Towards Systems Biology of Heterosis: A Hypothesis about Molecular Network Structure Applied for the Arabidopsis Metabolome
© Sandra Andorf et al. 2009
Received: 30 May 2008
Accepted: 4 August 2008
Published: 27 August 2008
We propose a network structure-based model for heterosis, and investigate it relying on metabolite profiles from Arabidopsis. A simple feed-forward two-layer network model (the Steinbuch matrix) is used in our conceptual approach. It allows for directly relating structural network properties with biological function. Interpreting heterosis as increased adaptability, our model predicts that the biological networks involved show increasing connectivity of regulatory interactions. A detailed analysis of metabolite profile data reveals that the increasing-connectivity prediction is true for graphical Gaussian models in our data from early development. This mirrors properties of observed heterotic Arabidopsis phenotypes. Furthermore, the model predicts a limit for increasing hybrid vigor with increasing heterozygosity—a known phenomenon in the literature.
"Biological function" is the core of biological research, but it is an ill-defined term. Geneticists, cellular biologists, structural biologists, biophysical chemists, and bioinformaticians all target different meanings in their respective research areas . However, as a unifying notion, biological function always refers to semantic features and, as such, is always context-dependent. A specific state of any biological molecule alone is not accomplishing any biological function . Rather, biological function resides in interactions [4–6]. The characteristics of such biological interactions, when analyzed on a genome-wide scale, are referred to as the structure of biological networks (including their dynamics). Relating structure of biological networks to biological function is therefore a major objective in biology, mirrored in recent developments such as systems biology.
A huge variety of biological networks exist; however, there are common characteristics. Biological network structure always arises as interaction of genetic determination and environmental influences, as well as internal systems dynamics. As pointed out by Somogyi and Sniegoski , interactions within specific representations of biological networks may either map directly to existing biomolecules, or may reflect rather indirect relations involving possibly many of hidden variables . Most types of biological networks can be interpreted also as regulatory networks, in the sense that they "respond" to environmental or developmental challenges by changing their state or dynamics. A frequent approach to search for important network structures at a rather global level of biological networks is statistical network modeling. It starts out by screening for significant measures from graph theory [9–11]. Distributions of such measures can then be compared between biological, technical, and random networks, as well as between different classes of organisms , regimes of environmental challenges, or developmental periods . If specific structures are discovered, their relation to a biological function of interest may be hypothesized and experimentally validated on further datasets.
Gjuvsland et al.  demonstrate how epistatic interactions within statistical genetics models can be translated into functional structures of regulatory biological networks. In our contribution, we focus on these molecular network structures and ask the following question. Which structures of biological networks could systematically lead to higher adaptability in heterozygotes, and thus to the heterosis phenomenon? For investigating this question, we choose to follow a conceptual modeling approach . Our model choice is based on a major result of statistical network modeling. Analyses of distributions of simple regulatory motifs both in prokaryotes and in eukaryotes point to similar results. The so-called multi-input motif is a significant and prominent part of regulatory biological networks . The properties of networks of this type were studied by Steinbuch already in 1961 . His studies were focusing on modeling and implementing models of associative learning. The so-called Steinbuch matrix is a two-layer feed-forward network. The information about which input vector is associated with which output vector is encoded within the pattern of presence/absence of connections between these two layers. We are going to use this Steinbuch network as a conceptual model for biological networks, and develop a hypothesis of heterosis based on biological network structure. We expect specific global structures in biological networks to be different between homozygotes and their heterozygous offspring.
To validate and further detail our network hypothesis of heterosis, we analyze partial correlation structures in experimental metabolite profile association networks from two different homozygous Arabidopsis thaliana lines and both reciprocal crosses as heterozygotes. These metabolite profiles were measured during early development of Arabidopsis, as during this time heterosis phenomena become manifest in this species . We refer again to Somogyi and Sniegoski  following their argument that not only the transcriptome but also the metabolome could be viewed as a special mapping of the extended biological regulatory network. Such a mapping would include many indirect regulatory interactions involving hidden molecular variables which are part of other levels of gene expression.
