Genes are grouped by their presence or absence in a selection of organisms (Table 1). In gene group A (Table 2), we expect to find the most recent additions to the genome through contributions from both lateral transfer and duplication. Because these genes are present only in the subject organism, it is probable that lateral transfer occurred after the last divergence. In the case of relatively recent divergence, such as for instance Salmonella typhimurium and S. typhi, the total number of genes (N) in category A is relatively small, whereas organisms which are farther apart (for example, Bacillus subtilis and B. halodurans) have larger numbers of unique genes. However, there is probably no constant rate at which genomes acquire genes. This is illustrated by the comparison between the two E. coli strains, where the A group is almost twice as large in O157:H7 than in K12. This suggests that strain O157:H7 is more active than K12 in gene acquisition either mechanistically or because of added selective pressure. Other pairs of organisms, such as E. coli K12 and S. typhimurium, have more similar patterns of gene acquisition. Imports in category B (Table 3) are on average older than in category A, and many genes are likely to have been lost in the outgroup genomes as a result of different lifestyles and requirements. Genes in category D are present in all four organisms in the group, and are probably not recently transferred genes.

In previous work 1819 we have developed a compositional measure (cT²; see Materials and methods) to find genes with atypical dinucleotide frequencies. This measure is used in combination with the gene categories. Through all categories and in all comparisons, the number of atypical genes is the highest in groups that are associated with recent lateral transfer. The proportion of deviant genes (Pdev; P(cT² > 38)) are consistently higher in categories A and B than in D (Table 4), which would support the notion that there are many candidates for lateral transfer in groups A and B. In B. subtilis and B. halodurans, which appear to have diverged long ago, Pdev in categories A and B is low, although still higher than in D. The numbers of deviant genes in category A are 51 and 40 in B. subtilis and B. halodurans respectively, similar in number to those of, say, E. coli or S. typhi. Thus the rates of import may be similar in magnitude and the low Pdev scores could be due to category A containing a large number of older genes that have either been ameliorated in Bacillus or lost in Clostridium. The older the A group, the lower Pdev will be, until it approaches the Pdev of category D. The comparison between the Staphylococcus aureus strains mu50 and mw2 shows an intermediate pattern, with a relatively high Pdev in categories A and B compared to category D. This would suggest that there is no great difference between the proteobacteria and the Bacillus/Clostridium group in lateral transfer patterns.

The distributions of guanine and cytosine at the third codon position (GC₃) in categories A and D are summarized in Figure 1 and Table 5. Category A is generally 'flatter' than the GC₃ distribution of category D. In some genomes, particularly in the Escherichia/Salmonella clade, GC₃ is almost uniformly distributed. Given that genes in category A are good candidates for lateral transfer, a flat distribution would suggest that the GC₃ range of imports is wide. Accordingly, the number of donor genomes must be large enough to incorporate a broad spectrum of gene GC₃ content. In the Bacillus/Clostridium clade, the GC₃ pattern of the A groups resembles that of the D groups to varying degrees. Particularly B. halodurans appears to have a narrower GC₃ range of potential imports, which could be explained by one of three reasons. First, B. subtilis and B. halodurans have been diverging for so long that the A group could be dominated by old imports that have had time to ameliorate to the genomic GC₃ distribution. Second, it is possible that the number of donor organisms is lower in the halophilic environment of B. halodurans. If so, there could be less variation in the GC₃ range of imports. Third, genes in category A could be primarily genes that have been lost in B. subtilis but present in the ancestor, although this explanation also means that the same genes have been lost in Clostridium.

To focus on the genes that were likely to be more recent imports in Clostridium acetobutylicum, B. subtilis and B. halodurans, which have long divergence times, we studied the GC₃ distribution of the subset of A genes with high cT² values. Genes with cT² scores grater than 30 were compared to the rest of the set (Figure 2). The numbers of genes with scores over 30 were 91, 97 and 78 for C. acetobutylicum, B. subtilis and B. halodurans respectively. The GC₃ distribution of the high-cT² genes in C. acetobutylicum differs neither from the rest of the genes in the A group, nor from those in the D group (Figure 1h). This could either be due to a lower rate of import (all imports have ameliorated), or a consequence of an already low GC content. B. subtilis and B. halodurans, on the other hand, have a wider GC₃ range in the subsets of high-cT² genes relative to the rest of the A genes. B. subtilis has a significant overrepresentation of high-cT² genes in the GC-poor terminus region, which could be an effect of either a higher recombination rate in this area 9 or local nucleotide variations, or both.

