It would be extremely useful if transcriptional regulatory architecture of bacteria could be predicted from the sequence of their DNA. Attempts to do so generally involve extrapolation from well-studied species such as E. coli, in cases where the regulators and target genes are orthologous and a binding site is conserved in the target promoter. However, such extrapolation relies on several implicit assumptions (see Introduction) that have not been well tested experimentally. We used the Lrp regulon as a model to carry out tests of these assumptions.

The pronounced conservation among Lrp orthologs in enteric bacteria was first noted over a decade ago 77, and the large number of subsequently-determined genome sequences has not altered that pattern. The Lrp ortholog in P. mirabilis differs from that in E. coli by only 4/164 amino acids (98% identity), while Lrp from V. cholerae shows 92% identity to E. coli Lrp. Importantly, none of the changes observed in P. mirabilis and V. cholerae occur in the helix-turn-helix motif responsible for DNA sequence recognition (Fig. 1, cartoon representation and boxed region), defined via mutation of the E. coli lrp gene and x-ray crystallography of an archaeal ortholog 7879 and recently of the E. coli protein itself 80. Similarly, the Lrp orthologs of these bacteria are completely conserved for amino acids implicated in coregulator recognition (boxed amino acids in lower panel of Fig. 1).

Lrp orthologs from another γ-proteobacterial family, the Pasteurellaceae (including Haemophilus influenzae and Pasteurella multocida), are much more divergent from E. coli (Fig. 1), with the differences specifically including the helix-turn-helix motif. It is interesting that the Pasteurellaceae appear to form an outgroup with respect to Lrp, as the Lrp differences are far more pronounced than their overall relationship to neighboring bacterial genera would suggest (see Fig. 2 in 81, and Fig. S1 in Additional file 1). This, and the fact that in H. influenzae Lrp controls only a small number of genes 82, led us to exclude the Pasteurellaceae from this study.

The first test of functional conservation is that if Lrp is having similar broad effects on gene expression in E. coli, P. mirabilis, and V. cholerae, then one would expect to see similar effects of a lrp null mutation on their growth. Fig. 2A shows the results of growth experiments for wild-type (WT) and lrp strains of these three species grown in MOPS glucose minimally supplemented medium ("MOPS glucose") or MOPS glucose defined rich medium ("MOPS rich") [both media including nicotinate as required by the P. mirabilis strains, and pantothenate and thiamine as required by the P. mirabilis Δlrp strain (see Methods)]. The plot shows the WT specific growth rate on the x-axis, and the rate for the lrp strain on the y-axis; thus where lrp mutation has no effect on growth rate the points fall on the diagonal line. In MOPS rich medium (closed symbols), lrp mutation had little effect on growth of any of the three species. However in MOPS glucose medium, P. mirabilis stands out as having a substantial growth rate decrease when lrp is mutated (193 min vs. 66 min doubling time for the WT strain). This might represent a lrp-dependent partial auxotrophy, in addition to the lrp-dependent requirements for pantothenate and thiamine that were satisfied by the medium. However that may be, it is clear that the lrp mutation has differential effects in these three species.

We cloned the lrp genes from P. mirabilis and V. cholerae downstream of PlacUV5 in the low-copy vector pCC1 (Epicentre); a consensus E. coli Shine-Dalgarno sequence 83 was introduced as well. We also cloned the lrp gene from E. coli O157:H7 (which is identical to that of E. coli K-12 at the amino acid level). Thus the three lrp alleles had identical expression sequences. The effect of the lrp allele on P. mirabilis growth in minimal medium was fully complemented by supplying the cloned P. mirabilis lrp gene on a plasmid, and was mostly but incompletely complemented by the lrp genes from E. coli and V. cholerae (Fig. 2B).

We next tested the ability of Lrp orthologs to complement a complex phenotype other than growth. P. mirabilis undergoes differentiation to form hyperflagellated swarmer cells >20-fold longer than nonswarmer cells, and yields concentric rings of growth on agar 5659; Fig. 3A). Like growth rate, swarming is sensitive to a variety of factors and thus also provides a sensitive indication of the cell's physiological status 59. It has been shown by others 84 that a lrp mutation abolishes swarming in P. mirabilis (Fig. 3B). We show here that the lrp orthologs from both E. coli and V. cholerae complement a P. mirabilis lrp mutant, and restore the complex swarming behavior (Fig. 3, panels E and F). The P. mirabilis lrp gene does not regenerate the exact WT swarming pattern when supplied in trans (Fig. 3A, D); this difference may reflect replacement of the native lrp expression sequences with PlacUV5 on the plasmid. However, despite the fact that the three plasmid-borne lrp alleles had identical expression sequences, they gave consistent differences in the P. mirabilis swarming patterns as reflected in growth ring measurements from triplicate experiments (Fig. 3G). In this experiment, only the regulator (Lrp) was varied; the P. mirabilis target genes and promoters/binding sites are identical between strains. Thus the significant differences in swarming (e.g., comparing the outer growth rings of strains complemented by lrp from Vibrio and Proteus in Fig. 3G) indicate functional differences in the Lrp proteins, some of which may be amplified by indirect effects.

