Web-Queryable Large-Scale Data Sets for Hypothesis Generation in Plant Biology

INTRODUCTION

The study of plant biology, as with all areas of biology, has undergone dramatic changes in the past decade. The development of technologically advanced, high-throughput methods for querying the expression levels of thousands of genes at once, for detecting interactions between proteins in a plant's proteome, or for simultaneously measuring the amounts of many metabolites has permitted unprecedented insight into many aspects of plant biology. Thousands of data sets encompassing millions of measurements have been generated, and importantly, most of these are freely available for use by any plant biologist worldwide to examine in the context of his or her biological question. While such large scale data sets may not provide complete understanding of a particular question, they are often an excellent starting point for planning experiments or generating hypotheses in silico or helping to make sense of one's own high-throughput data sets. These hypotheses can then be readily tested in the laboratory with the amazing variety of genetic resources and molecular techniques that have also been developed in the past 10 years.

This review provides an overview of the breadth and depth of data sets that are currently available, especially for, but not limited to, the model plant Arabidopsis thaliana. Many of these data sets were generated by researchers funded through the National Science Foundation Arabidopsis 2010 project in the U.S., the stated goal of which was to identify the functions of 25,000 genes in Arabidopsis by 2010 (Chory et al., 2000), and by the AtGenExpress Consortium, an international effort to uncover the Arabidopsis transcriptome. In this review, we emphasize Web-based tools that have integrated data from several sources. While many individual researchers have set up websites for their own data sets, resources that compare diverse data sets are often of more utility to a wider biological research audience. We describe well-developed sequence databases, focusing on transcriptome data sets, which are the most comprehensive of all of the large-scale data types, and discuss tools for querying these both in a directed manner and correlatively, using data mining tools for generating hypotheses or narrowing down search space. We also discuss databases of epigenetic modifications and small RNAs and survey metabolomic and proteomic resources. Tools for integrating disparate data types to improve function prediction are key to leveraging even more knowledge from these data sets, and two such tools will be reviewed. We conclude with some perspectives on what the future will bring in terms of queryable browsers for further understanding the plant as a collection of cellular systems and processes and of plant varieties at an ecophysiological level. Throughout this review, we provide bioexamples of how such large scale data sets have been used to expand our understanding of the processes described above, often at the cost of only a click of the mouse. An overview of the use of these tools and data sets for plant biology is given in Figure 1 , and programs and websites discussed in this review are listed in Table 1 .

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

How Can Queryable Browsers Be Used to Address Biological Questions?

Queryable browsers are represented in colored boxes. Left panel: How queryable browsers can be used to elucidate the function of a gene of interest. Right panel: How queryable browsers can be used to elucidate the molecular network within which a gene of interest participates.

GENE EXPRESSION DATABASES AND BROWSERS

Coming back to one's gene or genes of interest, a suitable next question is, what is the expression pattern of my gene of interest Expression patterns can then be used to guide further biological experiments. Additional questions might include: within my gene's family, are family members uniquely expressed in certain tissues or is one uniquely upregulated by a specific abiotic or biotic stress suggesting subfunctionalization or neofunctionalization Are there other genes, not currently known to be involved in my gene's given biological process, that exhibit similar patterns of expression Alternately, if one is not aware of a given gene's biological function, what are the functions of other genes that are similarly expressed with the gene of interest? Moving away from a single gene-centered approach, one might also ask what is the full set of transcriptional programs that occur in my tissue of interest or response condition?

