Bacterial genomes are known to be dynamic, consisting of genes with different evolutionary histories. Some genes are evolutionarily conserved while others can be gained and lost in a lineage-specific fashion, and by horizontal gene transfer events. Prior studies on yeast and vertebrates have suggested that genes that are the most evolutionary conserved and most highly expressed evolve at the slowest rate 67. Therefore to investigate the effect of selection on these various gene categories, we classified E. coli genes according to their evolutionary conservation into Core, Non-core and HGT genes (see Methods). In this regard, we hypothesize that the cumulative effect of selection acting on these different gene categories would leave footprints in their sequence and gene expression characteristics.

It is known that horizontally transferred genes can be inserted into existing operons and thus contribute to the dynamic nature of the gene order and membership of these operons 89101112131415. Although a few studies have investigated the evolutionary stability and the conservation of gene order of operons 1617, the relative contribution of HGT on the evolutionary composition of operons remains unclear. In this regard, we aimed to explore the prevalence of HGT genes in operons by cataloguing the presence of operons consisting of Core, Non-core and HGT genes.

Another area that has been missing in the study of HGT events is the aspect of protein-protein interactions and cellular networks. A few studies have concentrated on the impact of horizontal gene transfer on metabolic networks 1819. Unfortunately very little is known about the effect of horizontal gene transfer on the global protein interaction networks in this aspect, mostly due to the lack of cellular interaction data in bacteria until recently.

It has been suggested that the scale-free properties of biological networks may in part be due to a model of preferential attachment by means of gene duplication, whereby new nodes preferentially attach to existing highly connected nodes. In networks that have evolved via preferential attachment, older nodes should have a higher average connectivity than younger nodes 20. In this regard, horizontal gene transfer can be considered as an additional biological mechanism to the existing model of preferential attachment. Although distinctly different, a model of network growth and expansion that involves gene duplication results in a duplicate protein copy with exact same or similar function, whereas a mechanism involving HGT may represent novel functions. In this regard, proteins encoded by HGT genes can be seen as competing with resident genes in establishing and gaining protein interactions.

We investigated both operons and protein interactions as a means of detecting successful incorporation of putative horizontally transferred genes in the E. coli genome. We explored the possibility that successful HGT genes would require integration at the level of operons to be expressed and integration at the network level to establish fitness benefits to the organism. We found horizontally transferred genes exhibit lower gene expressivity and evolve at faster evolutionary rates than evolutionarily conserved core genes. In addition, although proteins encoded by horizontally transferred genes have lower network connectivity, they preferentially attach to resident Core proteins rather than Non-core proteins within the protein interaction network. We conclude that a small proportion of the low connectivity proteins may have arisen from HGT events.

Genome sequences available for the various E. coli strains were downloaded from the NCBI (Escherichia coli K12 MG1655 – NC_000913; Escherichia coli O157H7 – NC_002695; Escherichia coli O157H7_EDL933 – NC_002655; Escherichia coli CFT073 – NC_004431; Escherichia coli UTI89 – NC_007946).

Our primary data set consisted of horizontal gene transfer events that were identified using a combination of the gene phylogeny and the pattern of gene presence and gene absence 15. This approach is similar to gene presence/absence analyses 2122.

For detection of horizontal gene transfer events, a total of 326 complete bacterial genome sequences divided into 8 bacterial clades were downloaded from MicrobesOnline database 23. Using each protein sequence contained within the E. coli K12 genome as query, BLASTP sequence similarity searches are conducted against all 326 bacterial proteomes. Subsequent BLAST sequence hits are further categorized into "BestN" hit categories with the Best0 category referring to the E. coli K12 gene itself. Each gene is assigned to a relative age category (i.e. clade) based on the BLASTP hit with the highest score. The method classifies each gene within the E. coli K12 genome as belonging to either (i) a set of horizontally transferred gene set (named HGT), (ii) a native gene set restricted to the E. coli lineage (named Native) or (iii) a gene set with no known sequence homologs (named ORFan). Thus the BLASTP scores gradually decrease in groups with increasing phylogenetic distance from E. coli K12.

