Background

Accurate characterization of as many genes as possible is a central challenge of the post-genomic era and an essential precondition for progress in systems biology [1]. But this characterization is very far from completion. By various estimates, at least 30-50% of the genes of any given organism are of unknown function [2], incorrectly annotated [3, 4], or have only a generic annotation such as ‘ATPase’ [5]. This problem is particularly acute for eukaryotic genomes, which are in general less well annotated than prokaryotic ones [6, 7].

Moreover, with more than 6,000 genomes now (August 2010) in the pipeline, 1,354 of them eukaryotic (http://www.genomesonline.org), the numbers of unknown genes continue to increase [8] and annotation errors continue to increase even faster [9]. For some gene families up to 60% of the annotations are wrong [9]. Without specific functional annotation efforts, present and future genome information will become ever more corrupt and hard to analyze, and will thus be greatly underexploited.

The first step in linking gene to function is to define what constitutes a function, and this is not trivial. Full definition of a protein’s function requires a combination of two features or ‘dimensions’: (i) a molecular function (e.g. an enzymatic activity) and (ii) a functional context (e.g. a pathway) comprising other proteins involved in the same process. Currently most annotations in public archives convey only molecular functions, mainly assigned by homology. However, when an enzymatic activity has been annotated in this way, it may well be wrong if other genes of the same pathway are not in the genome [10]. To decide whether a protein has a truly known function, it is therefore essential to take into account both the molecular and functional context dimensions. Most automated annotation platforms use only the molecular function, but when metabolic reconstruction (i.e. pathway context) is included in the annotation process this greatly improves annotation quality [11–13].

We and others have previously emphasized the power of cross-kingdom comparative genomics approaches to link gene and function [8, 14]. This strategy was applied in the work presented here to families of unknown function shared by Arabidopsis thaliana and prokaryotes. Using the series of sieves summarized in Fig. 1, we combined comparative genomic and experimental validation approaches to discover the function of ‘unknowns’. Throughout this work, our primary comparative genomics platform was the SEED database and its tools [10]; the SEED is publicly available at http://www.theseed.org/Papers/20101120/.

Figure 1
figure 1

Project workflow. The overall strategy that combined in silico and experimental validation is presented showing the number of genes that were analyzed at each stage.

Full size image

Results and discussion

Selecting candidate hypothetical genes families conserved in plants and prokaryotes

Generation of the starting Arabidopsis gene set

The full set of 26,207 Arabidopsis genes was extracted from the re-annotated genome [15]. To predict functions for ‘unknown’ genes conserved among prokaryotes and plants, it is important to avoid large gene families because their members often have different functions. The Tribe [16] and TIGR [17] algorithms were therefore used to filter out genes belonging to families having four or more members in Arabidopsis, leaving 9,250 genes corresponding to 6,034 gene families (Table 1; personal communication, Dr. Brian Haas, The Institute for Genomic Research).

Table 1 Selection of candidate hypothetical genes families conserved in Arabidopsis (AT) and prokaryotes for in silico functional predictions and potential experimental verification – an overview.
Full size table

Selecting gene families conserved between plants and prokaryotes

A second filter was applied to the 9,250 genes to retain those whose products have prokaryotic homologs. BLASTP [18] searches were performed in summer 2008 against the approximately 650 complete or almost complete microbial genomes then available in the SEED database. The probability threshold (E value) of better than 10-10 was imposed to ensure sufficient functional conservation between amino acid sequences included in the analysis [19, 20]. Approximately one quarter of the 9,250 Arabidopsis genes tested were found to be similar to at least one prokaryotic gene (2,325 total, Supplemental Tables 1A, 1B, and 1C). Prokaryotic homologs for additional Arabidopsis genes would most probably be detected were this comparison to be repeated with the more numerous (~1,000) and more diverse microbial genomes now available.

Selecting hypothetical Arabidopsis/prokaryotic gene families

Several strategies were combined to extract from the set of 2,325 conserved Arabidopsis genes those whose functions are unknown or poorly known. The following sources of evidence (available as of summer 2008) were considered: (i) Arabidopsis gene annotations in the TAIR database [21]; (ii) SEED annotations of prokaryotic orthologs of Arabidopsis genes; (iii) the list of Arabidopsis proteins of unknown function (PUFs) [22] (http://bioweb.ucr.edu/scripts/unknownsDisplay.pl); and (iv) publications in PubMed and TAIR databases (or the absence thereof).