Summarizing the objectives of our study, we motivate the proposal of a network structure-based hypothesis of heterosis, and look for heterozygote-specific network structures as predicted by a Steinbuch network conceptual modeling approach. Analyses of metabolite profiles of early development in Arabidopsis thaliana and further observations of heterosis in plants will serve as to validate and further adjust our hypothesis.
Section 2 describes the experimental dataset and our preprocessing prior to statistical network analyses. In Section 3, we describe our modeling approach as well as a small simulation study. Its results motivated our choice of network statistics for global assessment of network structures described in the remaining part of this section. The first part of Section 4 reports the simulation results. In its second part, we develop our network structure-based hypothesis of heterosis and its predictions. In the last part of this section, results of experimental data analysis as motivated by our model predictions are presented. Finally, in Section 5 we discuss the main findings of our study, along with their relevance and benefits, and constraints of our approach as well as future prospects.
2. Experimental Data and Preprocessing
We investigate metabolite profiles (GC-MS data) of early development of Arabidopsis thaliana. More precisely, metabolite profiles of plants of the two homozygous lines C24 and Columbia (Col-0: depicted as Col in what follows) and the reciprocal crosses ColxC24 and C24xCol are studied. Metabolite profiles of the two homozygous genotypes C24xC24 and ColxCol and the two heterozygous genotypes C24xCol and ColxC24 were measured at 7 time points (0, 12, 24, 36, 48, 72, and 96 hours after sowing (HAS)). For each measurement, a Petri dish of seedlings was grown and fully harvested after the specific time of growing. In our balanced cross-factorial design, four replicates were assessed per genotype and time point, measured at three different measuring days, such that each genotype time point combination was measured at least once per measuring day. The raw data preparation was performed as in ; afterwards, the data were log-transformed. Overall, 210 metabolites have been measured. Eight of them contained more than 20% missing values, and were therefore excluded from further analysis.
3.1. Modeling and Simulation
Our conceptual modeling approach employs a model of association to simulate adaptability in regulatory networks. Adaptedness can be described as the ability to give a correct response (output) to an environmental or developmental challenge (input). Hence, an adaptation can be viewed as the correct association of a response to the input in question. Correspondingly, adaptability is the number of differentiated correct adaptations that a regulatory system is able to realize.
We conducted a small simulation study, employing an association matrix of size which is capable of correctly associating 4 pairs of input-output vectors. The model was trained to reproduce these predefined input-output pairs, which can be interpreted as some kind of crucial regulatory reply (regulatory step) to cope with a special environmental challenge. The study should reveal whether a partial correlation analysis of state profiles for the nodes of the network is a valid possibility to study the causal regulatory interactions in this network. 100 randomly generated input vectors and their corresponding outputs were stored as profile data and partial correlations calculated as detailed in what follows.
3.2. Network Statistics
The bases for these values are the normalized metabolite values for the seven time points from (2) for each genotype and each of the analyzed 202 metabolites. Thus, for any two metabolites of one of the four genotypes, partial correlations can be calculated based on the seven pairs of metabolite values corresponding to the seven time points. is the estimate of the partial correlation between the metabolites and . are the elements of the inverse covariance matrix which is estimated using a shrinkage estimator . The algorithm is implemented in the package GeneNet .
for all metabolites .
Here, denotes the respective heterozygous line ( ).
Distributions of were displayed and compared.
To investigate if the metabolites showing the largest values for had a specific distribution over metabolite pathways, we visualized the first thirty metabolites in a ranking of for each heterozygous line using MAPMAN . MAPMAN is a tool to display large datasets onto diagrams of metabolic pathways.
Not only global distributions of changes in partial correlations but also structural properties of partial correlation networks could be different between homozygous and heterozygous lines. In such networks, edges are significant partial correlations, computed according to Opgen-Rhein and Strimmer . P-values were corrected using the FDR correction described by Benjamini and Hochberg . Accordingly, nodes in partial correlation networks are the metabolites contributing to significant partial correlations.
The degree of such a node is defined as the number of edges it is part of. We characterized the partial correlation networks of the two homozygous and the two heterozygous lines by counting significant edges and the participating nodes, as well as calculating the mean degree values over all nodes of a network.
