted, amplified, cloned, and sequenced by using methods formerly reported 525354. The nuclear and mitochondrial sequences were PCR-amplified using primers reported elsewhere 65556 and are also given on the first author's web site 57. Sequencing was conducted using manual and automated methods (ABI Prism^® 377 automated DNA sequencer, Applied Biosystems) according to manufacturer's protocols.

The initial matR alignment incorporated all of the Rafflesiales parasites and the nonparasite sequences previously published 9 as well as our newly generated sequences. The 106-taxon matrix represented over 40 orders and contained three gymnosperm outgroup taxa (Ginkgo, Pinus, and Zamia), 28 monosulcates, 63 nonparasitic eudicots, and 15 Rafflesiales. For two taxa (Mitrastema and Rhizanthes), our sequences, as well as those previously published, were from the same species but different accessions to test for consistency. Taking into account codon information, an alignment of 2177 sites was constructed manually using SeAl version 2.0 58. The full matrix was used for parsimony analyses whereas another, truncated to 77 taxa by removing all but three monosulcate taxa (Laurales used as outgroup), was constructed to facilitate likelihood analyses. This operation was justified because monosulcates were never implicated as relatives of Rafflesiales in any analyses. A 71-taxon, 1265-site atp1 alignment was similarly constructed and included the same gymnosperm outgroup genera as above, 24 monosulcates, 32 nonparasitic eudicots and 12 Rafflesiales. All of the monosulcate genera in the atp1 alignment were also represented in the matR data set, whereas eudicot sampling for atp1 was constrained by sequences available on GenBank (12 of the same genera as with matR or placeholders from same family).

To test the position obtained for Rafflesiales taxa using mitochondrial genes with an independent data set derived from different compartments, a 4646-site "3-gene" matrix combining sequences from nuclear SSU rDNA and chloroplast rbcL and atpB was constructed that included 103 taxa (3 gymnosperms, 28 monosulcates, 58 nonparasitic eudicots, and 14 Rafflesiales). Sampling across angiosperm orders was very similar to the matR matrix, differing only by the presence of 11 placeholders and a second accession of Pilostyles. For the holoparasites, only nuclear SSU rDNA sequences were included; the chloroplast gene data for these taxa were coded as missing. The two chloroplast genes were included to add stability to the tree topology given that nuclear SSU has been shown to contain lower phylogenetic signal when used alone 15. As with matR, the 103-taxon matrix was truncated to 77 taxa by removing all but five monosulcate taxa to facilitate likelihood analyses. All alignments reported in this paper have been deposited with TreeBASE 59: study accession number S1177, matrix accession numbers = M2034–M2037.

All three data sets were analyzed using maximum parsimony (MP) and maximum likelihood (ML) methods in PAUP* 4.0b10 60 and Bayesian inference (BI) methods in MrBayes 3.0b4 61.