Originally, genome-wide expression measurements were limited to the use of cDNA or EST resources. Seki et al. (2004) used microarrays containing RIKEN Arabidopsis full-length cDNA sequences to identify many novel abiotic stress-induced genes. Of course, without the availability of a genome and cDNA sequences, the development of the widely used short oligonucleotide microarrays for measuring the transcriptome of Arabidopsis would not have been possible. Affymetrix's 8K At GeneChip (Zhu and Wang, 2000), subsequent 22K ATH1 GeneChip (Redman et al., 2004), and Arabidopsis Tiling 1.0R Array (Laubinger et al., 2008) all have been used to examine the transcriptomes of bulk tissues and specific cell types, both within the framework of the international AtGenExpress project and by individual researchers. Data sets associated with the AtGenExpress project profile a wide variety of developmental stages, tissues, cell types, hormone responses, and biotic and abiotic stresses (Schmid et al., 2005; Kilian et al., 2007; Goda et al., 2008). These extensive resources can be mined to generate hypotheses. Mining of such data can also result in the elucidation of putative transcriptional modules by identifying genes coexpressed with a gene of interest, in the inference of cell type-, tissue-, or context-specific expression of genes within large, seemingly redundant families and in the identification of genes potentially acting within complexes. More than 4400 data sets generated with the Affymetrix ATH1 platform have been deposited to GEO (the National Center for Biotechnology Information [NCBI] Gene Expression Omnibus) at www.ncbi.nlm.nih.gov/geo/ (Edgar et al., 2002), and these may be downloaded for further analysis by independent researchers using the open source BioConductor suite (Gentleman et al., 2004). Two Web-queryable databases, which have incorporated more than half of the these expression data sets and provide many useful tools, are commonly used within the Arabidopsis community due to their user-friendly interfaces and data mining capabilities: the Bio-Array Resource for Arabidopsis Functional Genomics (the BAR; BAR.utoronto.ca) and Genevestigator (www.genevestigator.com) (Zimmermann et al., 2004, 2005; Toufighi et al., 2005; Grennan, 2006; Geisler-Lee et al., 2007; Winter et al., 2007; Hruz et al., 2008; Wilkins et al., 2008). The BAR and Genevestigator provide many tools for analysis of Arabidopsis microarray expression data and for expression data from other species. The BAR additionally allows investigation of mouse, poplar, and Medicago truncatula expression data, while Genevestigator allows investigation of human, mouse, rat, barley (Hordeum vulgare), rice, and soybean (Glycine max). We will briefly describe the tools available on both these websites, highlighting unique aspects of each.

The BAR

Analysis tools at the BAR include the Expression Angler, Expression Browser, the electronic Fluorescent Pictograph (eFP) Browser, and Promomer tools. The Expression Browser tool takes as its input large lists of genes and queries expression across user-selected expression data sets, thus allowing gene expression levels to be determined during development or in response to stresses. These data can be displayed in plain text or hierarchically clustered and visualized (Toufighi et al., 2005). Of particular utility is the user's ability to output either absolute expression levels in various treatment and control samples or the ratio of response level in the treatment relative to the level in the control. Gene expression can be visualized using the eFP Browser tool. Here, expression of one or two genes is visualized in stylized pictographs of experimental samples used to generate the data sets, essentially allowing a digital in situ of gene expression (Winter et al., 2007; Figure 2 ). Different filters can be selected to visualize expression in absolute terms, while a stimulus response can be visualized in the relative mode. The ability to monitor expression of two genes in the compare mode at high spatial resolution is useful when inferring regulatory relationships between genes of interest. Subcellular localization of a gene product can also be visualized using the Cell eFP Browser, whereby a confidence score for the localization of a gene product within each distinct subcellular compartment or region is calculated and displayed as a color scale (Winter et al., 2007). This tool will also be discussed in the proteomics section. Expression Angler is of great use when a researcher wants to identify genes that are similarly expressed with his or her gene of interest. These similarly expressed genes may be involved in the same biological process of the query gene or found within the same transcriptional regulatory module under the guilt-by-association paradigm. Taking an individual gene as bait, the user can set a Pearson correlation coefficient threshold to identify genes closely correlated or anticorrelated with that gene's expression pattern (Toufighi et al., 2005). An additional tool at the BAR is Promomer, which can identify statistically overrepresented cis-elements within the promoter region of a single gene or a list of genes, perhaps obtained from Expression Angler or Expression Browser (Toufighi et al., 2005).

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Exploring Arabidopsis Gene Expression Data with the eFP Browser (Winter et al., 2007).