4.1. Simulation Results
Our small simulation study, where we recorded outputs for 100 random inputs to a association matrix reproducing 4 input-output associations, revealed that causal interactions between input and output layers lead to increased partial correlations of the respective nodes. As demonstrated in Figure 4(b), for our model, causal interactions can be deduced from observational profile data by calculating partial correlations. These properties of our conceptual model led to the development of a network structure-based model of heterosis as outlined in what follows.
4.2. Network Hypothesis of Heterosis
As suggested by Robertson and Reeve , heterozygotes are likely to possess a greater biochemical versatility by carrying a greater diversity of alleles. Heterosis would then result from a reduced sensitivity to environmental variations since there will be ways of overcoming such challenges. In other words, the heterosis phenomenon may be due to higher adaptability in heterozygotes.
Correspondingly, as illustrated in Figure 3(a), the molecular network of a heterozygous cross may contain a proportion of heterozygous loci, as for gene "b," for example. The additional alleles at this locus may lead to additional regulatory interactions in the molecular network (yellow arrows in Figure 3(a)). In our model, as shown in the simulation (see Figure 4(b)), additional causal interactions are the basis of an increasing number of associations in the repertoire of the Steinbuch network. It is known from earlier studies of system properties of the Steinbuch network that there exists a limit of associated pairs for a network of a given size . A Steinbuch network of a given size can be built to be able to differentiate between a certain number of inputs by "responding" with the (associated) belonging outputs, and not more. This is a known system property of this type of regulatory networks—but also for other types of neural networks. Moreover, if we measure an increasing amount of partial correlations within a molecular network, this might correlate with an increased amount of regulatory "challenge-response" pairs managed by this network, and hence with increased adaptability. Interpreting these properties as conceptual model for adaptation and adaptability in molecular regulatory networks leads to two predictions for the case of heterosis.
There should exist a limit for increasing hybrid vigor with increasing level of heterozygosity. Increasing the genetic distance of homozygous parental lines beyond a certain threshold should result in less hybrid vigor if these parental lines are genetically too different. When mating two similar homozygous genotypes, only few additional regulatory connections within the molecular networks can be expected. However, when mating homozygous genotypes which are genetically very different (with large genetic distance), the limit of the resulting merged molecular network structures may be exceeded—in the sense that regulatory interactions in the network of the resulting heterozygotes do not match and therefore do not lead to additional possibilities of valid regulatory answers.
Molecular interactions in regulatory networks of heterozygotes should be slightly enriched. This increased number of "challenge-response" pairs is modeled as a higher number of association pairs in our conceptual model, interpretable as increased adaptability leading to heterosis. As for the model, where we were able to measure interactions as increased partial correlations, we also expect an increase in partial correlations from homozygotes to heterozygotes for the experimentally observed dynamics of biological regulatory networks.
Regarding prediction 2, we studied our experimental dataset, the Arabidopsis metabolome of a developmental time series (see Section 4.3). From the perspective of our model, Figure 3(a) illustrates how the molecular network of heterozygotes contains additional regulatory possibilities. In the association network model, these correspond to additional connections (interactions) between input and output layers, enabling the network to add additional associations to its repertoire. These additional associations (input-output pairs) represent a grown repertoire of adaptations, or increased adaptability, enabling increased hybrid vigor. The objective of our experimental data analyses was to investigate if such increase in molecular interactions would be measurable as increase in partial correlations as a global network property for the metabolite profiles recorded during Arabidopsis development.
4.3. Analysis of Experimental Data
Our experimental data were metabolite profiles from development of Arabidopsis thaliana (see Figure 2). To test our hypothesis that heterosis comes as increasing adaptability and should result in increasing connectivity of molecular networks, we had first conducted a small simulation study (see Section 4.1). Its findings provide the basis for our investigation of partial correlation structures of the metabolomes of heterozygous and homozygous genotypes for the experimental data, as we want to test a hypothesis about increased regulatory possibilities in heterozygotes and the belonging structures of molecular profiles. Hence, we analyzed partial correlations according to Opgen-Rhein and Strimmer  for our experimental dataset.
Significant partial correlations (significance level: ).
