In this project, five pairs of plant-pathogen were selected (Table 1). The rust fungus Melampsora larici-populina is a major threat in European poplar plantations [24]. Biotrophic rust fungi are some of the most devastating pathogens of crop plants. The biotrophic bacterium Xanthomonas oryzae pv. Oryzae causes bacterial blight of rice (Oryza sativa L.), which can cause reductions of rice yields of as much as 50% in some areas [25]. The necrotrophic fungal pathogen Sclerotinia sclerotiorum causes stem rot or white mold on soil-borne plants in more than 500 species of plants globally [26], including the important feedstock soybean (Glycine max). The plant model species Arabidopsis thaliana is infected by the hemi-biotrophic pathogen Pseudomonas syringae pv. Tomato [1] and this interaction is widely used to study the underlying mechanisms of pathogen attack [6]. The biotrophic fungi Ustilago maydis can cause smut disease in maize [27], even though this pathogen is not a major pest in crop plants. All protein sequences used to build the species-specific metabolic networks for all 10 species were downloaded from the NCBI protein database (see description in Table 2). The non-redundant “NR” database used in the BLAST [28] analysis was downloaded as of Jan 9, 2011.

We built draft genome-wide metabolic networks for each plant and its pathogen based on the MetaCyc reference metabolic network [42], and the protein databases Pfam [43], UniProt [44], and NCBI. The annotation of enzymes in the genomes of the 10 species investigated here was based primarily on applying hidden Markov models (HMMs). First, all reactions from MetaCyc with at least one annotated protein sequence were extracted. If for such a reaction a corresponding Pfam domain was annotated, the associated profile-HMM as provided by Pfam was used for further computations. If no Pfam family was annotated, but a four-digit EC number for the reaction was given, all protein sequences annotated with the corresponding EC number were downloaded from UniProt and used to build HMM-models (see below). If fewer than 20 protein sequences were available for a given EC number, standard BLAST [28] with a score cut-off of 50 and an E-value threshold of 10^-10 was used to collect additional significant sequence hits from the non-redundant sequence database “NR” to allow creating HMM-models with a sufficient number of sequences. 591 HMMs were derived from Pfam, 1192 HMMs from UniProt associations with EC-numbers, and 399 HMMs needed additional protein sequences obtained by BLAST searches. Afterwards, all reaction-specific protein sequences were aligned using the multiple sequence alignment program MAFFT [45]. The resulting multiple sequence alignments were converted into a reaction-specific profile HMM using the HMMER program [46]. HMMER transforms a multiple-sequence alignment into a probability based position-specific scoring system. Finally, all protein sequences from each species were searched with HMMER using all reaction-specific HMMs. For every reaction HMM, the protein with the lowest E-value score and below the cut-off of 1 was assigned the annotation associated with the respective HMM. The workflow is shown in Figure 1. Thus, a given HMM is assigned to only one protein in a species’ genome at most. While in reality a particular enzyme may exist in multiple isoforms, for the purpose of network reconstruction a reaction can proceed as long as there is at least one enzyme catalyzing it. However, a single protein may carry more than one annotation as it may be identified as the best hit by more than one HMM. First, proteins may indeed carry out multiple reactions (e.g. acting on different substrates), and secondly, using the bioinformatics annotation protocol alone, ambiguous assignments cannot be easily resolved, but would require experimental verification. For the sake of network capacity, we decided to accept ambiguous, but significant assignments.

The draft metabolic network models derived by the method explained in the previous section were further curated by removing reactions that are stoichiometrically inconsistent in the sense that the sum formulas on both sides of the chemical equations do not yield the same number of atoms. Mass-balance was checked in reactions for which all involved compounds are annotated with a chemical formula. If elements were net-produced or net-consumed violating mass conservation (neglecting protonation state and water production/consumption), the reaction was removed. Reactions marked spontaneous in the MetaCyc database were added to all organism-specific metabolic networks. The resulting plant networks were then extended using the method introduced in [47]. Briefly, the metabolic networks are assumed to produce essential biomass precursors (amino acids, nucleotides) from autotrophic nutrients consisting of carbon dioxide, water, oxygen, protons and all other inorganic metabolites found in the MetaCyc database. Further, we added ribulose bisphosphate to ensure the functioning of the Calvin cycle. If the draft network is not able to produce the essential biomass precursors, a greedy approach identifies reactions from the MetaCyc database that should be added to the network to fulfill this requirement. It was not possible to apply this approach to the pathogen networks, because their nutrient requirements are presently largely unknown.