We relied mainly on the second of these sources, i.e. SEED annotations of prokaryotic orthologs of the candidate Arabidopsis genes. The nontrivial task of establishing gene orthology is greatly aided in SEED by the subsystem-based organization of annotations (described in Methods; [10, 23]). We considered Arabidopsis genes to be ‘known’ and excluded them from further analysis if they or any of their prokaryotic orthologs were associated with SEED subsystems that are classified as non-hypothetical; i.e. encoding established metabolic pathways, physiological processes, or structural complexes (as opposed to experimental or hypothetical subsystems that group uncharacterized genes based on various criteria, including co-localization, co-regulation, common phenotype, etc.)

The list of Arabidopsis PUFs [22] served as a secondary resource. Three distinct PUF identification methods, complementary to our approach, have been used to compile this list: (i) BLASTP searches against proteins of known function in Swiss-Prot; (ii) Hidden Markov Model-derived searches against the Pfam domain database; and (iii) retrieval of the 'unknown' annotations from the Gene Ontology system [22]. Finally, to keep abreast of discoveries of functions for Arabidopsis genes, publications associated with selected Arabidopsis genes were extracted from the TAIR database and PubMed repeatedly during the course of this project.

This analysis yielded a set of 630 hypothetical Arabidopsis genes, corresponding to about 360 gene families common to Arabidopsis and prokaryotes that were highly enriched in those specifying proteins of unknown function. This gene set was then prioritized for in-depth in silico analysis as described below.

Prioritizing Arabidopsis/prokaryotic gene families for detailed in silico analysis

General strategy

As the pool of 360 gene families was still too large for the labor-intensive process of in-depth comparative genomic analysis, we prioritized the candidates for further analysis based on several characteristics associated with each protein-encoding gene in the SEED database. The main such characteristic was the presence of ‘functional coupling’ or ‘conserved gene clustering’ [24–27] for a prokaryotic member of the family, but other criteria were also computed when available as detailed below.

Detecting and analyzing gene clustering

Physical gene clustering is the tendency of functionally associated genes to be located near each other on the chromosome. Although not entirely absent in eukaryotes [28], such clustering is far more marked in prokaryotes, in which functionally related genes are often arranged in operons [29] or divergently transcribed from the same promoter region [24], or are simply neighbours or near-neighbours [24, 30]. On average, ~35% of bacterial metabolic genes are in clusters [24]. A key point is that the more taxonomically diverse the genomes in which a cluster occurs, the more informative the cluster becomes [30]. A single gene family can be involved in different clusters in different taxa (all potentially diagnostic of its function), even if it is not clustered with informative genes in all taxa.

Several software tools in SEED that take advantage of gene clustering were used to select promising candidates, as well as to link unknown gene families to general metabolic pathways and to generate specific functional predictions during the next phase of the project:

(i) Strength of ‘functional coupling’ (FC) – measures the number of distantly related organisms (with 95% overall DNA sequence identity or less) in which two genes are located in each other’s vicinity. Close strains are not taken into account in this parameter, for example: all sequenced Escherichia coli genomes in SEED in which two particular genes are co-localized on the chromosome are counted as one when computing FC (see ref. [24] for a more formal treatment of this topic).

(ii) Length of cluster – reflects the number of genes involved in a specific cluster.

(iii) Evidence code ‘in c luster w ith n on-hypothetical’ (cwn) – indicates that a gene family is functionally coupled to (tends to co-localize with) at least one other gene family that has been assigned a function that is considered ‘non-hypothetical’. The functional coupling score must be five or more for this code to apply.

(iv) Evidence code ‘in c luster w ith h ypothetical’ (cwh) – as above, except it labels gene families that tend to co-localize with at least one other hypothetical gene family;

(v) Association of a gene family with a ‘clustering based’ subsystem in SEED (CBSS). These subsystems group hypothetical protein families solely on the grounds of co-localization patterns conserved across multiple genomes; however, manual subsystem encoding takes automatically pre-computed leads (such as evidence codes) to the next level. For example, comprehensive phyletic spread is determined for protein families that might have been labeled as ‘in cluster with hypothetical’ in merely a fraction of genomes. CBSS subsystems provided a useful starting point for the next phase of in silico analysis.