If closely-related Lrp orthologs are fully functionally conserved, as would be necessary for regulatory extrapolation, then the orthologous WT lrp alleles should cross complement to generate the same pattern of target gene regulation. We therefore tested the ability of heterologous Lrp proteins to properly regulate three Lrp-responsive genes in a lrp null mutant of E. coli. The lrp-bearing plasmids described above, which produce Lrp independently of the normal growth-dependent control associated with the native lrp gene 8586, were used to transform E. coli strains containing reporter fusions to known Lrp target genes. Western blot analysis of a constant amount of total protein, probed with a polyclonal anti-Lrp antiserum 87, revealed comparable accumulation of the various Lrp orthologs (Fig. 4A). The strain carrying the V. cholerae lrp gene appeared to accumulate ~75% as much Lrp protein as the E. coli control (Fig. 4A and data not shown), though this is a minimal estimate because the antiserum was generated against E. coli Lrp (92% identical to V. cholerae Lrp at the amino acid level). It is interesting that the Vibrio Lrp migrated slightly faster than the other Lrp proteins. Its calculated mass (18.79 kDa) is slightly smaller than that of the other two proteins (18.89 and 18.92), and its calculated pI is less basic (7.7 vs. 8.9 for both of the others); other factors can also affect migration of individual polypeptides in SDS acrylamide gels 88.

To begin testing the functional equivalency of the different Lrp orthologs, we co-transformed strain BE10.2 (ΔlacZ, lrp-Tn10) with a vector containing an E. coli Plrp-lacZ fusion and the various lrp-bearing plasmids (or vector control). Lrp directly represses its own promoter in E. coli, and this occurs whether or not leucine is present 89. Strains were grown in MOPS glucose, and the specific activity of β-galactosidase was determined and plotted against culture density to more quantitatively assess its level and to assure that the cultures were in balanced growth. Compared to the vector control, V. cholerae Lrp repressed Plrp-lacZ to the same extent as E. coli Lrp (~2 fold), while P. mirabilis Lrp (98% identical to E. coli Lrp) repressed about twice that much (Fig 4B).

Transcriptional activation is generally a more demanding process than repression, in the sense that the activator has not only to bind the DNA correctly but also (in most cases) to make productive contacts with RNA polymerase 9091. Strain BE3780 contains a chromosomal operon fusion to the gene for glutamate synthase (gltB-lacZ) in the E. coli BE10 background (ΔlacZ, lrp-Tn10) 74. Lrp directly and strongly activates gltBD transcription in E. coli 7492, in a process that also requires the global regulator IHF 93 and is antagonized by Crp and ArgR 94. We found that, relative to the vector control, activation of gltB transcription by P. mirabilis or V. cholerae Lrp was essentially indistinguishable from that of the E. coli lrp positive control (Fig. 4C).

Conserved function among regulators depends not only on DNA-binding (and RNA poymerase-contacting) properties, but in some cases also on responses to small molecule coregulators. This provides the basis for a third test of assumptions in regulatory extrapolation. Lrp transduces metabolic signals in the form of amino acid pool levels, in particular the amino acids L-leucine and L-alanine 64. The livKHGMF operon is one of two high-affinity branched chain amino acid transport systems in E. coli 95. The livK gene is repressed by Lrp when exogenous leucine is present 7096, and activated when leucine is not in the medium 66. Thus livK is a particularly sensitive indicator of the responses of Lrp orthologs to leucine. The amino acid residues previously demonstrated to be involved in leucine binding in E. coli Lrp 6497 are completely conserved in the P. mirabilis and V. cholerae orthologs (boxes in lower half of Fig. 1).

We prepared a PlivK-lacZ operon fusion (pRLIV2), and co-transformed it into E. coli BE10.2 (ΔlacZ, lrp-Tn10) together with plasmids bearing the heterologous lrp alleles under the control of PlacUV5 (Table 1). These strains were grown in MOPS glucose media containing isoleucine (I, Ile) and valine (V, Val), with or without leucine (L, Leu). Leu was not used alone as it can lead to starvation for Ile, via feedback inhibition of L-threonine deaminase 98. β-galactosidase activity was determined as described above.

In general the Lrp proteins from P. mirabilis and V. cholerae yielded livK regulatory patterns similar to that of the E. coli control, showing activation in the absence of leucine and repression in its presence (Fig 4D and 4G). However both P. mirabilis and V. cholerae Lrp gave threefold greater repression than E. coli Lrp (Fig 4G). We also looked at the effects of adding ILV on the regulation of E. coli Plrp and PgltB. Addition of ILV interfered to a moderate extent with the repression of Plrp by all three Lrp orthologs (Fig 4E), but the differences observed in the MOPS glucose cultures were maintained, with substantially greater repression by P. mirabilis Lrp. Also as expected from previous studies 7492, activation of PgltB by E. coli Lrp was moderately reduced in the presence of ILV; the two heterologous Lrp orthologs gave patterns essentially indistinguishable from that of E. coli Lrp (Fig 4F).