In B. subtilis group D, there are a number of genes with high cT² and low GC₃ around the ori region. These are most likely not imports, as they have defined (and often ribosomal) functions and a high codon adaptation index (CAI) 20, indicating that there is translational selection on these genes. Genes in group A around the terminus region also have high CAI, but this is probably due to the fact that low GC₃ coincides with the choice of major codons in B. subtilis 21.

Comparing GC₃ skews of category D in Figure 1, we note that all Escherichia/Salmonella clade genomes are skewed towards low GC₃. This relationship could be due to one or both of the following explanations: there may be local, although small, regions that are biased towards lower GC₃; or it may suggest that there are a number of pervasive low-GC₃ imports into the Escherichia/Salmonella clade.

Both E. coli and B. subtilis have distinct regions around the terminus with low GC content 1922. This regional bias around the terminus is not a general feature. The genome of B. halodurans also has a low-GC region around the terminus, although not as distinct as that of B. subtilis. The regional bias of S. aureus mu50 is intermediate between B. halodurans and B. subtilis, whereas Clostridium has no low-GC region around the terminus.

The fate of a recent duplication is selection or redundancy followed by deletion, with the vast majority of recent duplications being deleted 3. In several organisms it was found that the persistence of neutral material is extremely low. Few neutral paralogs survive long enough to undergo any significant mutational degradation. In all cases considered we find that the total number of duplications in the A groups are significantly overrepresented compared to category D (Table 6).

As a result of both stoichiometric and expression-level constraints on gene duplication, we expect the number of recent duplications among genes that are either highly expressed or involved in complex protein interactions to be low. While the interaction level can be hard to assess, expression can be estimated from the codon bias 20. In E. coli K12, we see that the paralogs in category A are all hypothetical proteins with low codon bias, CAI < 0.3. In category D, 33 of 47 recent duplications are hypothetical or putative genes, with a low average CAI score. Two exceptions are the tuf copies. In E. coli O157:H7, hypothetical genes dominate the duplications in category A and none of the paralogs has a known function. The duplications in category D of O157:H7 are similar to those in category D of K12, with mostly hypothetical genes. S. typhimurium has only five recent duplications in category A, and all are putative. In category D, we find 15 copies of a number of ABC transporter-related genes, along with putative genes and two tuf copies. C. acetobutylicum also follows this pattern, with mostly hypothetical genes, along with ABC transporters. In B. subtilis, all 16 duplications in category A are hypothetical. Ten of sixteen have CAI scores lower than average. Genes are better defined in category D, with only 6 of 10 duplicated genes having CAI scores lower than average. S. aureus mu50 has few recent paralogs in category D, and only one is annotated as hypothetical. In category A, four out of six are hypothetical. If duplication is mainly dependent on function, then it may in part explain the overrepresentation of duplications in category A for almost all organisms. This category should contain the highest proportion of weakly functional or nonfunctional genes, which are less disruptive to amplify.

In general, duplicated genes seem more likely to have high cT² scores. This correlation is dominated by the D category (Table 4) where there is a higher proportion of deviant genes among the duplications in all but three cases. This would suggest a general link between atypical sequences and duplication, which may be due to unusual DNA sequences such as transposon-like structures. Even though annotated transposons have been removed from the datasets, some such sequences may still persist. Furthermore, the overrepresentation could also be due to higher rates of recombination in certain chromosomal regions. Finally, it is possible that the paralogs were imported as a cluster - the duplication therefore taking place before the transfer. This is not supported by the chromosomal positions of paralogs in category A of the paired organisms in Table 6, suggesting that duplication is likely to have taken place after transfer. Another possibility could be a multiple and probably simultaneous import of the same gene from a single source. It would be difficult to tell this event apart from an indigenous duplication event.

The context bias (cT²) has been shown to be a useful measure to identify recently imported genes 1819. This is further corroborated in this study where the gene categories that are expected to contain more recently imported genes always have a larger fraction of deviant genes (high cT²). The connection is strengthened further by the observation that the deviant genes have a broader and more uniform GC₃ distribution in keeping with the expectation that recent imports have not had time to ameliorate to the genomic bias. Thus, the number of recent imports in each genome can be assumed to be related to the number of deviant genes in category A. The absolute numbers of deviant genes in the A and B groups are very similar for all genomes considered (Tables 2,3). A possible exception is S. aureus, which seems to have a lower number of deviant genes in category A. This suggests that the intensity of lateral transfer is roughly of the same magnitude in all cases.