Expression data for any one of ∼24,000 genes are “painted” onto a pictographic representation of the samples that were used to generate the RNA for expression profiling. In this view, gene expression data are from the Schmid et al. (2005) Developmental Atlas and from the Nambara lab. Here, the expression level of ABI3 (At3g24650) is seen to be highest toward the later stages of seed development, denoted by strong red coloration in the seed pictographs.

COEXPRESSION TOOLS

Often, genes that are coexpressed with one's gene of interest can provide an avenue for further exploration, particularly in terms of association with a particular biological process. In addition to the aforementioned coexpression tools in Genevestigator and at the BAR, Aoki et al. (2007) have reviewed several other prominent coexpression tools: ACT, ATTEDII, and AthCoR{at}CSB.DB. CressExpress (Srinivasasainagendra et al., 2008) and GeneCAT (Mutwil et al., 2008) are more recent tools that are also useful for identifying coexpressed genes. Genes of unknown function within a list of genes that are highly coexpressed with a gene of interest may also be involved in the biological process of the query gene. There are several recent examples of coexpression being used as a primary screen to identify novel genes associated with a given biological process. Koo et al. (2006) performed a coexpression screen with genes that are coexpressed with known JA biosynthetic components to identify a key step in the jasmonic acid biosynthetic pathway in Arabidopsis. Hirai et al. (2007) pinpointed the transcription factors MYB28 and 29 as regulators of glucosinolate synthesis by combining coexpression across publicly available expression data sets, along with transcriptional analyses of sulfur-starved Arabidopsis plants. d'Erfurth et al. (2008) used Expression Angler to search for genes coexpressed with known meiotic genes and then phenotyped T-DNA mutants of 138 candidate genes. Chromosome spreads in two independent mutant alleles of At1g34355 (At PS1) revealed that these plants were polyploid, indicating a role in meiosis. Two other genes with meiotic function were identified in the same screen. As a final example, three new subunits of NAD(P)H dehydrogenase were similarly identified using coexpression and reverse genetics (Takabayashi et al., 2009).

SMALL RNA DATABASES

The importance of small RNAs in controlling many aspects of plant growth and development is one of the most exciting discoveries in plant biology in the past decade (Johnson and Sundaresan, 2007). Their role in such processes is certain to be revealed to be even more far reaching. For instance, it was recently shown that there is widespread inhibition of translation by miRNAs and small interfering RNAs (siRNAs, Brodersen et al., 2008), in addition to their more familiar roles in gene silencing and natural antisense. Several research groups have aimed to document all the small RNAs in Arabidopsis (Llave et al., 2002; Xie et al., 2005; Axtell et al., 2006; Rajagopalan et al., 2006; Fahlgren et al., 2007; Howell et al., 2007; Kasschau et al., 2007; Zilberman et al., 2007; Lister et al., 2008), which in turn have been collated into the Arabidopsis Small RNA Project (ASRP) Genome Browser at asrp.cgrb.oregonstate.edu (Gustafson et al., 2005). With this Genome View of the ASRP resource, it is possible to see if one's gene of interest is being targeted by a specific small RNA or contains elements that encode small RNAs that are being identified by the ASRP. Small RNAs identified from floral buds and immature flowers using deep sequencing technology can also be visualized in the UCSC Arabidopsis Genome Browser (Lister et al., 2008). For species other than Arabidopsis, the Cereal Small RNA Database (sundarlab.ucdavis.edu/smrnas/) contains large-scale data sets of maize and rice smRNA sequences generated by high-throughput pyrosequencing and have been mapped to the rice genome and available maize genome sequence (Johnson et al., 2007).