When two metabolic networks interact, they can exhibit positive or negative effects on each other. If the networks “cooperate” and combine their biochemical resources, increased biosynthetic capabilities may result from the interaction. A concept to quantify this symbiotic effect was introduced in [20], where the ‘metabolic gain’ was defined. The gain exists for any combination of initially available metabolites, called the seed. To calculate the gain, two networks are combined by simply forming the union of their biochemical reactions. Then, the scope of the seed, which is defined as the set of metabolites that can be produced from the seed as determined by the method of network expansion [19], is calculated for the individual networks and the combined network. The gain is then defined as the number of metabolites in the scope of the combined network minus the number of metabolites in the union of the scopes of the single networks. The asymmetric gain describes the positive effect on the individual organism, by subtracting the number of metabolites in the scope of a single network from the number of metabolites in the scope of the combined networks intersected with all compounds from the single network. To calculate the gain, we used as seed the combination of nutrients on which the plant networks can grow (see above), i.e. produce all necessary biomass precursors.

Calculating negative effects of one network on another is less straightforward than capturing positive effects. To measure how the pathogen can impair the plant network, we first determined minimal combinations of nutrients on which the pathogen can survive (i.e. produce all necessary biomass precursors). Typically, the minimal nutrient seed compound set consisted of three to six compounds depending on the pathogen. In addition, water, oxygen, and protons were always provided. We employed the method described in [48] with the modification that the nutrient compounds are restricted to metabolites present in the plant network. In this way, possible nutrient combinations, which the pathogen may extract from the plants, were determined. The method was based on a statistical sampling procedure and is therefore repeated 10,000 times for each pathogen network. To favor small molecules as nutrients, the sampling was not performed according to a uniform random number distribution, but by using a Boltzmann-factor-based shuffling procedure (for details see [49]), using the compound masses, m, as "energies" and a “thermodynamic beta” of 0.05 1/Da. Two compounds with the mass difference Δm are then exchanged with the probability of min(e^-beta*Δm,1). For each solution, i.e. a set of nutrients on which the pathogen can thrive, the possible negative effect on the plant network caused by the withdrawal of the corresponding metabolites was determined as follows: All reactions that use as substrate at least one of the determined pathogen nutrient compounds were removed from the plant network. This corresponds to the extreme scenario in which the pathogen extracts the entire compounds from the plant and reactions requiring this compound in the plant are no longer possible. Employing network expansion [19], this reduced network was then used to determine, which biomass compounds that are essential for the plant can still be produced. The relative impairment is then defined as the fraction of minimal nutrient combinations, for which a particular essential biomass compound can no longer be produced. For example, a relative impairment of 0.3 for the plant biomass component lysine means that in 30% of all calculated minimal nutrient combinations one or several nutrients have been removed by the pathogen from the plant, which render lysine not producible. It is in the nature of this analysis that only the effect of the pathogen on the host is evaluated. Therefore, no ‘symmetric’ score, as in the case of the positive effects discussed above, exists.

Vectors containing the impairment scores were defined to characterize impairment patterns of host-pathogen interactions and pairwise distances between these vectors were calculated. Based on these distances, we applied Multidimensional Scaling (MDS) to visualize similarities of the impairments in a two-dimensional diagram. The 'cmdscale' routine of R was used for the MDS computations.

The metric for the comparison of metabolic networks was based on Jaccard coefficients (JC) applied to sets of present reactions. The Jaccard coefficient between two sets is defined as the number of elements in the intersection divided by the number of elements in the union of the two sets. It is one for identical sets and zero for completely disjoint sets. We defined the Jaccard distance between two networks as d(N1,N2) = 1-JC(N1,N2), where JC is the Jaccard coefficient as defined above.