Other filtering factors

Other factors that were considered included:

(i) Phyletic spread – the number of distinct microbial species that harbored members of each hypothetical family under consideration, whether or not they were functionally coupled (see above). Widely distributed families were preferred over narrowly distributed ones.

(ii) Whether or not well studied model organisms such as E. coli, Bacillus subtilis, Pseudomonas aeruginosa, cyanobacteria, or yeast contained a member of the gene family in question, indicating likely availability of functional genomics data (e.g. expression arrays, gene essentiality data, protein interaction datasets) that could provide clues linking candidate families to general metabolic areas and aid specific functional predictions.

Comprehensive tables summarizing all types of association evidence (available as Additional Files 1 and 2) were used for manual sorting and evaluation to prioritize hypothetical plant/prokaryote gene families and to select candidates for further detailed in silico analysis.

Linking unknown gene families to general metabolic areas

Our first goal was to link the prioritized gene families to a particular metabolic or functional area. For this, we built on the gene clustering associations captured as described above, using these to construct a corresponding experimental subsystem in the SEED database for each family (see Methods). Each such subsystem included all members of the focus gene family across all genomes available in the database, as well as gene families potentially associated with it (as implicated by gene clustering in at least a fraction of prokaryote genomes). Such integration of biological functions with genome sequences provided by subsystems allowed us to discard or strengthen the clustering associations and to evaluate the phylogenetic co-distribution between the gene family of interest and the associated families. We then further explored the associations by extending the analysis to Arabidopsis. Indeed, organization of genomic data in subsystems allows accurate extrapolation of functional associations between genes detected in microbial genomes onto other prokaryotes and even eukaryotes (e.g. Arabidopsis), albeit with caution. For example, in our study the degree of similarity between an Arabidopsis gene and its nearest prokaryotic homolog involved in gene clustering played an important role in evaluating the validity of such cross-kingdom projections. Additional ‘checks and balances’ were used in projecting functional leads and hypotheses developed via comparative analysis of prokaryotic genomes back to plant genes, which might not have preserved the function of their prokaryotic counterparts. For example, when linking unknown gene families to general metabolic areas, or to individual protein families via gene clustering, the validity of such associations in the context of plant physiology and biochemistry was considered, as well as its correspondence to Arabidopsis expression array data [31–33], protein localization [34, 35], mutant phenotypes, and other functional genomics data (as illustrated in the case studies below). This first analysis yielded a list of 78 gene families linked to diverse metabolic areas (Additional file 2), including fatty acids, terpenes, vitamins, aromatic compounds, and sulfur, as well as iron-sulfur cluster assembly, oxidative damage protection, glutathione S-transferase-dependent detoxification, DNA repair, plastid/cell division, signalling systems, and metal homoeostasis. However, a clear bias reflecting the investigators’ areas of expertise was observed, with some ten in vitamin/cofactor metabolism and another fourteen in tRNA/RNA modifications. This emphasizes the value of combining multiple types of expertise to accurately predict and validate gene function.

Predicting and testing precise molecular and biological functions

General strategy

The next step in the pipeline was creating a functional hypothesis that could be tested by genetic and/or biochemical experiments. This is by far the most labor intensive and intellectually challenging step in the pipeline. Multiple types of data need to be queried and integrated with biochemical insights in order to make reasonable and testable predictions. Clues can come from high-throughput data (protein complexes, phenotypes, microarrays) from any organism, from the literature (where data may be buried in supplemental tables and contain no reference to the gene family), or from analysis of the structure of a member of the family (e.g. from a structural genomics effort). Biochemical insight can come from cataloguing globally or locally missing genes, i.e. those that encode enzymes for which a gene has never been identified in any species or is absent in certain species [36–38].