EPIGENETIC MODIFICATIONS

Transcription factor–mediated regulation of gene expression is only one component that determines the final level of gene expression. The marking of genes by the methylation of cytosines or by the methylation/acetylation of histones of the encompassing chromatin also can dramatically alter their level of expression. In the case of the FLOWERING LOCUS C (FLC) gene in Arabidopsis, dimethylation of Lys residues 9 and 27 on histone H3 of regions of the FLC locus serves to generate a memory of winter so that flowering does not occur until after winter is over (Bastow et al., 2004). Again, ingenious technologies, including chromatin immunoprecipitation (ChIP)/whole-genome tiling arrays and shotgun bisulphite sequencing, are allowing unprecedented insight into the epigenome of this species. Querying such data can allow researchers to determine the full complement of regulatory mechanisms that determine the expression of their gene of interest. For example, Zhang et al. (2007) performed ChIP/Chip using whole genome tiling arrays to examine the H3K27me3 patterns in Arabidopsis. Prior to this study, only seven genes had been shown to be H3K27me3 methylated, namely, FLC, AGAMOUS, MEDEA, SHOOT MERISTEMLESS, PHERES1, FUSCA3, and AGAMOUS-LIKE19 (Zhang et al., 2007). Clearly, the above mentioned genes are developmentally important, and mutants unable to H3K27me3 methylate have severe developmental phenotypes. This whole-genome histone methylation study found that up to 4400 genes may be regulated by histone methylation. It is possible to search for the methylation pattern of one's gene of interest using the UCSC Genome Browser at epigenomics.mcdb.ucla.edu/H3K27m3/. In a similar manner, Zhang et al. (2006) used a whole-genome array approach to assay cytosine methylation status and found that genes that are cytosine methylated in their promoters are typically expressed in a tissue-specific manner, while those that are body methylated are expressed at higher levels. These cytosine methylation marks can also be visualized in the TAIR Genome Browser. In another recent study, Cokus et al. (2008) perfected a method called BS-seq, combining the bisulphite treatment of genome DNA with Illumina short-read sequencing technology to generate a breathtaking base pair resolution map of cytosine methylation. These data have also been loaded into the UCSC Genome Browser at epigenomics.mcdb.ucla.edu/BS-Seq/. While it is certain to emerge that the epigenome is dynamic, these initial snapshots can provide insight into a researcher's genes of interest with respect to potential additional regulatory mechanisms.

PROTEOMICS

While expression data can tell a researcher that a given gene is expressed (transcribed) under certain conditions or in certain tissues, whether or not the transcript is translated into a protein is another matter. Additionally, questions such as where the gene product might be localized within the cell, if there are any posttranslational modifications (e.g., phosphorylation), or if it interacts with other proteins, are important to answer to understand a given protein's function and activity. Arabidopsis proteomic data sets can be subdivided broadly into those attempting to quantify and document the proteome in different tissues and growth conditions, those that delimit subcellular localization, and those that tabulate protein–protein interactions.

A novel proteomic data set generated by linear trap quadrupole ion-trap mass spectrometry, which profiled protein presence in six organs and identified proteins for nearly 50% of annotated Arabidopsis gene models, is currently represented as a track on the Genome Browser at TAIR, and in the PRIDE BioMart (www.ebi.ac.uk/pride/prideMart.do) and is available in the queryable AtProteome server (fgcz-atproteome.unizh.ch). Many of these proteins were used to identify presumed organ-specific biomarkers based on approximate abundance values across different organs (Baerenfaller et al., 2008). Interestingly, some of these biomarkers were identified in a recent proteomic analysis of guard cells, demonstrating how important cell type resolution is in the generation of large-scale data sets (Zhao et al., 2008). Another large-scale proteomic data set that was acquired using two-dimensional liquid chromatographic fractionation followed by linear trap quadrupole ion-trap mass spectrometry on peptides from four different organs was published recently (Castellana et al., 2008). The authors also used TiO₂ to enrich for phosphopeptides, thus expanding our current data set with a sampling of the phosphoproteome. Interestingly, both of these approaches (Baerenfaller et al., 2008; Castellana et al., 2008) identified novel, previously unannotated proteins, enabling refinement of existing gene models, although neither obtained full proteome coverage. The Castellana et al. (2008) data set is deposited in the Tranche database (tranche.proteomecommons.org). Additionally, more than 6000 phosphopeptides from 10 published Arabidopsis studies are available from the PhosPhAt database at phosphat.mpimp-golm.mpg.de (Heazlewood et al., 2008). Including these data in TAIR would be of great use to the community. Few high-quality, quantitative proteomic data sets have been deposited in publicly available databases, with only one isotopic labeling mass spectrometry experiment deposited in 2005 to GEO (www.ncbi.nlm.nih.gov/projects/geo/) for Arabidopsis and two data sets at the Proteomics database (proteomics.Arabidopsis.info) set up by the Nottingham Arabidopsis Stock Centre (NASC). Several 2D gel data sets for a handful of developmental stages are available through the German federal government funded GABI (Genomanalyse im Biologischen System Pflanze) Project Primary Database website, www.gabipd.org (Riano-Pachon et al., 2009). This is in contrast with the thousands of gene expression data sets available for Arabidopsis. Jorrín et al. (2007) and Thelen and Peck (2007) both give good overviews of types of data sets from a variety of methodologies currently being applied to a number of different plant species and perhaps an idea of why, in spite of the many data sets generated, there isn't the equivalent of a Genevestigator or BAR eFP Browser tool available for them. Several reasons exist for this: first, proteomic data are much more complex than transcriptomic data, particularly in terms of the host of potential posttranslational modifications that could exist. Second, proteomics technologies are rapidly developing, for example, the iTRAQ method has only been in existence for a few years. Finally, a large number of proteomic experiments are required to obtain full proteome coverage for a sample of interest. Obtaining full proteome coverage over the large number of conditions available for Arabidopsis gene expression would be beyond the funding level of most plant research grants. That said, a MIAPE (for Minimum Information About a Proteomics Experiment) specification for proteomics data has been developed (Taylor et al., 2007) so that the aforementioned details (i.e., experimental metadata) are reported for published experiments.