Multiple sequence alignments based on protein sequences are then constructed using the MUSCLE sequence alignment software 24. Fast neighbour-joining trees 25 are then subsequently constructed for each protein sequence alignment. Genes that lack "close" homologs in consecutive groups of related bacteria are then confirmed using a quartet test available within the software package TREE-PUZZLE 26. To infer a horizontal gene transfer event; gene trees are compared with the MicrobesOnline specie tree (see above). If a strongly supported clade in the gene tree was present in disparate genomes, so that three or more deletion events would be required to explain the distribution of the subfamily on the species tree, then an HGT event was assigned.

In addition, we have included a comparison of horizontally transferred genes obtained by various HGT detection methods which comprised three surrogate (non-tree based) methods namely, HGT-DB 27, the method published by Mrazek and Karlin 28 and a support vector machine-based method (HGT_SVM) developed by Tsirigos and Rigoutsos 29 versus our data derived from a combined phylogenetic and gene presence/absence based method 15, both in terms of overall counts but also in terms of their distribution of Cluster of Orthologous (COG) functional categories (see Additional File 1).

On the overall, the method developed by Price et al. predicts more HGT genes in E. coli K12 than the three surrogate methods. It is known that base compositional differences between resident and invading genes are "ameliorated" over a few million years 30. Surrogate methods that use a compositional approach may preferentially detect recent horizontal gene transfer events and genes with atypical base compositions 21. Thus, a cross-phylum approach using phylogenetic tree based methods combined with gene phyletic profiles are more likely to detect ancient but also recent horizontal gene transfer events.

Our operational definition of a Core set of genes was meant to reflect the evolutionary retention of a set of common genes in all E. coli strains. In this regard, our Non-core set reflect genes that are found in at least one strain but not all strains, and HGT genes correspond to genes which are derived from putative recent horizontal gene transfer events. Thus, this distinction between Core and Non-core genes serves to illustrate the difference between a stable and invariant Core component and a variable Non-core component that is specific to E. coli strains. In this regard, the Non-core genes represent genes with a restricted phylogenetic distribution limited to one or more E. coli strains. These genes can be lost or gained in a strain-specific manner. Thus to ensure that there is no overlap between any evolutionary gene categories we have filtered the Core, Non-core and HGT gene sets to ensure a non-overlapping set of each gene category is maintained.

We derived an evolutionary Core set of 2158 E. coli genes based on the criteria of using phylogenetic gene conservation and genomic context (positional gene conservation). Starting with an all-vs-all protein sequence comparison consisting of the five E. coli genomes, we grouped E. coli K12 genes based on their phylogenetic gene conservation profiles within all five strains. To ensure a high quality Core gene set, we extracted and compared the chromosomal locations of all Core genes. It is known that genes which evolve vertically between closely related species can be divided into those that retain homologous chromosomal positions (positional orthologs) and those that do not 31. In addition, phylogenetic trees were constructed based on selected protein sequences to verify the phylogenetic relationship between the five E. coli strains.

Our Non-core gene set was obtained by post-process filtering the BLAST sequence comparison results of E. coli K12 genes which had BLAST hits in at least one or more E. coli genomes, but not present in all genomes. We also extracted and compared gene chromosomal locations of this gene set and constructed phylogenetic trees for further investigation. Since this gene set showed lower phylogenetic conservation, they were also positionally conserved to a lesser extent.

In addition, results from correspondence analysis of codon usage also revealed a distinction between our Core, Non-core and HGT gene categories (Additional File 2). The Core,Non-core and HGT gene lists can be found in Additional Files.

Data pertaining to Escherichia coli operons and transcriptional units were downloaded from RegulonDB version 5.7 3233. RegulonDB is a manually curated database that focuses on transcriptional regulation in E. coli with information extracted from literature as well as sequence databases such as GenBank. Its basic structural unit is the operon, which describes the elements and properties of transcriptional regulation. Thus in keeping with this definition, we refer to an operon as a poly-cistronic transcribed unit with its associated regulatory sites, whereas a regulon is defined as a group of operons controlled by a single regulator. As of RegulonDB version 5.7, 4570 E. coli genes have been annotated and organized into 2684 operons.