The general flatness of the distributions of GC₃ content in category A may support the notion that recently laterally transferred genes dominate in this group. This is the kind of distribution we would expect under the assumption that the multitude of potential donors have a wide range of GC₃ content. The distribution of GC₃ content in category D is almost always more peaked and has more weight in its tails, reflecting a non-Gaussian probability distribution. Among the organisms with approximately 50% GC₃ content, there is a slight skew to low GC₃. Furthermore, in the two organisms with low genomic GC₃ content, both categories A and D are skewed slightly to higher GC₃.

Whereas recent imports generally seem to have broad, nearly uniform, GC₃ distributions, there seem to be few genes with high GC₃ content. This is particularly evident in the genomes with low overall GC₃ content. Thus, the recent imports in the Escherichia/Salmonella clade with intermediate overall GC₃ content have broad tails to low GC₃, but the low-GC₃ genomes have no corresponding tails to high GC₃ (Figure 1). Either the potential donors for these genomes are all of low GC₃ content or low-GC₃ genes are more easily exchanged. E. coli has regions in the genome with low GC₃ content 1922 as does B. subtilis. Possibly such regions are more likely to accept imports of similar GC content.

In the Escherichia/Salmonella clade, the chromosomal position of category A genes was studied. For E. coli K12 and S. typhimurium, there is a tendency to have a high number of A genes at gene position 2,000 to 3,000. E. coli O157:H7 has a peak at 1,000 to 2,000. In general, there is no overrepresentation of potential laterally transferred genes in the immediate vicinity of the low-GC terminus region in contrast to previous proposals 1018.

Genes that are likely to have been imported into the organism from an outside source are also more likely to be duplicated. This is the overall impression from studying bacteria both from the proteobacteria group and from the Bacillus/Clostridium clade, although the relation between cT² scores and propensity to duplicate is weaker in the latter group. This observed bias may be due to the following:

(1) Genes that have been imported into the genome may by their nature be more mobile or 'promiscuous'.

(2) Imported genes that are required in a new niche may be needed in larger amounts until control of expression levels has become established.

(3) Duplications are more likely to be retained or occur in genes that are poorly optimized. Thus, amplification is more likely to succeed in category A since there is a lower risk of counter-selection and a greater chance of discovering a weak but novel function.

(4) Genes may be imported in multiple copies into a genome. Even though this is not an effect of the duplication mechanism of the new host, the multiple imports still act as if they were actual paralogs, with the same restrictions. This explanation would be largely indistinguishable from the others.

Although all genes or open reading frames (ORFs) annotated as being transposon- or phage-related have been removed from this analysis, the possibility always remains that some such sequences remain that are as yet unrecognized. Therefore, explanation (1) cannot fully be ruled out. However, we believe that explanation (3) is more pertinent from observations in category D, where duplication occurs more often among genes with poorly understood function, and also from previous observations 4 where we find that highly expressed genes are less likely to be duplicated. This suggests that amplification of a gene product with a weak function is generally less disruptive and costly than amplification of a stronger gene product. If this is the case, the bias for duplications in category A need not be explained by areas of high recombination or increased mobility, but simply by virtue of a larger proportion of poorly optimized genes and gene functions among the potential imports.

As virtually all duplicated recent imports are annotated as hypothetical or putative, it is not easy to suggest functions. It is possible, and even likely, that some recent imports have strong gene functions that are required in a new niche and could be duplicated to provide an appropriate gene-dosage effect. However, recent imports that have identifiable functions are rarely duplicated. Thus it seems reasonable to suggest that most of the duplicated genes among the recent imports have poorly optimized functions and that the apparent bias for duplication is due to selection for gene dosage. Nevertheless, we expect that most of these duplicated imports are also transient, as are most other imports and other duplications. The main point is that neutral and near-neutral genes have a limited persistence in the genome and that the pool of weakly functional or nonfunctional genes is dominated by the recent imports. This is where new functions seem most likely to be discovered.

If every duplication event has roughly the same chance of being selected by amplifying a gene product, then the contribution of lateral transfer to gene innovation can be estimated. Here we assume that the observed duplications are essentially neutral or weakly selected, and that deleterious amplifications are quickly purged from the population. In E. coli K12 categories A and B, when compared to both E. coli O157:H7 and S. typhimurium, there are roughly half as many duplication events as in category D. In the corresponding S. typhimurium comparisons, a quarter of all duplications are in category A or B. For the proteobacteria, E. coli O157:H7 tops the list, with almost half of its recent duplications in category A or B. In the Bacillus/Clostridium group, more than half of all duplications are in these categories.