No. of significant edges
Hence, results of Figures 6 and 9 point towards the same tendency, supporting the "increasing-connectivity" prediction of our network hypothesis of heterosis. This tendency is strengthened as most of the 30 metabolites that show the largest differences between the heterozygotes and the midparent value also have significant edges. In more detail, for genotype C24xCol, 25 of the top 30 metabolites and, for genotype ColxC24, 27 of the top 30 metabolites have significant edges. Total numbers of nodes with significant edges are 45 and 40, respectively (see Table 1). On average, for either heterozygous genotype, 86.7% of the top 30 metabolites show significant edges.
We have developed a network structure-based hypothesis of heterosis. It is a systems biological approach to relate biological function to molecular network structure. Our hypothesis results in the following predictions. First, system properties of our network modeling approach suggest the existence of an upper limit for the heterosis effect when genetic distance of crossed homozygous parental lines becomes too large. Second, molecular networks of heterozygotes should contain additional interactions compared to those of their homozygous parents. These additional interactions should lead to increased partial correlations in molecular networks of heterozygotes. For the first prediction, we found support in the literature suggesting an upper limit for the heterosis effect. However, as we do not have sufficient additional own experimental evidence, no final conclusion can be drawn for this case. Further investigations seem promising and necessary. Regarding the prediction of increased connectivity of molecular networks in heterozygotes, for our own experimental metabolome dataset of Arabidopsis, such increased connectivity was observable for both heterozygous crosses. It was this phase of early Arabidopsis development in which the heterosis effect is established. The predicted pattern is visible for the majority of metabolites. However, also for the second part of our network hypothesis of heterosis, we call for additional experimental evidence, preferably on additional levels of molecular regulatory networks, such as proteomics or transcriptome data. To summarize, we present a conceptual frame for explaining the heterosis phenomenon from a molecular network perspective together with two hypotheses and their predictions, for which we were able to find the first supporting evidence from the literature and own experimental data.
We are convinced that research towards understanding the biological phenomenon of heterosis can particularly gain from a systems biological approach focused on interactions of molecular building blocks and global structures of molecular biological networks. Towards elucidating the genetic basis of heterosis, Melchinger et al.  have already shown that, taking a statistical modeling approach, epistatic interactions of individual loci with the entire genetical background constitute a major component of genetic variation important to explain heterosis. However, the mapping of interaction terms in models of quantitative genetics to structures in molecular regulatory networks is nontrivial . Our approach to investigate global network structures in molecular interaction networks for this reason is to be taken as complementary to the quantitative genetics view.
Meyer et al.  report for Arabidopsis thaliana development that it is the early phase of development (till one week of seedlings' growth) during which the heterosis phenotype for biomass is established. In later phases of the plant life, relative differences between heterozygotes and homozygotes are not further growing. The first observation coincides with our results. We observe increased connectivity in partial correlation networks during this period of development. It would be interesting to see—this is planned as future experimental study—if during the later phase, when according to  biomass heterosis is visible but no longer increasing, there is no indication of increased connectivity in the metabolome any longer.
The majority of metabolites investigated showed an increase in interaction connectivities. We tried to find common functionalities for the top 30 metabolites with most obvious changes. However, we were not able to detect evidence towards an accumulation of such metabolites within certain pathways or modules (MAPMAN categories). We hypothesize that it may be these metabolites that during the early phase of Arabidopsis development are mainly involved in regulatory interactions—to enable adaptation to the climatic chamber during the first contact with this environment.
Only part of the observed changes in partial correlations between heterozygous lines and the midparent value of both homozygotes can be based upon significant partial correlations (compare Figures 6 and 9). However, the same tendency is apparent for the global view as well as for the restriction to significant correlations. It is the sparsely designed experimental data that does not allow for a more precise analysis. Seven time points are clearly the lower limit of correlation analyses involving around two hundred metabolite species. We look forward to more generously designed experiments for testing our network structure-based hypotheses for heterosis.