E. coli K12 MG1655 microarray gene expression data were downloaded from the NCBI GEO microarray database 34. We selected the GDS2600 data set, which closely approximates growth under normal conditions. This data set contains a time course which monitors the expression of 4405 E. coli genes using spotted cDNAs in stationary phase using LB media. Log2-transformed gene expression values were used and we excluded genes with missing data from the analysis. For each gene, mean gene expression values across time points were calculated and used for subsequent analysis.

For construction of the E. coli interaction network, we extracted the protein-protein interaction data from a recently published mass spectrometry study 35. We examined this data set carefully to confirm that it was not biased towards particular pathways or functional categories using the KEGG pathways and COG functional classification databases respectively. The whole analysis was also re-performed using the protein interaction data from Arifuzzaman et al. 36 (Additional Files 3 and 4).

Detection of orthologs was performed using a reciprocal best-hits approach as implemented in the RSD-algorithm 37. Multiple sequence alignments were constructed from protein sequences using the ClustalW package 38. Phylogenetic tree reconstructions were performed using the neighbour-joining method 39. Evolutionary substitution rates were estimated using the CODEML program available from the PAML package40. Network analyses were performed using algorithms implemented in the NetworkX package 41 and visualised using PAJEK 42. Statistical analyses were performed using the R-programming language environment.

To investigate the selective pressure acting on organizational units, we classified E. coli genes according to their evolutionary conservation into three categories, namely, (i). Core Set: a conserved core set of genes that exist in all E. coli strains. (ii). Non-core Set: genes that are found in at least one strain but not in all strains, and (iii). HGT Set: genes that are derived from putative recent horizontal gene transfer events after the divergence of E. coli and Salmonella. By delineating genes according to their evolutionary conservation, we can more clearly identify the evolutionary forces to which the various evolutionary classes of genes are subjected.

Direct measurements of E. coli gene expression were obtained from microarray gene expression experiments (see Methods). In addition, we have also used the codon adaptation index (CAI) as a proxy for gene expression data, which we referred as "gene expressivity" 43.

Figure 1 shows that the Core genes have higher CAI gene expressivity levels (Figure 1A) as well as log2 expression values (Figure 1B) than Non-core and HGT genes (t-test and Wilcox rank test, p-value < 0.001). This can be explained by the different evolutionary histories of these three groups of genes. The Core set of genes, being the oldest resident genes in the genome have thus had sufficient time to adapt and optimise their codon usage patterns, explaining the higher levels of gene expressivity; whereas the recent horizontally transferred genes may need an adaptation period during which their base composition and codon usage patterns may need to be optimised to their new resident genome.

Figure 2 shows that amongst the three categories of gene sets, the Core set of genes evolve at the lowest substitution rates (dN/dS) and HGT genes evolve at the fastest rates, using E. coli K12 as reference for comparison (Wilcoxon-Mann-Whitney test, p-value < 0.001). The high evolutionary rates observed for HGT genes may be explained by either one of the following two hypotheses: (i) result of reduced negative selection pressure, which enable the invading genes to be purged from the genome, or (ii) result of increased positive selection whereby HGT genes contribute to the phenotypic character of E. coli strains 14. Accordingly, it is thought that the strain-specific Non-core genes and HGT genes may contribute to the pathogenic character separating the enterohemorrhagic and uropathogenic from the benign E. coli K12 strain, therefore these genes are under positive selection pressure.

There is increasing evidence to suggest that the chromosomal gene order in organisms is not always random 44. It is known that proteins of linked genes evolve at comparable rates, and that natural selection may promote the conservation of linkage of co-expressed genes 45. Accordingly, genes in the same operon occur in close physical proximity and are often known to be co-transcribed as units. In addition, genes encoding subunits of protein complexes also need to be expressed at similar times.