The functional conservation of Lrp orthologs was more broadly assessed by using microarrays to analyze gene regulation, when lrp alleles from E. coli, P. mirabilis, or V. cholerae (all fused to the same expression sequences) were used to complement the lrp null mutation in E. coli K-12. In these experiments, the only variable is the Lrp itself – the target genes and promoter binding sites are identical. The data for all significantly-affected genes are available [see Additional file 2].

Only 16% of the genes differentially regulated by one or more of the three Lrp orthologs were regulated in common by all three of them (Fig. 5), but this group included recognized Lrp-controlled genes such as gltBD, ilvG1, lysU, and osmC 677499100101 (these results are available in spreadsheet form [see Additional file 2]). Roughly half of the genes regulated by E. coli Lrp under these conditions, directly or indirectly, were regulated by either of the other two orthologs (51% by P. mirabilis Lrp, and 42% by V. cholerae Lrp). Similar proportions were seen when activated and repressed genes were considered separately (not shown). [Note: "repression" and "activation" as used here include both direct and indirect effects.]

A potential concern with these experiments is that some apparent differences between the gene sets responsive to the three Lrp orthologs might be artifactual. For example, if the typical significantly-affected gene varies twofold, but this change is only counted as being significant in half of the genes that truly change twofold, then 1/4 of genes would appear to be unaffected by a Lrp ortholog, even if there were no real functional differences between the Lrp proteins. We addressed this concern in two ways.

First, the NULL hypothesis for observed overlaps under our experimental design is that two different Lrp proteins should produce identical (100% overlapping) sets of differentially-expressed genes. We are not aware of any distribution test statistic that can be used to evaluate this hypothesis (if the NULL hypothesis had been that two sets are random, then we could have used either hypergeometric distribution or Fisher's exact test). Instead we performed a bootstrap procedure to enumerate all possible outcomes of overlaps, given the data. If our actual NULL hypothesis were true, two complementation comparisons (e.g. E. coli Lrp vs. V. cholerae Lrp) would be indistinguishable. To test this, we generated a NULL distribution of overlaps by randomly exchanging columns of ratio values between two sets.

Each set contains three ratio values (R1–3), as a result of the experiments having been carried out in triplicate. For each set we determined differentially-expressed genes at α = 0.05 using a t-test, and then determined the size of the intersection. Next, we "permuted" the sets as shown in Fig. 6A. There are 20 combinations in which three ratios can be selected out of six. Thus C36xC36 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaem4qam0aa0baaSqaaiabiodaZaqaaiabiAda2aaakiabdIha4jabdoeadnaaDaaaleaacqaIZaWmaeaacqaI2aGnaaaaaa@33AC@ pairs of sets were t-tested, "differentially-expressed" genes were identified, and the number of such genes in the intersection was determined. The number of times the size of the overlap is greater than that observed in the real comparison, divided by 400 (the number of "permutations") indicates the significance of the observed overlap – effectively a simulated p-value. Table 2 shows the results of such simulations. The overlap between Eco and Pmi is significantly smaller than chance, and the small size of the overlap between Eco and Vch is very close to being statistically significant. Overall, this analysis indicates that the sizes of the intersections are substantially smaller than what would be expected by chance, and that the different Lrps are having significantly different effects on expression of the E. coli genome.

Second, we used intra-set comparisons to qualitatively evaluate the observed overlaps. We compared the fractions of genes, differentially expressed by both Lrps in a pairwise comparison, at different ratio cut-offs. We also compared lrp vs. WT (chromosomal lrp⁺), and lrp(ParaBAD-LrpK12) vs. lrp. This allowed us to compare the overlaps between LrpK12 on two plasmids and between a plasmid and the WT chromosomal gene. The overlap fraction was calculated as the number of regulated genes in both of two sets, divided by the total number of unique genes in both sets. Fig. 6B illustrates that, as expected, the overlaps are consistently greater between identical complementation sets than between heterologous sets. Thus the microarray results reveal statistically significant functional differences between the Lrp orthologs.

Predictions of regulatory connections between a regulator and a target gene are useful in themselves, but substantial understanding of a cell's gene regulation also requires knowing the sign and strength of the regulation. Accordingly, we next examined whether the subset of genes that was regulated by both E. coli Lrp and one of the other orthologs showed a similar magnitude of regulation by each ortholog (Fig. 7). The effects of E. coli lrp (pEcoLrp) on gene expression are shown on the x-axis in every panel. Column A shows the set of genes for which transcript levels gave statistically significant decreases when the lrp mutation was complemented. Column A thus represents genes that are repressed (directly or indirectly) by E. coli Lrp. Column B includes gene showing direct or indirect activation by E. coli Lrp. Column C shows the set of 57 genes recognized in RegulonDB 7172 as being directly controlled by Lrp, whether the control is positive or negative. This set includes genes that are only controlled by Lrp under growth conditions that differ from those used by us 6669, so the cluster of genes showing little or no effect of Lrp is not surprising.