Our modeling approach is conceptual as advocated for by, for example, Wissel  and Shubik . It builds upon the understanding of the heterosis phenomenon as increased adaptability. This understanding has its roots already at the beginning of the 20th century in maize genetics  and since then has been expressed also within the context of hybrid vigor observed for other plant species as well as model animals (see, e.g., ). We make use of a model for adaptability which was originally designed to model associative memory (the Steinbuch matrix) . Within our model, being adapted means to respond in a correct way when confronted with a certain environmental or developmental stimulus, while adaptability means the potential to respond to a number of different stimuli with differentiated correct responses. The simplicity of this conceptual modeling implies rather general predictions. In our case, these are the limit-of-heterosis increase prediction and the increasing-connectivity prediction. These are predicted for a huge class of interaction networks, independent of molecular species. Motif analyses in different molecular interaction networks as well as within organisms of different kingdoms (prokaryotes and eukaryotes) have shown that certain motifs are always present. The "multi-input motif" is a prominent example. Here, we refer to the work by Milo et al.  and Lee et al. . The multi-input motif has the same structure as our association network model, which was first proposed in 1961 by Steinbuch . Furthermore, molecular interactions are often modeled based on a sigmoidal relationship as approximated by the boolean kind of interaction in the Steinbuch model (discussed in ).
A central assumption underlying motif analyses as well as our modeling approach for this work is that neglecting the diversity of different kinds of molecular species that interact within real molecular networks does not cause harm at the rather general level of conclusions of our conceptual investigations. It is clear that natural molecular networks cannot be reduced to a very simplistic model in all their structural and dynamical properties. However, we chose to follow Shubik's call for the most parsimonious modeling approach . Also, heterosis is a very general biological phenomenon together with its counterpart inbreeding depression. Both phenomena are occurring over a broad variety of sexually reproducing organisms. For this reason, approaches towards understanding the systems biological foundations of these phenomena should be independent of all organism-specific parameters, in other words as simple as possible.
Choosing the metabolome level, as in our study, is just one possibility. With , we argue that the extended regulatory network of an organism can be mapped to any of its levels of gene expression ("omics" levels). However, the modeler has to be aware of all possible hidden variables constituting each of the investigated interactions. These hidden variables are representations of the molecules from the "omics" levels which were not modeled. In our case, for example, regulatory interactions between metabolites have no direct correspondence to metabolic pathways. Moreover, as is true for gene expression studies for the case of transcription factors, also in metabolomics it is not at all possible to assess all molecules, but only a small fraction. The measurable fraction may or may not be a biased sample from the entire metabolome, and for this reason inferring network structures from such a sample has always to be taken with care (for an example concerning network statistics in protein interaction networks, see ). Also, we are aware of the problem of cell-type heterogeneity in our samples which are basically whole embryo/plant homogenates. Measured profiles in our case represent metabolite levels of the major cell type. In addition, it is important to take into account the fact that those 202 metabolites in our investigation are just around 10% (possibly less) of the metabolites that are supposed to be present in Arabidopsis thaliana . Thus, our network structure-based hypothesis of heterosis was validated only for the core carbon metabolism. These small molecules (e.g., sugars, amino acids, and carbon acids) act mostly within energy metabolism and as precursors for building the larger biomolecules, proteins, and nucleic and fatty acids. These metabolites represent what is currently measurable with the GC-MS metabolite profiling experiments.
For future investigations of molecular network structures with respect to the heterosis phenomenon, it will be an interesting challenge to extend the time series design of the current study in several aspects. To enable a more general conclusion regarding the two predictions from our network hypothesis of heterosis, it would be worth comparing several different homozygous lines and their reciprocal offspring. Also, genetically very different lines should be included to approach a direct test of the limit-of-heterosis increase prediction. Moreover, time points should be set more dense (e.g., as 10-hour intervals) and over a longer time scale (e.g., at least along the first four weeks of Arabidopsis thaliana development). Such a design would enable a higher precision for both estimating partial correlation structures as well as assessing a possible change of such structures during later phases of development, for which according to Meyer et al.  no additional heterosis effects are arising. Furthermore, studies are already planned to analyze transcript data measured under the same conditions as our metabolome dataset. This would enable us to show, first, how two levels of the extended regulatory network act together taking an integrative bioinformatics approach (see, e.g., ). Second, it would be possible to test the increasing-connectivity prediction of heterosis also for the level of the transcriptome.