For all five pairs of plant and pathogen, we mapped the reactions onto the KEGG [48] reference metabolic pathway map based on the MetaCyc annotation. If there was no corresponding KEGG reaction annotation in MetaCyc, the EC number was used to map between KEGG and MetaCyc. Using the KEGG html-based visualization, all pathway maps can be zoomed in and out and can be queried. All metabolic network maps for all five plant-pathogen pairs are available as Supplementary Material 1; the related reactions for all the species are available as Supplementary Material 2.

Our established annotation workflow (see Figure 1) resulted in 10 genome-scale metabolic draft networks, five representing the metabolism of plants and five the metabolism of phytopathogens, respectively.

The applied extension method suggested adding one reaction (catalyzed by imidazoleglycerol phosphate synthase) to all plant networks in order to produce histidine. No alternative reactions were found. Because no sequence information for enzymes catalyzing this reaction is available, a closer inspection whether these enzymes are indeed coded in the genome was not possible. For Arabidopsis thaliana, several reactions were additionally suggested in order to produce thymidine triphosphate (TTP). For the metabolic networks of Glycine max and Zea Mays, more reactions had to be added to enable the network to produce required biomass precursors from carbon dioxide and inorganic material. The numbers of added reactions are given in Table 3.

The sizes of the networks, measured in terms of numbers of reactions and numbers of metabolites, are summarized in Table 3. On average, plant networks are larger than pathogen networks (p_t-test = 0.01) with 3082 curated reactions present on average in plants, 2906 in fungi, and 2690 in bacteria. Despite its significance, the difference in size of the plant and bacterial metabolic networks is surprisingly small. It has to be borne in mind that the employed network reconstruction procedure relied on mapping known enzyme sequences onto novel protein sequences. As a consequence, these networks are biased to well-known and common enzymatic reactions. Thus, due to limitations in our current knowledge, the large secondary metabolism of plants is hugely underrepresented.

Beyond the size differences, the question whether the content of the networks are rather similar or display significant differences; i.e. whether the set of enzymes and reactions are overlapping or disjoint. Judged by the Jaccard distance of present reactions as a measure to compare two metabolic networks (see Methods), the five plant metabolic networks are most similar to each other (Figure 2), whereas the pathogen networks are much more heterogeneous. The metabolic networks of the bacterial pathogens (Xo and Ps) clearly differ from those of fungal species (Um, Ml, Ss) with the bacterial species representing the most distinct networks compared to all other species considered here. In fact, the metabolic reaction-based similarities of the networks reflect the respective evolutionary origin of the 10 different species (Figure 2).

A central goal of our work was to establish draft networks for various plant-pathogen pairs, which may offer a platform for further investigations. To facilitate the inspection of how the host and the pathogen networks are overlapping and interacting, we generated navigable pathway maps that are based on the KEGG [48] metabolic maps. As an example, Figure 4 displays the visualization of the interacting network pair Pseudomonas syringae pv. Tomato and Arabidopsis thaliana. Green edges represent Arabidopsis thaliana specific reactions, yellow edges represent Pseudomonas syringae pv. Tomato specific reactions, black reactions are common to both networks. Grey edges represent reactions not contained in any of the two species. All network pairs are accessible as interactive maps with links to the corresponding KEGG metabolite and reaction entries in the Supplementary Material 1.

The pathogen gaining its resources from the host, thus inflicting damage to it, constitutes the hallmark of any host-pathogen relation. This relation is asymmetric in the sense that the host can survive better in the absence of the pathogen, while in general, the pathogen is dependent on nutrients or other components provided by the host. At the metabolic level, it is therefore plausible to assume that the presence of the metabolism of the pathogen has a negative effect on the plant metabolism, whereas the existence of the plant metabolism is positive for the pathogen. To analyze and quantify negative and positive effects of interacting metabolic networks, we calculated a ‘gain’ measuring the positive effect and an ‘impairment’ measuring the negative effect that one network has on the other. Essentially, the gain measures the number of metabolites, which can be produced more by a combined network compared to the sum of the single networks and the impairment is assessed by calculating the negative effect that removal of required nutrients of the pathogen has on the host’s capability to produce essential biomass precursors (details are given in the relevant section in Methods).