Despite the lack of quantitative proteomics data sets in public repositories, such as GEO, several qualitative proteomics experiments using 2D gels followed up by mass spectrometric identification have been conducted to document the subcellular localization of proteins in Arabidopsis. The Arabidopsis Subcellular Database (SUBA) at www.plantenergy.uwa.edu.au/suba2/ (Heazlewood et al., 2007) has collated data from >1000 publications, documenting the subcellular localization of >6743 Arabidopsis proteins mainly based on mass spectrometry and green fluorescent protein fusion experimental data. The query interface allows the use of Boolean operators to look for overlaps in proteins identified in various published data sets, which is sometimes surprisingly low even for two proteomes from ostensibly the same subcellular compartment. Whether this is indicative of dynamic proteomes, difficulties in obtaining complete proteome coverage, or experimental error is unclear. Furthermore, predictions run with 10 common subcellular localization prediction programs have been applied to the entire Arabidopsis proteome with the results that most Arabidopsis proteins, if not documented, at least can be inferred to be in a certain compartment or compartments. Knowing the subcellular localization of a protein is vital for understanding its function. The BAR's Cell eFP Browser (Winter et al., 2007) at BAR.utoronto.ca displays SUBA's documented and predicted subcellular localizations in a pictographic manner, according to the confidence of the localization method, as described earlier.

No comprehensive data set exists for the Arabidopsis interactome, although several NSF 2010 projects are currently underway to document this, for example, the interactions between membrane proteins and proteins in the “unknowneome.” Two studies have attempted to predict interactions based on orthology to interacting proteins in other species (Geisler-Lee et al., 2007; Cui et al., 2008). In the case of the AtPID at atpid.biosino.org (Cui et al., 2008), the authors have also used coexpression matrices and protein domain co-occurrence to infer interaction. The 19,979 predicted interactions described in the Geisler-Lee et al. (2007) publication are available through the Arabidopsis Interactions Viewer (AIV) at the BAR (BAR.utoronto.ca), as well as 50,000 more from a more recent iteration of the approach using more organisms. The AIV also contains >2000 literature-documented, biochemically or genetically assayed interactions for Arabidopsis, some from studies with an individual protein and others from experiments conducted in a more high-throughput manner, such as those using protein microarrays to detect interactions between calmodulin-related proteins or mitogen-activated protein kinases and their protein targets (Popescu et al., 2007; 2009).