Regarding alternative approaches to measure differential network structures in molecular networks of homozygotes and heterozygotes, there exist a number of possible choices. An alternative type of networks used for inference of biochemical interaction networks is, for example, the so-called relevance network. Butte et al.  base their method on a pairwise Pearson correlation of all features. A serious limitation of relevance networks is that they contain many indirect correlations because they cannot distinguish between direct and indirect interactions. For our kind of observational data, Werhli et al. have shown that it is preferable to use association networks to infer regulatory interactions . For this reason, we decided to analyze partial correlations as proposed by Opgen-Rhein and Strimmer . We also favored the regularized inference of the covariance matrix they proposed, which is applicable for data with a small sample size and a comparatively large number of variables, as in our metabolome dataset. Our simulation study was able to demonstrate that, when observing a number of partial correlations from the Steinbuch model, these could be used to identify the nodes of input and output layers which were connected in the regulatory architecture of the network model to reproduce four predefined input-output patterns. Hence, for our conceptual model, regulatory interactions could be deduced from partial correlations. A possibly promising way to extend our analyses could be oriented along the lines of the work by Saul and Filkov  who proposed to use the so-called exponential random graph models. They demonstrate their utility in modeling the architecture of biological networks as a function of a number of different measures of local network structure, not only a single measure as in our case. The flexibility, in terms of the number of available local feature choices, and scalability possibly make this approach a suitable alternative for statistical modeling of biological networks.
To summarize, in our work we followed the call of Barabási and Oltvai  who conclude their review on network biology by stating that structure, topology, network usage, robustness, and function are deeply interlinked, forcing us to complement the "local" molecule-based research with integrated approaches that address the properties of regulatory networks at a systems biological level. In our study, we have done so by proposing a network structure-based model of heterosis and investigating its predictions for an experimental omics dataset. Heterotic phenotypes of Arabidopsis are mirrored as increased connectivity in metabolome partial correlation networks. A limit of hybrid vigor increase for increasing genetic distance of crossed parents is also correctly predicted. These results hold for the measured part of the metabolome, mostly central carbon metabolism.
Our conclusions cannot be more than an illustrative example of how a hypothesis can be built about a possible relation of biological network structure to biological function (in our case, the heterosis phenomenon). We advertise our approach as a way of investigating heterosis complementary to the quantitative genetics approach, and look forward to future unifying approaches to these two fields.
This work was supported by the German Research Council (DFG) under Grants RE 1654/2-1 and SE 611/3-1.
- Lambert D, Hughes T: Misery of functionalism. Biological function: a misleading concept. Rivista di Biologia 1984, 77(4):477-502.Google Scholar
- Ge H, Walhout AJM, Vidal M: Integrating 'omic' information: a bridge between genomics and systems biology. Trends in Genetics 2003, 19(10):551-560. 10.1016/j.tig.2003.08.009View ArticleGoogle Scholar
- Bohm D: Wholeness and the Implicate Order. Routledge, London, UK; 1980.Google Scholar
- Strogatz SH: Exploring complex networks. Nature 2001, 410(6825):268-276. 10.1038/35065725View ArticleGoogle Scholar