We used the slope of a least-squares fit between two ratios as a measure of overall regulatory concordance (Fig. 7). One ratio is the expression level of genes in the E. coli Δlrp mutant over the level in that strain complemented by a plasmid carrying the E. coli lrp gene (on the x-axis in all panels). The second ratio (y-axis) is the expression in E. coli Δlrp over that in the same strain complemented by a test plasmid (vector control, or lrp from P. mirabilis or V. cholerae). Full complementation relative to that by pEcoLrp would yield a slope of 1.0 for the linear fit. For each column, the top row indicates the effects of "complementing" the Δlrp allele with the vector alone (pCC1). Not surprisingly, the negative control of "complementing" the lrp mutation with the vector gives slopes of <0.2 for all three gene sets. As a positive control, we carried out a similar analysis comparing on the one hand the E. coli lrp mutant to the mutant complemented with plasmid-borne E. coli lrp [i.e., lrp/lrp(pLrpK12)], and on the other hand the mutant to the WT lrp⁺ strain (lrp/WT). The resulting slope between these datasets was within error of 1.0 (0.94 ± 0.12; not shown), indicating that native Lrp – even with the heterologous promoter and translation initiator – fully complements the effect of the chromosomal lrp mutation under the growth conditions used.

Supplying the lrp allele from P. mirabilis (pPmiLrp, bottom row) gave substantial, though not full, complementation of the gene expression pattern, with slopes ranging from 0.81 to 0.93. The lrp allele from V. cholerae (pVchLrp), which is more divergent from that of E. coli than is that of P. mirabilis (Fig. 1), yielded lower overall regulatory complementation – the slopes range from 0.49 to 0.69. Thus small changes in the Lrp sequence, outside of the fully-conserved helix-turn-helix motif, have substantial effects on the magnitude of Lrp-dependent effects on gene expression.

We find that, if anything, this analysis underestimates the differences between effects of the different Lrp orthologs. Since the power of inference cannot be adequately estimated in microarray experiments, we fixed the false discovery rate (FDR) at 5% without setting a minimum limit on fold change. However when we added a fold-change limit to the fixed FDR the differences between regression slopes were increased (not shown). In these experiments, genes were assigned to columns A or B (Fig. 7) based on having significant responses (positive or negative) to E. coli Lrp. The measurements on which this selection was based include noise, so even without a real difference in Lrp function the data might regress back towards no change in the complementation experiments. However, since the correlations used only genes that were significantly affected in the same direction in both sets in each pair, no systematic degradation of the signal is expected.

We evaluated the significance of differences between the slopes in Fig. 7 as follows. To avoid making assumptions about the nature of the distribution under the NULL hypothesis, we bootstrapped the slopes and estimated 95% confidence bands for each slope. The coordinates of each point in the correlations are estimates of corresponding ratios of transcript abundances, obtained as means of ratios observed in independent biological experiments. Therefore, we estimated the width of slopes of the regression lines by permuting the series, choosing N points (corresponding to the number of genes used in the correlation) at random from one of the three ratios. For example, for gene 1 we can take the ratio from biological replicate 1, for gene 2 from replicate 3, etc. N-member series were permuted in this way 1000 times, and the spread of 95% confidence intervals around the mean was calculated relative to the regression coefficients from Fig. 7. We demonstrated that 95% confidence intervals do not overlap between regression lines of interest. This supports the conclusion that differences between Lrp orthologs can explain the different amounts of variance observed in the various complemented strains.

If orthologous regulatory proteins are to generate the same expression patterns for orthologous target genes, having the same intrinsic properties (DNA binding specificity and equivalent interactions with small molecules and with other proteins) would be neccesary but not sufficient. The orthologs would also have to share extrinsic properties, including accumulation to similar levels under various growth conditions. This provides the basis for a fourth test of implicit assumptions underlying regulatory extrapolation. Potential differences in Lrp levels were minimized, in the complementation experiments described above, by providing each lrp ortholog with a common promoter and translation initiation region. However E. coli growing exponentially in a minimal glucose medium accumulates three- to four-fold more Lrp than in rich medium 8586. To determine whether this pattern of Lrp accumulation is conserved, we used western blot analysis to measure the levels of Lrp throughout a batch growth cycle in E. coli, P. mirabilis and V. cholerae grown in MOPS glucose and MOPS defined rich media (supplemented as described in Methods).

Our results confirm the earlier studies of E. coli, in that we saw several-fold higher Lrp levels when cells were grown in MOPS glucose than when they were grown in MOPS rich medium (Fig. 8, compare panels D and J). When grown in MOPS glucose, P. mirabilis (Fig. 8E) and V. cholerae (Fig. 8F) produced levels of Lrp similar to that in E. coli (Fig. 8D). Furthermore, all three species showed severalfold lower Lrp levels in rich than in minimal medium. There was, however, one substantial difference among the cultures. In the rich medium, P. mirabilis (Fig. 8K) produced up to twice as much Lrp as E. coli or V. cholerae, with levels highest in stationary phase.