An example of globally missing gene identified in this work is the Sua5/YrdC family involved in the universal carbamoylthreonyladenosine (t6A) modification in tRNA (case number 5 in Table 2 and summarized below). An example of a locally missing gene is the PTPS-III family (case number 4 in Table 2) that replaces the folate biosynthesis enzyme FolB in many bacteria and certain eukaryotes [39, 40]. Biochemical insights can also come from noting the presence of two gene families annotated as fulfilling the same role, suggesting a possible duplication followed by functional divergence [41]. One such example is the COG0354 family, previously annotated in many genomes as the folate-dependent glycine cleavage system T protein (GcvT), but in reality a protein involved in the repair of iron-sulfur clusters (case number 1 in Table 2 and summarized below). For some families very precise functions could be predicted, e.g. methylation of a specific position in ribosomal RNA (At4g28830, case number 31 in Table 3) whereas for others the prediction remained more general but testable nonetheless. For instance, we were able to link certain members of the COG0523 family (case number 7 in table 2) to zinc homeostasis [42] and this general prediction was borne out by demonstrating that some members of the family have a role in survival in low zinc conditions ([43] and C. Blaby-Haas and V. de Crécy-Lagard, unpublished results). We were able to make testable functional predictions for 41 families (Additional file 2); the rationales for these predictions are summarized in the subsystem notes for each family in the SEED database. Table 2 lists the 19 families for which the prediction has been experimentally confirmed by us or others. Table 3 lists four families that were experimentally invalidated and another eight for which experiments are well advanced, for a total of 31 families. Three illustrative examples of validated predictions are described briefly below. The first and second of these are fully described elsewhere [44, 45].

Table 2 Status of the experimentally validated families: cases 1-9 verified by us; cases 10-19 verified by others.
Full size table
Table 3 Status of the families invalidated by us (cases 20-23) or in progress (cases 24-31).
Full size table

COG0354 (At4g12130, At1g60990)

Bacterial genes encoding COG0354 (case number 1 in Table 2), the GcvT paralog noted above, often cluster with diverse iron/sulfur (Fe/S) proteins (shown in red in Fig. 2A), and proteomic data [46, 47] show induction by oxidative stress and confirm an Fe/S association. Moreover, the COG0354 protein is required for full activity of certain Fe/S enzymes in E. coli[48] and yeast [49]. We therefore predicted that COG0354 is a folate-dependent enzyme (based on its homology to the folate-dependent protein GcvT) involved in assembly or repair of Fe/S proteins, particularly under oxidative stress. Consistent with this prediction, deleting the gene encoding COG0354 in E. coli (ygfZ) increased oxidative stress sensitivity, and the stress-sensitive phenotype was complemented by expressing a plant COG0354 protein (Fig. 2B). Folate-dependence was established by using NMR to demonstrate stereoselective folate binding by recombinant E. coli COG0354, and by showing that in vivo activity of the E. coli Fe/S protein MiaB is as seriously impaired by deleting the folate synthesis gene folE (which eliminates folates) as by deleting ygfZ, i.e. removing folates had the same impact as removing COG0354 [44].

Figure 2
figure 2

Clustering arrangements of genes encoding COG0354 and functional complementation of an E. coli COG0354 deletant by an Arabidopsis COG0354. (A) Clustering of COG0354 genes with Fe/S-related genes. Blue, COG0354; red, Fe/S proteins; rose, proteins in same complex or pathway as Fe/S proteins; turquoise, Fe/S cluster assembly proteins. Rx, Rubrobacter xylanophilus; Sm, Stenotrophomonas maltophilia; Pu, Pelagibacter ubique. (B) Growth of an E. coli COG0354 (ygfZ) deletant harboring plasmid-borne E. coli ygfZ, Arabidopsis mitochondrial COG0354, or vector alone on LB medium or LB plus the oxidative stress agent plumbagin (OX) (30 μM), arabinose (0.02% w/v), and appropriate antibiotics.

Full size image

COG3643 (At2g20830)

The histidine utilization (Hut) pathway occurs in certain bacteria and animals, but not plants. The Hut pathway up to the intermediate N-formiminoglutamate is invariant, but thereafter there are three routes to the end-product glutamate, one of which involves a formiminotransferase, COG3643 (Fig. 3A). Comparative genomics analysis showed that bacteria that have a formiminotransferase-type Hut pathway generally lack the ygfA gene encoding 5-formyltetrahydrofolate cycloligase, the key enzyme required to recycle 5-formyltetrahydrofolate, which inhibits various folate-dependent enzymes and is formed by a side reaction of serine hydroxymethyltransferase in the presence of glycine (Fig. 3B). This striking observation led us to predict that formiminotransferase or formiminotransferase paralogs can replace YgfA. This prediction fits with classical biochemical data showing that mammalian formiminotransferase can mediate formyl transfer from 5-formyltetrahydrofolate to glutamate, albeit at a low rate [50, 51]. The prediction was supported by showing that various prokaryotic COG3643 genes (highlighted in Fig. 3B) complement the growth phenotype of an E. coli ygfA deletion mutant (which cannot use glycine as sole nitrogen source, presumably because 5-formyltetrahydrofolate accumulation inhibits the folate-dependent glycine cleavage reaction). Representative data for the Acidobacterium COG3643 gene are shown in Fig. 3C. Folate analysis of the ygfA deletant with and without complementing COG3643 genes confirmed that the deletant accumulated 5-formyltetrahydrofolate and that COG3643 genes reversed this accumulation [45]. Furthermore, characterization of recombinant COG3643 proteins showed their kinetic characteristics to be consistent with an in vivo role in 5-formyltetrahydrofolate recycling [45]; this biochemical corroboration is important since functions carried out by ectopically overexpressed genes do not necessarily reflect their native function. Taken together, this evidence suggests that COG3643 paralogs in plants may likewise replace YgfA. Consistent with this possibility, the ygfA knockout in Arabidopsis has a mild phenotype, pointing to the existence of an alternative route for disposal of 5-formyltetrahydrofolate [52].