- Somogyi R, Sniegoski CA: Modeling the complexity of genetic networks: understanding multigenic and pleiotropic regulation. Complexity 1996, 1: 45-63.View ArticleMathSciNetGoogle Scholar
- Noble D: Modeling the heart—from genes to cells to the whole organ. Science 2002, 295(5560):1678-1682. 10.1126/science.1069881View ArticleGoogle Scholar
- Perrot M, Guieysse-Peugeot A-L, Massoni A, et al.: Yeast proteome map (update 2006). Proteomics 2007, 7(7):1117-1120. 10.1002/pmic.200600952View ArticleGoogle Scholar
- Tresch A, Markowetz F: Structure learning in nested effects models. Statistical Applications in Genetics and Molecular Biology 2008., 7(1): article 9Google Scholar
- Barabási AL, Albert R: Emergence of scaling in random networks. Science 1999, 286(5439):509-512. 10.1126/science.286.5439.509View ArticleMathSciNetGoogle Scholar
- Milo R, Shen-Orr S, Itzkovitz S, Kashtan N, Chklovskii D, Alon U: Network motifs: simple building blocks of complex networks. Science 2002, 298(5594):824-827. 10.1126/science.298.5594.824View ArticleGoogle Scholar
- Saul ZM, Filkov V: Exploring biological network structure using exponential random graph models. Bioinformatics 2007, 23(19):2604-2611. 10.1093/bioinformatics/btm370View ArticleGoogle Scholar
- Lee TI, Rinaldi NJ, Robert F, et al.: Transcriptional regulatory networks in Saccharomyces cerevisiae . Science 2002, 298(5594):799-804. 10.1126/science.1075090View ArticleGoogle Scholar
- Matthäus F, Salazar C, Ebenhöh O: Biosynthetic potentials of metabolites and their hierarchical organization. PLoS Computational Biology 2008, 4(4):1-13. e1000049View ArticleGoogle Scholar
- Shull GH: The composition of a field of maize. American Breeding Association Report 1908, 4: 296-301.Google Scholar
- Robertson FW, Reeve ECR: Heterozygosity, environmental variation and heterosis. Nature 1952, 170(4320):286.View ArticleGoogle Scholar
- Birchler JA, Auger DL, Riddle NC: In search of the molecular basis of heterosis. The Plant Cell 2003, 15(10):2236-2239. 10.1105/tpc.151030View ArticleGoogle Scholar
- Crow JF: Dominance and overdominance. In Heterosis. Iowa State College Press, Ames, Iowa, USA; 1952:282-297.Google Scholar
- Tsaftaris SA: Molecular aspects of heterosis in plants. Physiologia Plantarum 1995, 94(2):362-370. 10.1111/j.1399-3054.1995.tb05324.xView ArticleGoogle Scholar
- Gjuvsland AB, Hayes BJ, Omholt SW, Carlborg Ö: Statistical epistasis is a generic feature of gene regulatory networks. Genetics 2007, 175(1):411-420.View ArticleGoogle Scholar
- Wissel C: Aims and limits of ecological modelling exemplified by island theory. Ecological Modelling 1992, 63(1–4):1-12.View ArticleGoogle Scholar
- Shubik M: Simulations, models and simplicity. Complexity 1996, 2(1):60. 10.1002/(SICI)1099-0526(199609/10)2:1<60::AID-CPLX12>3.0.CO;2-VView ArticleGoogle Scholar
- Shen-Orr SS, Milo R, Mangan S, Alon U: Network motifs in the transcriptional regulation network of Escherichia coli . Nature Genetics 2002, 31(1):64-68. 10.1038/ng881View ArticleGoogle Scholar
- Steinbuch K: Die Lernmatrix. Kybernetik 1961, 1(1):36-45. 10.1007/BF00293853View ArticleMATHGoogle Scholar
- Meyer RC, Törjék O, Becher M, Altmann T: Heterosis of biomass production in arabidopsis. Establishment during early development. Plant Physiology 2004, 134(4):1813-1823. 10.1104/pp.103.033001View ArticleGoogle Scholar
- Lisec J, Schauer N, Kopka J, Willmitzer L, Fernie AR: Gas chromatography mass spectrometry-based metabolite profiling in plants. Nature Protocols 2006, 1(1):387-396. 10.1038/nprot.2006.59View ArticleGoogle Scholar
- Opgen-Rhein R, Strimmer K: From correlation to causation networks: a simple approximate learning algorithm and its application to high-dimensional plant gene expression data. BMC Systems Biology 2007, 1: 1-10. article 37 10.1186/1752-0509-1-1View ArticleGoogle Scholar