These differences between E. coli and P. mirabilis, in Lrp protein levels, could well have substantial regulatory significance but needed confirmation. Samples were taken from parallel cultures of E. coli and P. mirabilis during early logarithmic, mid logarithmic, late logarithmic and stationary phases in defined rich medium. Equal amounts of total protein were resolved side-by-side via SDS PAGE. The results of the subsequent western blot (Fig. 8M) confirm that P. mirabilis produces substantially more Lrp protein throughout the growth phases, with the greatest difference (roughly twofold) seen in stationary phase. Thus Lrp provides an example of related bacteria with nearly-identical regulator proteins producing significantly different amounts of those regulators.

As regulator levels are often inferred from microarray measurements of mRNA levels, we determined whether the level of lrp mRNA also varies with the growth medium in all three organisms (Fig. 9A). At least four QRT-PCR determinations from each of two independent experiments were averaged to generate each plotted value. In E. coli, lrp mRNA levels are profoundly lower in rich than in minimal medium, irrespective of growth phase, and similar to what was seen for Lrp protein (compare black bars to Fig. 8 panels D and J). Also like the protein results, P. mirabilis lrp mRNA rises in stationary phase (compare gray bars to Fig. 8 panels E and K), but the mRNA shows this rise only in rich medium. The V. cholerae lrp mRNA results resemble the protein data in that there is no significant growth phase effect, but differ in that the mRNA shows no growth medium effect (compare white bars to Fig. 8 panels F and L).

We also measured the levels of lrp mRNA for all three organisms during log-phase growth, using a more highly-quantitative method, and the results are consistent with the protein data (Fig. 9B–D). We used a sensitive dilution-response approach that makes use of the fact that our three species-specific pairs of QRT-PCR primers for lrp amplify with the same efficiency but are completely specific for their respective template DNAs (data not shown). A standard amount of reference E. coli RNA (from a mid-log phase culture in MOPS glucose plus nicotinate) was mixed with varying amounts of test RNA (from cultures grown in either the MOPS glucose or MOPS rich medium). The various mixes were reverse transcribed and the resultant cDNA was used as template for simultaneous amplification with the three primer pairs, with real-time fluorescence monitoring (see Methods). Where the proportion of total mRNA as lrp mRNA is equal to that in the reference sample, the slope should be 1.0 (dotted lines in Fig. 9B–D). For E. coli-derived cDNA, the resulting slope is about 0.75 (MOPS glucose culture; Fig. 9B) or 0.4 (MOPS rich culture), indicating lower lrp mRNA levels in rich than in minimal medium and consistent with the medium-dependent effect on Lrp protein levels shown in Fig. 8. V. cholerae (Fig. 9D) gave a pattern similar to that of E. coli, though with less of a medium-dependent effect. However – also consistent with the protein data – P. mirabilis had substantially more lrp cDNA as a proportion of total cDNA than did E. coli, with slopes of about 2 (Fig. 9C).

We next tested an explicit assumption of regulatory extrapolation, by determining if the expression of orthologs of E. coli Lrp target genes are Lrp-responsive in their native hosts. We measured the mRNA levels from orthologs of two genes previously shown to be Lrp responsive in E. coli 69: adhE and gltB. These orthologs were chosen based on percent identity to the E. coli protein, and presence of at least one predicted Lrp-binding site using PRODORIC 102 (Fig. 10).

Wild-type and lrp strain pairs of E. coli, P. mirabilis, and V. cholerae were grown in MOPS glucose defined rich medium. Samples were taken in early logarithmic phase (OD_{600 nm} = 0.3), and early stationary phase (1 h after the culture OD vs. time semilogarithmic plot diverged from linearity). Real-time RT-PCR analysis was used to determine the relative levels of adhE, gltB, and (as a control) recA mRNAs. The experiment was performed in triplicate and the relative levels of mRNA were determined using the standard curve method 103 and by normalizing to recA. There is no effect of Lrp on recA expression, at least in E. coli and V. cholerae under our conditions 69 and N. Dolganov, pers. commun.).

Finally we tested whether promoter regions from orthologous genes, where the E. coli gene is Lrp-controlled, are regulated by Lrp in heterologous hosts. This was done by preparing lacZ operon fusions to a set of ortholog promoters cloned from E. coli, P. mirabilis, and V. cholerae, and then introducing each of these fusions into both the WT and lrp strains of all three species. Relative LacZ activity was measured using the approach shown in Fig. 4 (determining the slope of a LacZ activity vs. culture density plot). These experiments are reciprocal to those shown in Fig. 4, where heterologous lrp alleles had been moved into an E. coli background to test for control of E. coli target genes. However the two sets of experiments share the feature that the target promoters/binding sites being assessed are identical (the set of E. coli promoters in the case of Fig. 4, and the promoter for one of the three lrp orthologs in this experiment).

Regulatory extrapolation relies on a tacit assumption that regulatory proteins with high amino sequence identity are functionally equivalent. We took closely-related Lrp orthologs (all >92% identity) from three species, gave them identical expression sequences for transcription and translation, put them into the same very low-copy vector (pCC1), and introduced them into the same E. coli K-12 lrp and P. mirabilis lrp backgrounds. Of the 164 aa in all three Lrp orthologs, P. mirabilis Lrp differs from E. coli Lrp at only four positions, while the V. cholerae and E. coli orthologs differ at 12; none of these differences affects the known DNA-binding helix-turn-helix or the coregulator binding sites (Fig.1).