As a conservative estimate, under the assumption that novel gene functions arise through modifying paralogs that are retained through an amplification of primarily weak functions, we propose that at least a quarter of all gene-innovation events are a direct consequence of lateral transfer. The estimate is conservative as there may be duplications in D that are 'stuck' in the amplified function, such as elongation factor and, possibly, ABC transporter genes. In any case, lateral transfer in combination with duplication may have a considerable contribution to gene innovation. This contribution is not always apparent when focusing solely on the transfer of established gene functions.

A number of genomes were downloaded from GenBank, and abbreviated as follows: Escherichia coli K12 (ecol) 23; E. coli O157:H7 EDL 933 (edl) 24; Shigella flexneri (sflex) 25; Salmonella typhimurium (stym) 26; S. typhi (styp) 27; Bacillus subtilis (bsub) 28; B. halodurans (bhal) 29; Clostridium perfringens (cper) 30; C. acetobutylicum (cace) 31; Staphylococcus aureus strains mu50 (samu) 32 and mw2 (samw) 33; and Pseudomonas aeruginosa (paer) 34. Furthermore, contigs from Klebsiella pneumoniae (klebs), which is nearing completion, were downloaded from the Genome Sequencing Center 35.

ORFs annotated as related to either transposons or phage sequences were removed, because occurrences of these sequences may prove confounding. Only genes longer than 400 nucleotides were considered, because of small-sample effects when calculating cT² scores (see below).

The bacteria studied here fall loosely into two clades: the Bacillus/Clostridium and the Salmonella/Escherichia clades. Although it is often difficult to produce good estimates of bacterial divergence, 16s RNA trees (data not shown) suggest that the bacteria in the Salmonella/Escherichia clade diverged much later than the bacteria in the Bacillus/Clostridium clade, with the possible exception of the S. aureus strains, which may be comparable to the Salmonella/Escherichia clade.

As a measure of lateral transfer into various genomes, we use a phylogenetic approach among a group of related bacteria. For instance, as the divergence of Salmonella and Escherichia is more recent than the divergence of their ancestor to Klebsiella and Pseudomonas, we assume that a gene present only in Escherichia is a likely lateral transfer (gene import) that occurred after the divergence of Salmonella and Escherichia, or a transfer that occurred before divergence but was subsequently lost in Salmonella. This category of genes contains likely candidates for lateral transfer. As support, we use a method developed to detect atypical nucleotide context biases (cT² 1819). As defined, this context bias is independent of the nucleotide usage of the gene considered and is expected to reflect a mutation bias that deviates from the average of the genome. The values of cT² are low when genes appear typical and high when they appear atypical. When a foreign gene is first introduced, it may appear atypical because it has been adapted to a different mutation bias. In time it is expected to ameliorate and approach the context bias of its new host.

Organisms are organized into groups of four. The first group comprises the subject organism (for example, E. coli K12) and a close relative (S. typhimurium). These are then compared with two organisms that are more distant (K. pneumoniae and P. aeruginosa). There are four categories of genes that are considered: A, B, C and D, according to Table 1 (compare 18). Category A is assumed to include very recent gene imports, entering the genome after the last divergence. Category B is presumed to also include imports, although to a lesser extent than category A, as there is the increased probability of genes being lost in the outlier genomes. Category C, which is only briefly addressed in this study, is likely to be composed of genes that have been lost in the paired genome and retained in the others. Category D consists mainly of genes that have been present for a long time and would include imports only if they are very pervasive, that is, they have been lost and reimported or imported several times.

As a threshold for determining presence or absence, we used the blastx package 36, which gives six-frame translations of nucleotide sequence to amino acid sequence. An E-value of less than 10^-10 indicates presence in the studied genome. The value was set to a relatively low level of significance because of the highly varying degrees of divergence of the organisms studied. Furthermore, we chose the same threshold for all organisms, as any variable value would imply that the relative ages of divergence were known.

As a simple measure of recent duplications, genomes were searched against themselves using blastn 36. Genes that could fulfill both a more stringent E-value of lower than 10^-20 and a nucleotide identity of ≥ 95% were considered to be recent duplications. There is always the possibility that such genes are not recent paralogs, but highly conserved, through gene conversion for instance. However, if these paralogs are so highly conserved, they would more likely be found in categories C or D.