To investigate the relative contributions of the various evolutionary gene categories on organizational structures, we surveyed both operons and the protein interaction network for their content of Core, Non-core and HGT genes. The Core set form a predominant portion of operons with 47% (2129 out of 4506 genes catalogued in RegulonDB version 5.7) of the operons consisting of Core genes, whereas 21% (948 out of 4506) of Non-core genes and 23% (1020 out of 4506) HGT genes, respectively, accounted for the remaining gene constituents of operons (Figure 3A). Similarly, proteins encoded by Core genes account for a 67.5% (852 out of 1262) of the protein interaction network as reported by Butland et al 46 whereas Non-core genes and HGT genes account for 14.1% (178 out of 1262) and 18.4% (232 out of 1262) respectively (Figure 3B).

The tendency of operons to be enriched in Core genes can be explained by a need to simplify regulation, since genes residing in operons known to be under control of the same promoter (Chi-squared test, p-value < 0.001). This may facilitate horizontal gene transfer by enabling genes to be inherited as a physical and functional cohesive group rather than separate individual genes. In regard to the protein interaction network, it is thought that the Core genes form the ancestral backbones of the protein interaction network to which new functionalities are added via protein nodes and thus strengthens a model by which pathways expand 47.

To understand the impact of higher order organization of genes (i.e. operons) and proteins (i.e. interaction complexes) on properties such as expression or evolution, we investigated the gene expressivity characteristics and evolutionary substitution rates of the various categories of gene sets. We found that Core genes in organizational clusters (both operons and protein interaction network or PIN) have higher gene expressivity (CAI) values (Figure 4A and 4B) and as well as log2 expression values (Figure 4C and 4D) relative to Non-core and HGT genes (t-test and Wilcox-test for both operons and PIN, p-value < 0.001). For the PIN, this trend was robust against removal of ribosomal proteins.

The overall trend from surveying operons and the protein interaction network indicates that Core genes tend to be found more often in organizational units such as operons and networks. The evolutionary composition may be the reason that highly clustered proteins in the protein interaction network display apparently high gene expressivity and low substitution rates.

We have analyzed and compared the distribution of the Cluster of Orthologous (COG) functional categories of the Core, Non-core and HGT genes within the E. coli K12 genome, protein interaction network and operons (Figures 5, 6 and 7). The various gene categories differ significantly in their COG distribution in the genome, the protein interaction network and operons (Scheirer-Ray-Hare test, p-value < 0.001, see Additional File 5).

In the overall gene comparison of the E. coli K12 Core, Non-core and HGT gene sets, the Core genes constituted the major evolutionary gene set present in all COG categories (Figure 5). The Non-core gene set in comparison to the HGT gene set was markedly abundant in the two COG categories: O (Posttranslational modification, protein turnover, chaperones) and T (Signal transduction mechanisms). The HGT gene set was more abundant than the Non-core gene set in the COG categories C (Energy production and conversion), F (Nucleotide transport and metabolism), G (Carbohydrate transport and metabolism), I (Lipid metabolism), K (Transcription) and V (Defense mechanisms). For the operons, the Core genes occur predominately in all COG functional categories, whereas the Non-core genes are over-represented in COG categories S (Function unknown) and U (Intracellular trafficking, Secretion, and vesicular transport) and the HGT genes are over-represented in comparison to the Non-core genes in COG functional categories C, E (Amino acid transport and metabolism), G, H (Coenzyme metabolism), R (General function prediction only) and V (Defence mechanisms) (Figure 6).

For the protein interaction network, the HGT genes are over-represented in COG functional categories most notably C, G, H, and V (Figure 7). A most notable example in this regard is the COG category V in which the HGT gene set within the E. coli protein interaction network has a significantly higher mean node degree than the Core and Non-core genes sets. The overall statistical difference in distribution of COG functional categories between the Core, Non-core and HGT gene sets therefore seems to argue against the notion of a Core-versus-Non-core or Core-versus-acquired gene category consisting of Non-core and HGT genes, but rather strengthens the notion of a distinct separate category for Non-core genes.