The Lrps exhibited similar overall behavior, supporting extrapolation in general, particularly where the only concern is whether a regulatory link exists at all irrespective of its sign or strength. However there were significant functional distinctions between the tested Lrp orthologs. In E. coli, the native Plrp (fused to lacZ) was repressed equivalently by E. coli and V. cholerae Lrp, but about twice as much by P. mirabilis Lrp (Fig. 4B). These differences were magnified in the presence of leucine, where P. mirabilis Lrp was unique in showing virtually no effect (Fig. 4E). In contrast, the three Lrp orthologs gave equivalent activation of PgltB (Figs. 4C, F). PlivK, which in E. coli is activated by Lrp in the absence of leucine and repressed in its presence, was regulated equally by all three Lrp orthologs with one exception: in the presence of leucine, the E. coli Lrp represses less than the others (Fig. 4D, G).

We used microarray analysis to more globally assess the ability of orthologous Lrp proteins to properly control the E. coli K-12 Lrp regulon. Our results confirmed that minor changes in the Lrp amino acid sequence had substantial effects on the targets (Fig. 5) and magnitude (Fig. 7) of Lrp effects. Of the genes whose expression was significantly changed by E. coli Lrp, over a third were not significantly affected by either of the other Lrp orthologs. Conversely, about half of the genes significantly affected by P. mirabilis Lrp, and about a third of those affected by V. cholerae Lrp, were also affected by E. coli Lrp in the E. coli background. Looking only at the subset of genes that were significantly affected by all three Lrps, we found that whether we examined genes significantly repressed by Lrp, activated by Lrp (in both cases including both direct and indirect effects), or directly controlled by Lrp in E. coli, the results were consistent. Specifically, the transcriptional effects of P. mirabilis Lrp were most similar to those of E. coli Lrp (correlation of 80–92%), followed by V. cholerae (correlation of 67–70%), with vector alone showing no statistically significant similarity (0–20%), and the chromosomal vs. plasmid-borne E. coli lrp comparison giving 94% concordance. It is important that in these experiments, the only variable is the Lrp ortholog present – the target genes and binding sites are identical between strains (as they are all based on the same E. coli host).

Another assessment of Lrp functionality involved P. mirabilis swarming over a solid surface. Swarming is a complex phenomenon; for example, in Salmonella about a third of all genes showed swarming-associated changes in expression 114. For the purposes of the present study, swarming thus represents a sensitive indicator of Lrp action. In a P. mirabilis background, we found that all three Lrp orthologs restored swarming, but gave repeatable differences in the resulting swarming patterns (Fig. 3).

We are currently exploring the molecular bases for these intrinsic differences in the three Lrp orthologs. In theory, a combination of differences in DNA specificity (due to changes outside the helix-turn-helix motif), in cooperativity, in response to coregulatory amino acids, or to interactions with RNA polymerase or other regulatory proteins 115 could be involved. For the purposes of this report, however, the main point is simply that such differences exist even between proteins that are 98% identical (the P. mirabilis and E. coli Lrps).

We also examined the native regulation of lrp in E. coli, P. mirabilis, and V. cholerae. Regulatory extrapolation relies on a second tacit assumption: that levels of the conserved regulator are similar in the organisms being compared, and change similarly in response to growth conditions. Target gene regulation is, not surprisingly, affected by the level of the regulatory protein; this is specifically true for Lrp 6692116.

Lrp protein levels in all three species were reduced in rich medium relative to glucose minimal medium (Fig. 8). For two of the three species, lrp mRNA levels are also lower in rich medium (Fig. 9). However we found that, during growth in defined rich medium (especially at higher cell densities), P. mirabilis levels of Lrp protein and lrp mRNA were about double those in E. coli or V. cholerae (Figs. 8, 9). For the purposes of this study, the important point is that Lrp levels differ significantly between the species, so that sequence analysis of the Lrp open reading frame and target gene promoter is not sufficient to predict expression patterns of the target gene. This regulatory variation is not an idiosyncracy of Lrp; for example, similar species-specific variation in regulation, among Enterobacteriaceae, has also been reported for the global regulator Fis 117, though as with Lrp the most basic patterns of expression are conserved 118.

A third tacit assumption underlying regulatory extrapolation is the reciprocal of the one described immediately above: that orthologous target genes moved into a common background will be regulated in the same way by orthologous regulators. We prepared lacZ fusions to both PgltB and Plrp promoters from E. coli, P. mirabilis, and V. cholerae, and introduced them into the lrp and lrp⁺ strain pairs for all three species in all combinations.

Once again the assumption is supported in general – most Plrp combinations are unaffected or repressed by Lrp (Fig. 11D), while all but one of the PgltB combinations are unaffected or activated by Lrp (Fig. 11E). [Note: repression and activation by Lrp have been demonstrated for these target genes in E. coli, but have not been proven to occur in the other backgrounds (where the effects might be indirect), and we use these terms for brevity.]