- Werhli AV, Grzegorczyk M, Husmeier D: Comparative evaluation of reverse engineering gene regulatory networks with relevance networks, graphical Gaussian models and Bayesian networks. Bioinformatics 2006, 22(20):2523-2531. 10.1093/bioinformatics/btl391View ArticleGoogle Scholar
- Schäfer J, Strimmer K: A shrinkage approach to large-scale covariance matrix estimation and implications for functional genomics. Statistical Applications in Genetics and Molecular Biology 2005, 4(1):1-30. article 32View ArticleMathSciNetGoogle Scholar
- Opgen-Rhein R, Schäfer J, Strimmer K: GeneNet: Modeling and Inferring Gene Networks. R package version 1.2.0., 2007
- Thimm O, Bläsing O, Gibon Y, et al.: MAPMAN: a user-driven tool to display genomics data sets onto diagrams of metabolic pathways and other biological processes. The Plant Journal 2004, 37(6):914-939. 10.1111/j.1365-313X.2004.02016.xView ArticleGoogle Scholar
- Benjamini Y, Hochberg Y: Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B 1995, 57(1):289-300.MathSciNetMATHGoogle Scholar
- Nadal J-P: Associative memory: on the (puzzling) sparse coding limit. Journal of Physics A 1991, 24(5):1093-1101. 10.1088/0305-4470/24/5/023View ArticleMATHGoogle Scholar
- Moll RH, Lonnquist JH, Fortuno JV, Johnson EC: The relationship of heterosis and genetic divergence in maize. Genetics 1965, 52(1):139-144.Google Scholar
- Link W, Schill B, Barbera AC, et al.: Comparison of intra- and inter-pool crosses in faba beans ( Vicia faba L.). I. Hybrid performance and heterosis in Mediterranean and German environments. Plant Breeding 1996, 115(5):352-360. 10.1111/j.1439-0523.1996.tb00933.xView ArticleGoogle Scholar
- Melchinger AE: Genetic diversity and heterosis. In The Genetics and Exploitation of Heterosis in Crops. ASA-CSSA, Madison, Wis, USA; 1999:99-118.Google Scholar
- Falconer DS, Mackay TFC: Introduction to Quantitative Genetics. Longman, Essex, UK; 1996.Google Scholar
- Melchinger AE, Utz HF, Piepho H-P, Zeng Z-B, Schön CC: The role of epistasis in the manifestation of heterosis: a systems-oriented approach. Genetics 2007, 177(3):1815-1825. 10.1534/genetics.107.077537View ArticleGoogle Scholar
- Gibson G: Epistasis and pleiotropy as natural properties of transcriptional regulation. Theoretical Population Biology 1996, 49(1):58-89. 10.1006/tpbi.1996.0003View ArticleMathSciNetMATHGoogle Scholar
- Solomon KF, Labuschagne MT, Viljoen CD: Estimates of heterosis and association of genetic distance with heterosis in durum wheat under different moisture regimes. The Journal of Agricultural Science 2007, 145(3):239-248. 10.1017/S0021859606006551View ArticleGoogle Scholar
- Harrison GA: Heterosis and adaptability in the heat tolerance of mice. Genetics 1962, 47(4):427-434.Google Scholar
- Kauffman SA: The Origins of Order. Oxford University Press, Oxford, UK; 1993.Google Scholar
- de Silva E, Thorne T, Ingram P, et al.: The effects of incomplete protein interaction data on structural and evolutionary inferences. BMC Biology 2006, 4: 1-13. article 39 10.1186/1741-7007-4-1View ArticleGoogle Scholar
- Cui Q, Lewis IA, Hegeman AD, et al.: Metabolite identification via the Madison Metabolomics Consortium Database. Nature Biotechnology 2008, 26(2):162-164. 10.1038/nbt0208-162View ArticleGoogle Scholar
- Steinfath M, Repsilber D, Scholz M, Walther D, Selbig J: Integrated data analysis for genome-wide research. In Plant Systems Biology. Birkhäuser, Basel, Switzerland; 2006:309-329.Google Scholar
- Butte AJ, Tamayo P, Slonim D, Golub TR, Kohane IS: Discovering functional relationships between RNA expression and chemotherapeutic susceptibility using relevance networks. Proceedings of the National Academy of Sciences of the United States of America 2000, 97(22):12182-12186. 10.1073/pnas.220392197View ArticleGoogle Scholar
- Barabási AL, Oltvai ZN: Network biology: understanding the cell's functional organization. Nature Reviews Genetics 2004, 5(2):101-113. 10.1038/nrg1272View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.