We calculated the metabolic gain for each plant-pathogen pair by assuming the photoautotrophic plant seed consisting of carbon dioxide and inorganic nutrients. As described in the relevant paragraph in the Methods, we calculate a symmetric gain, which describes the overall increase of producible metabolites, and asymmetric gains, which specify the advantages for each partner by only including metabolites occurring in the network of the respective interacting partner. This plant asymmetric gain describes how many new compounds could in principle be synthesized by the plant if all metabolic reactions in the pathogen could be used constructively. While this may seem irrelevant considering that pathogens do not help their hosts but rather exploit them, it is nevertheless an interesting theoretical exercise, because a transition from pathogenic exploitation to symbiotic mutualism can occur, and vice versa. For example, type III secretion systems that are crucial for the pathogenicity of P. syringae, are also active in the plant growth promoting Pseudomonas fluorescens SBW 2528 [50], and haustoria are established by pathogenic fungi, e. g. powdery mildew, as well as by symbiotic mycorrhiza (here called arbuscules). Some symbionts may have evolved from pathogens and adapted to their hosts by providing them with valuable chemical substances [51]. It is plausible to assume that pathogens, at least to some degree, may also be physiologically beneficial to their host’s metabolism.

The plant asymmetric gain varies considerably for the five pairs. By far the highest value is observed for Glycine max with Sclerotinia sclerotiorum. Therefore, the metabolism of the parasite S. sclerotiorum appears to have a high potential to be of use for its host. It can be speculated that in evolutionary terms this parasitism has some chance to evolve into a symbiotic relationship, because both partners can – at least at the metabolic level – in principle benefit considerably from each other. Except for one pair (Gm/Ss) where it is relatively even, the asymmetric gain is considerably higher for all pathogens relative to the plant hosts in all pairs. This is not surprising considering that we performed our calculation under the assumption that only inorganic nutrients are available. Clearly, the parasites cannot utilize this combination without presence of the plants. Interestingly, the symmetric metabolic gain and the plant asymmetric gain do not seem to be correlated with network distance. Intuitively, one would expect that the larger the network overlap (small Jaccard distance), the smaller the gain, because the networks have essentially the same metabolic capacity, and conversely, the smaller the overlap (large Jaccard distance), the larger the gain. However, no significant correlation, even slightly negative rather than the expected positive correlation, was found between the gain or asymmetric gain and the Jaccard distance (with the Pearson correlation coefficients and associated p-values of r=-0.25, p=0.68, and r=-0.10, p=0.87, respectively). For example, Glycine max and Sclerotinia sclerotiorum exhibit the largest gain amongst the five pairs (191 (symmetric), 238 (asymmetric)), whereas the network distance of 0.145 is rather moderate. By contrast, Arabidopsis thaliana and Pseudomonas syringae exhibit a rather large Jaccard distance of 0.216, but the metabolic gain is only 1 (symmetric) and 34 (asymmetric).

To assess the negative impact of pathogens on the hosts, we calculated how the presence of the pathogens impairs the production of essential biomass precursors in the host networks (see Methods). To allow for a systematic analysis, we determined the impairment not only for the specific interactions of pathogens on their natural hosts, but extended the analysis to all combinations of host-pathogen pairs. Effectively, the non-natural plant-pathogen pairs produced in silico serve as a null-model to which any specific effects of the actual plant-pathogen combinations can be contrasted. The result is depicted in Figure 4 in which the shadings of the squares indicate the impairment scores in a logarithmic scale. A white square represents a score of less than 0.01, a black square of a score close to 1.