Figure 3
figure 3

COG3643 in relation to the Hut pathway. (A) Hut pathway; note the three different routes. (B) The distribution of histidine utilization genes among bacterial and eukaryal genomes in relation to that of the ygfA gene for 5-formyltetrahydrofolate disposal. Gene colors correspond to different parts of the pathway as in part A. Lines between boxes denote gene fusions. (C) Growth of an E. coli ygfA deletant harboring plasmid-borne E. coli ygfA, Acidobacterium COG3643, or vector alone on minimal medium with NH4Cl or glycine as sole nitrogen source. The medium contained 1 mM IPTG and appropriate antibiotics.

Full size image

YrdC/Sua5 (At5g60590)

The universal base modification t6A occurs at position 37 in a subset of tRNAs decoding ANN codons. The biogenesis of this complex modification is yet to be elucidated but is known to require threonine, ATP, and bicarbonate [53–55]. COG0009 was predicted as a possible candidate for a missing t6A biosynthesis family because it occurs in all genomes sequenced to date, is known to bind double-stranded RNA [56], and has been linked to defects in translation in both prokaryotes and eukaryotes [57, 58]. This conjecture was supported by sequence homology with the [Ni-Fe] hydrogenase maturation protein HypF, which catalyzes a reaction analogous to the one expected for a t6A enzyme [59]. The COG0009 family can be further split based on sequence comparison into three subfamilies: YrdC, Sua5 (YrdC with an extra domain termed Sua5), and YciO. One or two members of this family are present in each genome; for example, the Arabidopsis and E. coli genomes contain two, YrdC (At5g60590) and YciO (At3g01920), while the Saccharomyces cerevisiae genome contains only one, Sua5 [60]. We showed that (i) tRNAs from S. cerevisiae strains lacking sua5 do not contain t6A and that this phenotype is complemented by transforming with a plasmid encoding the wild type gene (Fig. 4A); (ii) the homologs from B. subtilis, M. maripaludis, E. coli yrdC, but not E. coli yciO also complement the phenotype; (iii) the yrdC homolog is essential in E. coli, whereas yciO is not; (iv) S. cerevisiae, B. subtilis, M. maripaludis yrdC genes but not E. coli yciO are able to complement the lethality phenotype of yrdC in E coli (Fig. 4B); and (v) E. coli yrdC is able to bind t6A apomodified tRNAThr but not unmodified transcript [60]. Therefore, members of the YrdC/Sua5 family are involved in t6A biosynthesis. In Arabidopsis, At5g60590 and At3g01920 are annotated as related to the YrdC family. However, based on comparative sequence analysis combined with genetic orthologous complementation tests these can be distinguished. As depicted in Fig. 4C, the YrdC family is characterized by the KxR/G …SxN signature sequence; At5g60590 is therefore most probably part of the YrdC family while At3g01920 is not.

Figure 4
figure 4

Separation of the COG009 family into two subgroups YrdC and YciO based on motifs and functional assays. (A) Complementation of the t6A- phenotype of the yeast Δsua5 (YGN63) strain by the E. coli yrdC gene but not the E. coli yciO gene. (B) Complementation of the yrdC essentiality phenotype in E. coli by yrdC subfamily members from E .coli (EcyrdC), Bacillus subtilis (BsywlC), Methanococcus maripaludis (MmyrdC) and yeast (Scsua5) but not by yciO from E. coli (EcyciO). All genes were cloned in pBAD24 [95] and were therefore expressed in the presence of arabinose (Ara, 0.2%) and transformed in an E.coli strain carrying the chromosomal copy of yrdC under PTet control [96] that does not grow in the absence of anhydrotetracycline (ATc, 50 ng/ml). (C) Signature motif of the functional homologs of YrdC (KxR/SxN) that are not found in the YciO subfamily. In green are the two homologs from Arabidopsis and their distribution.