However the assumption is not supported by the specifics – e.g., the E. coli PgltB is well-expressed and Lrp-activated in all backgrounds, while the P. mirabilis PgltB ranges from nonexpression to Lrp-activated expression in the different backgrounds (Fig. 11). It is particularly interesting that, unique to the V. cholerae Plrp in the V. cholerae background, Lrp activated rather than repressing the promoter. There are some intriguing and distinctive sequence characteristics of the V. cholerae Plrp that may explain this behavior, and we are investigating these further. However V. cholerae Plrp gave low expression in the V. cholerae background even when Lrp was present, while this same promoter gave much higher (and Lrp-responsive) expression in the E. coli and P. mirabilis hosts. This, and the fact that the E. coli and P. mirabilis Plrp fusions were both well expressed in the Vibrio background, suggests that V. cholerae negatively regulates its Plrp via some Vibrio-specific factor, and that Lrp may in this case act as an anti-repressor.

Comparisons of independent microarray studies on distinct platforms are problematic 119120, and this report has the benefit of direct comparison using the same experimental and statistical methods. Nonetheless, our results are supported by earlier analyses of the somewhat differing effects of gene disruptions in the global regulators H-NS 121122123124, IHF 125126, and Fis 127128 in the closely-related genera Escherichia and Salmonella.

One possible interpretation of these results is that global regulators are more likely to have greater recognition plasticity than local regulators, in which case a study of Lrp may represent something of a "worst case scenario" for regulatory extrapolation despite its remarkable conservation (Figs. 1, S1). It has been suggested that global regulators bind a large number of sites with a wide range of affinities, affecting chromosome superhelical density and providing a continuous "analog" regulation, in contrast to the more "digital" regulation by more specific regulators 129. It is also true, however, that more local and sequence-specific regulators (such as LexA) show considerable range in their in vivo DNA binding 130. There are certainly local regulators that bind unique sites with extreme specificity 131, though the value of predicting their regulatory roles across species is correspondingly limited.

The bacterial strains used for this study are listed in Table 1. In all cases cells were grown in baffled flasks shaken at 37°C. Morpholinopropane sulfonic acid (MOPS) glucose minimal medium, and MOPS-based defined rich medium 132 were purchased from Teknova (Hollister, CA). In experiments comparing E. coli and V. cholerae with P. mirabilis, media for all strains were supplemented with 0.01 mM nicotinic acid, which is required for the growth of Proteus mirabilis 56 and of the lrp mutant of Vibrio cholerae (REL, unpublished observation). When lrp mutants were part of an experiment, minimal media also contained 0.01 mM each of pantothenate and thiamine, which we found to be additional requirements of the P. mirabilis lrp mutant, and in some cases 0.1 mM L-cysteine with 0.2 mM L-methionine which were not required but improved growth of this mutant (REL, unpublished observation). For PlivK-lacZ analyses, additional amino acids were used at the following final concentrations: 10 mM L-leucine, 0.4 mM L-isoleucine and 0.4 mM L-valine. Antibiotics were used, where indicated, as follows: 100 μg ampicillin/ml, 15 μg chloramphenicol/ml, 100 μg kanamycin/ml, and 10 μg tetracycline/ml.

The lrp alleles are as follows. For E. coli and V. cholerae, all but the first six and last six codons of the lrp ORF were replaced by the gene for chloramphenicol acetyltransferase (cat) (our unpublished result; N. Dolganov and G. Schoolnik, unpublished result), with confirmation by PCR amplification and sequencing. Some experiments made use of strains carrying an E. coli lrp::Tn10 allele (lrp-35, 67). The P. mirabilis allele is a lrp::miniTn5 disruptant 84, provided by G. Fraser). The E. coli MG1655 lrp mutant has the entire lrp ORF replaced by the gene for kanamycin, and was constructed using λred recombinase gene replacement system 133. Other strain information is in Table 1.

Overnight cultures in MOPS glucose or MOPS rich medium were inoculated from (respectively) M9 glucose or LB agar plates containing 0.01 mM nicotinic acid, and grown to early stationary phase. These cultures were then used to inoculate fresh media (1:32). OD_{600 nm} was measured following sample dilution as needed to maintain OD within the range of 0.08–0.3.

Samples for real-time RT-PCR analysis were isolated at the indicated times by removing an equal number of cells (estimated from culture density) from the flask and immediately adding it to two volumes of RNA stabilization buffer (RNA Protect Bacteria Reagent, Qiagen, Valencia, CA). This prevents the rapid changes in mRNA content that otherwise occur when bacteria are harvested. Samples were mixed, left at room temperature for 10 min, and stored at 4°C for no more than 5 days.

Samples for microarray analysis were isolated at an OD₆₀₀ of ~0.4, at which point 20 ml of culture was mixed with 2.5 ml of ice-cold 5% water-saturated phenol (pH < 7.0) in ethanol 134. After 10 min on ice, cells were pelleted, supernatant was removed, and pellets were frozen in liquid nitrogen and stored at -80°C, if necessary.