METABOLOMICS

The output of the plant proteome is in part a huge diversity of small molecules, which is apparently many times more diverse than the small molecule component of mammalian proteomes (compare ∼200,000 different small molecules in the plant kingdom space [Fiehn, 2002] to the ∼6500 for humans as documented in the Human Metabolome database [www.hmdb.ca]). For a researcher, it is important to know what, if any, small molecule could be produced by a given gene product of interest (or if there are any small molecules that act upon it, which could be answered using BRENDA at www.brenda-enzymes.org if it is an enzyme) (Schomburg et al., 2002) or if a given stimulus/mutation causes an overall perturbation of the metabolome. Unfortunately, the scope of metabolomic experiments in plants is very small, with only a limited number of biological conditions examined to date and large gaps in our knowledge of biosynthetic pathways. This reflects the fact that many metabolomic methods are still in development, in part limited by the capabilities of current instrumentation, the development of a comprehensive set of library standards, and in the laborious annotation of as yet unidentified metabolites. Identification of these metabolites will complete our picture of biological processes occurring within plants by helping us to characterize metabolic pathways and their intermediates and signaling molecules more definitively.

The Golm Metabolome Database (csbdb.mpimp-golm.mpg.de/csbdb/gmd/gmd.html) contains several metabolomic experiments conducted on Arabidopsis plants grown, for example, under different light intensities (Kopka et al., 2005). It is possible to query the database for a given compound and to identify experiments for which the compound of interest was found to be higher or lower than a given threshold. MetNetDB (MetNetDB.org), out of Iowa State University, documents compounds in metabolic pathways and links these to gene products, in a manner similar to AraCyc at www.Arabidopsis.org/biocyc/ (Mueller et al., 2003), MapMan at www.gabipd.org/projects/MapMan/data.shtml (Thimm et al., 2004), KEGG Atlas at www.genome.jp/kegg/atlas/metabolism/ (Okuda et al., 2008), Reactome at reactome.org (Tsesmetzis et al., 2008), and other pathway databases. As metabolomic data sets become more prevalent, it would be highly desirable for GEO or some other larger database to serve as the primary repository for the raw data generated by these experiments. Other more specialized tools could then be developed based on subsets of data from the primary repository, a model that has worked very well for minimum information about a microarray experiment (MIAME)-compliant (Brazma et al., 2001) transcriptome data sets. Fiehn et al. (2005) operate BinBase at http://eros.fiehnlab.ucdavis.edu:8080/binbase-compound/, which documents >1000 small molecules from plants, and they and others are actively involved in the creation of the Metabolomics Standards Initiative to bring MIAME-like standards to metabolomics experiments (Fiehn et al., 2007).

INTEGRATIVE RESOURCES

The inclusion of each of these types of large-scale data sets in easy-to-use, queryable browsers is of great importance for hypothesis generation. In particular, genome browsers, like the TAIR Genome Browser, include multiple sources of data that users can integrate within their queries. This allows the user to identify genetic variation at the sequence level that may lead to alteration in regulation of gene expression or protein function across diverse accessions. Expression browsers, like the BAR and Genevestigator, which permit visualization or interrogation of gene expression at the anatomical level, or at the level of response to a stimulus, can be used to generate hypotheses about gene function. The identification of genes coexpressed with one's gene of interest, or clusters of genes that are coregulated, and the mining of these gene groups for functional association via overrepresentation of Gene Ontologies allows for in silico prediction of gene function. Further mining of these lists for overrepresented upstream regulatory sequences can identify putative regulatory factors. Ultimately, however, integration of the types of data sets described here toward the generation of multilevel regulatory networks (multinetworks) for hypothesis generation is desirable. Furthermore, development of methods to query these networks in a statistical manner that also assesses and weighs the validity of the data sources is necessary.