Full size image

Propagating validated and predicted gene functions to other organisms

Subsystem-based organization of genomic data in SEED implies delineation and maintenance of isofunctional gene groups [61]. This strategy greatly facilitates not only the development of functional predictions for uncharacterized genes, but also accurate projection of this knowledge, once verified, to other species. For example, annotations for the 19 families developed in the course of this study and experimentally confirmed by us or others have been propagated to a total of 6,297 genes in some 900 complete or nearly complete bacterial, archaeal, and eukaryotic genomes currently available in the SEED database (Table 2). Furthermore, we believe there is merit in accurate and exhaustive propagation of yet untested predictions to all orthologs in all available genomes, early in the process of in silico analysis. This allows complete and accurate cataloguing of functional homologs for each gene family under study, thus revealing its phyletic spread, co-occurrence with known gene families, potential associations with specific features of an environmental niche – all of which can serve as additional clues for developing specific functional hypotheses. For this reason we have built subsystems in the public SEED database for 78 families (Additional file 2) so that readers can produce and test their own predictions for gene families that fall in their area of expertise. The total number of annotated genes in these families exceeds 22,000.

Comparing results to those from automated functional prediction platforms

Several recently developed platforms seek to automatically integrate comparative genomics, high-throughput experimental data, and literature reports to make gene function predictions [46, 62]. For ten gene families that were predicted and validated, we analyzed the accuracy of the corresponding predictions from the two most relevant predictions platforms, eNet (E. coli) [46] and AraNet (Arabidopsis) [62] (Table 4). The predictions from these automated platforms were mostly wrong. At best, they produced a general annotation that was in the right functional area such as ‘folate dependent regulatory protein’ by eNet and ‘iron-sulfur assembly protein’ by AraNet for the COG0354 family. However, both of these came straight from the literature, not from associations, and can thus hardly be called predictions; furthermore, these two correct ‘predictions’ were buried in long lists of incorrect ones. Thus, whenever a prediction was possible, the automated platforms failed to make a correct and precise one.

Table 4 Comparison of functional predictions in eNet and AraNet for 10 of the protein families.
Full size table

Conclusions

The analysis presented here shows that combining comparative genomics with expert intellectual input enabled correct functional annotation of 19 gene families by only a few researchers, in a short time (three years), and at a moderate cost (<$1M). This number of successful functional predictions is roughly comparable to the number made through the entire structural genomics effort [63], involving many more people, a much longer period, and far greater expense. The cost-effectiveness of our approach is thus perhaps its most striking feature.

Our analysis also underscores the imperative of combining molecular function and biological context to annotate function as shown in the COG3643 example. Homology and even in vitro assays would have labeled this family – correctly, but incompletely and misleadingly – as a formiminotransferase. Only by interpreting phylogenetic distribution data with biochemical insight and then applying complementation tests was this family correctly annotated as an alternative to 5-formyltetrahydrofolate cycloligase to metabolize 5-formyltetrahydrofolate. It is noteworthy that in this and most of our other successful predictions, there is a strong bias towards the authors’ areas of expertise – from which the obvious inference is that other experts would, equally easily, have been able to predict additional sets of functions. For this reason we have built subsystems in the public SEED database for 78 families and made available the raw comparative genomic data for all gene families shared between Arabidopsis and prokaryotes (Additional file 1A, 1B, and 1C) so that other experts can bring their insight to our analysis in order to make and test their own predictions.

Finally, the only plant genomes available when this effort started were Arabidopsis, rice, and poplar, and rich post-genomic resources were available only for Arabidopsis. We accordingly focused our work on gene families common to Arabidopsis and prokaryotes. However, now that other plant genomes are pouring in (some 20 are available already and many more are in the pipeline) it is clear that almost all of the families we investigated have orthologs in other plants, making our work of immediate value in annotating other plant genomes. Furthermore, the rapid growth of microarray databases and other post-genomic resources for plants besides Arabidopsis (e.g. [64–66]) is providing many sources of association evidence to reinforce the approach that we have pioneered here.