To investigate the mode and mechanism of integration of horizontally transferred genes into the E. coli protein-protein interaction network, we systemically investigated the network characteristics of proteins encoded by the various evolutionary categories of genes (Table 1). We found that proteins corresponding to the Core gene set represent the most highly connected protein nodes, which have an average connectivity of 11.0 interactors (Chi-squared test, p-value < 0.05). In contrast, Non-core proteins and proteins encoded by HGT genes have on average lower connectivities of 4.0 and 3.0 interactors respectively. This is consistent with our hypothesis that Core genes being the most highly conserved genes have resided in the genome for much longer, and thus had more opportunities to evolve interactions. The result of the network analysis is consistent with this theory.

We also analyzed two additional network properties: betweenness centrality and clustering coefficient (Table 1). Betweenness centrality characterizes how essential a node is in maintaining communication between each pair of nodes in a network 48. Depending on its position within the network, removal of a node can have very different effects on the connectivity, topology and flux of the network. Some nodes can be removed without any harmful effect, while others separate a connected network into disconnected sub-graphs. Betweeness centrality is a measure devised to describe the fraction of shortest paths going through a given node, with high values indicating that a node can reach many other nodes. Removal of nodes with high centrality will make it difficult to reach from one node to another, thus lengthen the path between nodes. The clustering coefficient describes the local transitivity in a network, with two nodes having a common neighbour in a network being more likely to be neighbours 49.

Table 1 shows that the HGT genes have lower betweenness centrality than the Core and Non-core genes, which suggests that they are less important in cellular communications. Interestingly the Non-core genes have higher betweeness centrality than the Core genes, the implication of which need to be further explored. On the other hand, Core genes have the highest clustering coefficients, with any two Core genes having a common neighbour being more likely to be neighbours of each other. The results in Table 1 indicate the HGT genes are the least important in maintaining the overall connectivity of the protein interaction network, in other words they are more likely to be peripheral nodes.

Our analysis of the distribution of COG functional categories of the Core, Non-core and HGT nodes within the E. coli protein interaction network reveal that the Core genes are the most abundant and cover all the major COG functional categories in comparison to the Non-core and HGT gene sets (Figure 7). Although, the Non-core and HGT genes show similar COG distribution profiles within the protein interaction network, differences exist in COG categories C, G, H and V in which the HGT genes are relatively more abundant than Non-core genes. A most notable result in this regard is the COG defense category (V) in which the HGT gene set within the E. coli protein interaction network has a significantly higher mean node degree than the Core and Non-core genes set.

We further investigated the evolutionary profiles of the interaction partners in the network. Table 2 shows that about 74% of all the interactions are between a pair of Core genes, 11.2 % of the interactions are between a Core gene and a Non-core gene. In other words, in total about 85% of the interactions involve at least one Core gene. Among the interactions involving HGT genes, a large percentage (89%) was between a HGT genes and a Core gene, while interactions between Non-core and HGT genes only account for 1%. This is surprising since the ratio between Core genes and Non-core genes is only ~5:1, much smaller than the 9:1 ratio (89%: 10%) that we observed in the network. This discrepancy in ratio implies that an HGT gene have a higher propensity to establish interaction with a Core gene than with a Non-core gene. Indeed, the proportions of HGT-Core interactions are higher than expected by chance (Chi-squared test, p-value < 0.001).

Such a model of preferential attachment has previously been proposed to explain the growth of protein interaction networks in S. cerevisiae 205051 and was also suggested recently for E. coli 52; however it has remained mostly unproven since it was difficult to trace back the evolution history of protein networks. Along this line, the HGT genes in E. coli offer a unique opportunity to test this theory since these genes are indeed "new genes" that were only added to the network after the HGT event ~100 millions ago 5. Our observation provided direct evidence and support for this model, which has not been reported previously.