Generation of a multinetwork that incorporates multiple sources of data in Arabidopsis has been accomplished in a queryable Web browser named VirtualPlant at VirtualPlant.org (Gutierrez et al., 2007a). This multinetwork incorporates data for metabolic pathways, known protein–protein, protein–DNA, miRNA–RNA, and predicted protein–protein and protein–DNA interactions (Gutierrez et al., 2007a). Resulting gene networks are visualized using Cytoscape, and regions of high connectivity can be identified using Antipole, a graph clustering algorithm (Ferro et al., 2003). The original VirtualPlant multinetwork contained 6176 gene nodes, 1459 metabolite nodes, and 230,900 edges (or interactions) between these nodes (Gutierrez et al., 2007a). This network has also recently been expanded to include bioinformatically identified protein–DNA interactions (Gutierrez et al., 2008). Subnetworks can be identified by querying multinetworks with a list of genes, often identified from gene expression analysis and statistically tested for significance. Functional annotations can also be overlaid upon identified subnetworks of the multinetwork to help infer subnetwork function. In one interesting approach (Thum et al., 2008), supernode networks were generated by collapsing genes from a subnetwork into a category according to both their metabolic pathways and the first two words of their gene annotation, although the statistical significance of these supernode annotations was not tested. The resulting size of the node is proportional to the number of genes annotated to that node (Thum et al., 2008). The VirtualPlant system has been used successfully to define gene networks in various signaling pathways as further described in Bioexample 5.

While VirtualPlant's integration of multiple data sources into a cohesive queryable system is an important advancement in our ability to make sense of and use large-scale data sets, attributing measures of confidence to an edge between two nodes, as defined by experimental evidence, would greatly improve accuracy in defining network interactions. For example, a predicted protein–DNA interaction should not be weighed as heavily as an experimentally verified protein–DNA interaction. Furthermore, for a set of experimentally verified interactions, interactions with multiple sources of experimental support should be given greater confidence than an interaction with a single source of experimental support. Attributing such measures of confidence is not a simple task, as many of these data sources are heterogeneous and require explicit knowledge of how each data set was obtained experimentally. Access to information, such as the statistical methods used to define a gene as expressed, a polymorphism as a deletion based on array hybridization signal, a promoter as marked by an epigenetic modification, a metabolite or protein as present and properly annotated, and the in planta relevance of interactions between plant proteins detected in yeast two hybrid assays should also be required when synthesizing multinetworks or when using data from these multinetworks.

Integration of diverse data types in a statistical framework to infer gene function or to identify gene or protein interactions has been accomplished for a wide variety of organisms, including yeast, mouse, and humans (Myers et al., 2005; Lee et al., 2007; Guan et al., 2008; Kim et al., 2008; Mostafavi et al., 2008; Ramani et al., 2008). Methods and guidelines to integrate and correlate such heterogeneous data have been described by Lee and Marcotte (2008) and provide good principles that should be taken into account in the plant community, especially as integration of multiple data sets has been shown to outperform individual functional genomics data sets in accuracy and coverage in hypothesis validation.

As an example, algorithms like GeneMANIA tackle this computationally complex problem at the level of the individual network using data from several levels (expression profiles, protein–protein interactions, subcellular localization, etc.). Functional prediction analyses are possible for several organisms, now including Arabidopsis (Mostafavi et al., 2008). The authors assign a weight to each network derived from a single data source that reflects its usefulness in predicting a given function of interest. To construct the final composite network, they then take the weighted average of the combined association networks. Furthermore, at the node level, GeneMANIA incorporates genes positively associated with a label from a particular network, genes that are unlabeled, and genes that are negatively labeled. This method has proven more accurate in prediction of GO category association than leading methods on mouse and yeast functional data, using the area under the curve for the resulting receiver operating characteristic curves (Pena-Castillo et al., 2008). For the mouse data, this is primarily due to their inclusion of genes that are negatively labeled. GeneMANIA is also available on a Web server for easy access (see Table 1). The application or development of such algorithms to available plant large-scale data sets is greatly needed. In addition, the concept of a competition for critical assessment of Arabidopsis gene function prediction, similar to that held for mouse (Pena-Castillo et al., 2008), in which several groups submit their best predictions on a benchmark data set assembled by organizers, might be attractive, especially considering some of the novel data types available in this species.