Methods

Bioinformatics

The SEED genomic database and software suite [10], publicly available at http://theseed.uchicago.edu/ (see http://TheSEED.org for access to data relating to the SEED Project) was the main comparative genomics platform of this study. This database hosts all validated and proposed functional predictions developed in the course of this study for 78 genes families analysed in this study.

The SEED organizes genomic data in the form of subsystems (typically metabolic pathways or structural complexes) covering all organisms rather than on an organism-by-organism basis. Subsystems are developed and maintained by experts to capture the current status of knowledge of specific biological processes in model, well characterized organisms and to project this knowledge to other species via comparative genomics and metabolic reconstruction techniques [10]. Each subsystem includes a set of functionally related protein families (jointly encoding a specific pathway, process, or structural complex) across all available genomes (874 bacterial, 58 archaeal, and 29 eukaryotic complete and nearly complete genomic sequences as of July 2010). In SEED large homology-based protein families are broken into isofunctional subfamilies (‘functional roles’) based on genome context, functional context, phyletic profiling, shared regulatory sites, and other homology-independent clues. Association of each functional role with the corresponding subsystem(s) provides rich two-dimensional functional/phylogenetic context for each subfamily, leading to far more accurate annotations than the usual approach of annotating the genes within a single organism. Furthermore, the subsystem spreadsheet is used in SEED as a framework for integration of various types of functional data organized as gene attributes (e.g. gene clustering on a chromosome, expression array data, gene essentiality, etc.) and organism attributes (oxygen requirement, motility, pathogenicity, etc), which provide valuable non-homology based clues for functional predictions for uncharacterized genes.

All Subsystems created in this study, as well as over 1300 resident subsystems in SEED encoding all aspects of microbial physiology and metabolism are available on the public SEED server at http://theseed.uchicago.edu/FIG/SubsysEditor.cgi. They are regularly updated to accommodate newly sequenced bacterial genomes as well as novel experimental data and other relevant data as they become available.

Phylogenetic occurrence profiles were analysed using the Signature Genes tool on the NMPDR server (http://www.nmpdr.org/FIG/wiki/view.cgi/FIG/SigGenes). This tool identifies gene families that are common to a selected group of genomes, or those that differentiate one group of genomes from another. Annotations for paralog families were made using physical clustering when possible or by building phylogenetic trees using the ClustalW tool [67, 68] integrated in SEED or deriving specific motifs.

We also used the bioinformatic tools and resources at NCBI (http://www.ncbi.nlm.nih.gov) and KEGG (http://www.genome.jp/kegg) [69], BRENDA (http://www.brenda-enzymes.info/) [70], PHYRE (Protein Homology/analogY Recognition E ngine http://www.sbg.bio.ic.ac.uk/phyre/) [71], the Pfam database (http://pfam.sanger.ac.uk) [72], and specialized genomic resources and collections of functional genomic data for Arabidopsis, yeast, and various bacterial species, including: TAIR (The Arabidopsis Information Resource http://www.arabidopsis.org/) [21]; SGD (Saccharomyces Genome Database http://www.yeastgenome.org) [22]; MicrobesOnline (http://www.microbesonline.org/) [73]; EcoGene (http://ecogene.org/) [74]; EcoCyc (Encyclopedia of E. coli genes and metabolism (http://biocyc.org/ECOLI/) [75]; Cyanobase (http://genome.kazusa.or.jp/cyanobase) [76]; Pseudomonas genome database (http://www.pseudomonas.com/) [77]; and Rhodobase (http://rhodobase.org/index.php). Collections of Arabidopsis global expression and proteomics data with on-line tools for visualization and analysis: ATTED (http://www.atted.bio.titech.ac.jp/) [78]; Golm Transcriptome database (http://csbdb.mpimp-golm.mpg.de/csbdb/dbxp/ath/ath_xpmgq.html, [31–33]); PED, Plant Gene Expression Database (http://bioinfo.ucr.edu/projects/Unknowns/external/express.html); Genevestigator (https://www.genevestigator.com) [31–33]; PPDB, The Plant Proteome Data Base (http://ppdb.tc.cornell.edu/) [79]; PDB, The Protein Data Base (http://www.rcsb.org/pdb/home/home.do) [80].

Experimental validations

Methods for the three experimental validation vignettes described above are already, or soon will be, described in the authors’ publications.