Interestingly, only few biomass precursors show a high score while most precursors are only marginally impaired. Only for histidine, lysine, methionine and thymidine triphosphate (TTP) scores over 0.05 are observed. Histidine is strongly impaired in all investigated plant-pathogen pairs with scores between 0.25 and 0.85. Methionine and cysteine are also rather uniformly impaired in all pairs, but with considerably lower scores (between 0.05 and 0.08 for methionine and between 0.03 and 0.04 for cysteine). By contrast, the impairment of lysine and TTP is very heterogeneous throughout the investigated plant-pathogen pairs. While the production of thymidine triphosphate is strongly impaired by the pathogens X. oryzae (0.40-0.65), U. maydis (0.95-0.96) and S. sclerotiorum (0.99), it is only weakly impaired by the other two pathogens P. syringae (0.03) and M. larici-populina (0.03-0.06). This means that for U. maydis and S. sclerotiorum for almost all predicted nutrient combinations, removal of the nutrients leads to a disability of the host network to produce the essential nucleotide phosphate TTP. Interestingly, their metabolic networks exhibit a low pairwise distance (0.144) so it can be speculated that the mechanism is similar, despite the fact that S. sclerotiorum is a nectrotrophic parasite, while U. maydis is biotrophic. On the other hand, the network of M. larici-populina, the third fungal pathogen which is also biotrophic, also exhibits a low distance to S. sclerotiorum and U. maydis, but the impairment on host networks is considerably different. Similarly, lysine is considerably impaired by the pathogens U. maydis (0.18-0.21) and M. larici-populina (0.25-0.29), but only marginally by the other three pathogens (scores below 0.02).

The observation that only a subset of biomass precursors is susceptible to impairment may result from a general, non-plant-specific, vulnerability of the respective synthesis pathways. To investigate the general fragility of these synthesis pathways, we have included the network comprising all reactions from MetaCyc [52] as a hypothetical host network. The impairment of the pathogens on the MetaCyc network (marked MC) is displayed in the top row of each panel in Figure 4. Histidine, for example, is as strongly impaired in the MetaCyc network as it is in the plant networks, suggesting that the full set of reactions does not provide a higher robustness for histidine synthesis when compared to the plant-specific synthesis pathways. Lysine production is impaired in the MetaCyc network only by the two pathogens U. maydis and M. larici-populina, which also impair lysine production in the plant networks. However, the considerably lower impairment scores (0.12 and 0.09, respectively) indicate that the full set of MetaCyc reactions provides more alternative synthesis routes and thus displays an increased robustness against competition by the pathogens. An interesting pattern is observed for the impairment of TTP. The pathogens U. maydis and S. sclerotiorum, which strongly impair TTP production in plant networks, also exhibit the strongest effect on the full network, albeit with a lower impairment score (0.64 and 0.27, respectively). By contrast, X. oryzae, which also strongly impairs TTP in plant networks, has only a negligible effect on the full network (<0.01). This indicates that in the case of X. oryzae other metabolic routes exist in MetaCyc, which could circumvent the removal of the required nutrients, while this is not the case for U. maydis and S. sclerotiorum.

As a general tendency, the impairment patterns appear to be largely determined by the pathogens and rather independent of the host species. For a systematic investigation of the similarities in the impairment patterns, we quantify the overall effect of a pathogen on a host by the vector containing as elements the 28 impairment scores for the biomass precursors and calculated the pairwise Manhattan (1-norm) distances. These distances were used to perform a multi-dimensional scaling [53] to visualize similarities and differences. The resulting plot (Figure 5) highlights that pairs containing the same pathogens (same colors) are always grouped together, confirming that the impairment pattern is largely determined by the pathogen. Furthermore, it can be observed that in most cases the impairment on the full MetaCyc network (squares) is clearly distinguished from the impairment on the plant networks. The only exception is P. syringae, which displays a similar effect on the full network as on the plant-specific networks. This exception can be explained by the fact that P. syringae mainly impairs the production of histidine. However, as discussed above, histidine production is almost equally impaired in the full MetaCyc network as in the plant-specific networks.