FUTURE DIRECTIONS

Comparisons across species are often key to understanding a biological process. To make accurate cross-species comparisons, it is necessary to have a vocabulary representing the similarity of form and function in the species under consideration. For this reason, the Plant Ontology resource was developed based on anatomical and functional features of Arabidopsis, rice, and maize, with others being added to describe other crop species (plantontology.org). Ontology terms include those describing tissues and cell types, organs and organ systems, and those denoting particular stages, such as senescence or germination. To permit the incorporation of data from as yet unsequenced genomes, a standardized staging and developmental state for each plant organ was developed (Jaiswal et al., 2005; Ilic et al., 2007). Careful annotation of the tissue samples using the Plant Ontology system would make it easy to query, for example, the response of orthologous genes in similar tissues in related or more distant species. In a similar manner, a recently funded National Institutes of Health Genome Research Resource Grant (P41) to establish a Pathway Commons to facilitate the exchange, integration, and distribution of biological pathway information will maintain and extend the BioPAX exchange language for biological pathways and develop improved software for querying these.

The iPlant Collaborative (iplantcollaborative.org) recently decided on which Grand Challenge proposals to assist Plant Cyberinfrastructure will be financially supported. These are Assembling the Tree of Life for the Plant Sciences, which is focused on the design and creation of a phylogenetic cyberinfrastructure, and Cyberinfrastructural Support for the Genetic and Ecophysiological Decipherment of Plant Phenological Control in Complex and Changing Environments.

A looming challenge, presumably in part to be addressed by the second iPlant Collaborative Grand Challenge listed above, presents itself with next-generation sequencing initiatives, which have revolutionized genome sequencing abilities in terms of time and cost. Within the next 10 years, thousands of plant genome sequences will be released, using a variety of different sequencing platforms, including the Roche 454 pyrosequencing system, the Illumina Genome Analyzer, and the Applied Biosystems SOLiD system. Plant research is benefitting from these technologies, with several Arabidopsis accessions recently sequenced and many more plant genomes in the sequencing pipeline (Ossowski et al., 2008). Once these large data sets are publicly released, the next challenge is in how to deal with the resulting bioinformatic bottlenecks. In particular, base calling software, methods to score sequence quality, and alignment software can all differ depending on both the sequencing platform and the researcher's personal choice and should be accounted for in sequence browsers that publicly display these data. Available bioinformatic software is well described by Shendure and Ji (2008). Community accepted guidelines for compiling these and other metadata associated with these data sets, in a manner similar to that of MIAME guidelines (Brazma et al., 2001), will be extremely important in the future. A draft version of this guideline, termed MINSEQE (minimum information about a high-throughput sequencing experiment) has been proposed (www.mged.org/minseqe/). NCBI has established a short read archive (SRA) (http://0-www.ncbi.nlm.nih.gov.lib1.npue.edu.tw/Traces/sra/) that accepts next-generation sequencing data from a variety of platforms and includes data from de novo sequencing experiments, resequencing experiments, structural variation discovery, and SNP calling experiments. The SRA tracks metadata associated with each experiment and should help improve database efficiency by normalizing data structures. GEO accepts next-generation sequencing data from mRNA sequencing, ChIP sequencing, bisulfite sequencing, and small RNA discovery and profiling experiments (www.ncbi.nlm.nih.gov/projects/geo/info/seq.html). Metadata are also associated with these submissions. A remaining challenge is in displaying this wealth of sequence data in a user-queryable, Web interface format in a manner that allows the user to extract biologically meaningful information from these data sets.

Clearly, the trend toward generating more and more data, especially sequence and expression data, will necessitate the development of new computational tools for viewing, querying, and analyzing such data. How will the average “wet lab” scientist be able to use data from 1001 genomes, let alone view them? Interestingly, it would seem that the generation of such data will lead to the reunification of the ecology and evolutionary, and cell and molecular fields of plant biology.

In the meantime, so-called SOAP (Simple Object Access Protocol) services, such as various BioMOBY resources (Wilkinson and Links, 2002), are being developed by plant bioinformatic groups worldwide. These services promise to allow databases and Web-based tools to “talk” to one another, thereby automating specific aspects of creatively thought out analyses now becoming possible with ever more published large-scale data sets. For the average plant biologist, there is already a wealth of information available with existing Web-based tools, and he or she would be wise to embrace the computer as the “new molecular biology.”