For RT-PCR experiments total RNA was isolated using the RNeasy miniprep kit (Qiagen) using their protocol with an added sonication step. Briefly, cells in the stabilization buffer were harvested by centrifuging at 4°C for 15 min at 5,000 rpm. Supernatants were removed and the pellet was resuspended in 1× TE buffer containing lysozyme (400 μg/mL). Lysis buffer was added and the cells were sonicated 3× for 15 s in a cup horn attachment to enhance lysis. Following ethanol precipitation, RNA was bound to the column provided, washed and eluted. To eliminate DNA, the RNA was treated with RQ1 RNAse-free DNAse (Promega, Madison, WI) as directed. cDNA was synthesized using total RNA as template, random hexamers (Invitrogen, Carlsbad, CA), and ImPromII reverse transcriptase (Promega). The random primers were annealed at 25°C for 5 min, and the first strand was then extended at 42°C for 1 h. The reverse transcriptase was inactivated by heating to 70°C for 10 min. cDNA samples were stored at -20°C.

For microarray experiments total RNA was extracted by the hot phenol-chloroform method 135, and treated with DNase I in the presence of RNase inhibitor for subsequent labelling by reverse transcription with Cy3-dUTP and Cy5-dUTP fluorescent dyes (Amersham, Little Chalfont, United Kingdom). The RNeasy miniprep kit (see above) was also used in some cases.

Primer sets (Integrated DNA Technologies, Coralville, IA) were designed for the adhE, gltB, lrp and recA genes for each strain (Table S1). Before each new experiment dilutions of cDNAs were tested to determine the concentration that gave maximally-efficient amplification, and to determine the efficiency for each primer set (23). Cycle threshold (C_T) values were determined by Roche Lightcycler detection of SYBR green fluorescence. Melting curve (Roche Lightcycler software) and agarose gel analyses were used to confirm the formation of specific products, which ranged in size from 192–202 bp. The standard curve method was used to determine relative amounts of mRNA and levels were normalized to recA 103.

RNA was extracted from mid-log cells grown in MOPS glucose or MOPS rich media (plus nicotinate). RNA from the various samples was quantitated spectrophotometrically, diluted such that all samples had the same total RNA concentration, and then mixed 1:0, 1:1 and 1:7 with a standard level of RNA taken from E. coli grown in MOPS glucose (e.g., 1 μl E. coli MOPS glucose RNA mixed with 0, 1, or 7 μl P. mirabilis RNA). The RNA mixtures were then reverse transcribed (see above), and RT-PCR was performed using a 1:1:1 mixture of the lrp primer sets specific to each organism. Actual C_T values were then plotted against the C_T values expected if all original samples had the same proportion of lrp mRNA to total RNA. The resulting slopes indicate the fraction of lrp-specific cDNA relative to that in the reference E. coli sample.

For each sample equal volumes of cells were centrifuged at 16,000 – g for two min. The supernatants were removed and the cell pellets were stored at -80°C until analysis. Pellets were resuspended in 1× SDS buffer (Novagen, Madison, WI). Cells were lysed by heating to 98°C for ten min, and total protein concentrations were determined using the RC/DC kit and protocol (BioRad, Hercules, CA). Equal amounts of protein were loaded on a 12% acrylamide SDS gel and electrophoresed at 110 V in 1× tris-glycine buffer. Proteins were then electroblotted to polyvinylidene difluoride (PVDF) membranes at 30 V for 1 h using the Xcell blot apparatus (Invitrogen). Proteins were detected by fluorescence using the ECL-plus Western Blotting Detection System (GE Health Sciences, Piscataway, NJ) per the manufacturer's protocol, with a 1:125 dilution of rabbit anti-Lrp polyclonal serum (gift of Dr. Joseph Calvo 87), and a 1:25,000 dilution of horseradish peroxidase-conjugated goat anti-rabbit IgG (gift of Dr. Darren Sledjeski). Protein bands were visualized on a Storm 840I phosphorimager (Molecular Dynamics, now GE Healthcare). Densitometric analysis of Lrp bands was performed using the Molecular Dynamics software, and the amount of Lrp in each sample was determined by comparison to a standard curve from purified Lrp dilutions included on each gel.

Starting with freshly-streaked single colonies, cultures (supplemented with Ile, Leu, Val and thiamine as described above) were aerobically grown overnight at 37°C and then diluted 20-fold into 20 ml of fresh medium. Recombinant cultures were propagated in the presence of chloramphenicol, and growth was monitored via OD_{600 nm}. Cultures were maintained in exponential growth for at least 10 generations by dilution.

Relative mRNA abundances between the lrp mutant and the same strain carrying a lrp gene on plasmid pCC1 (or carrying only the vector) were determined, using at least three biological replicates. Each replicate culture was grown on a different day, and inoculated with a mix of 2–3 average-sized colonies less than a day old. This analysis employed E. coli K-12 whole-genome DNA microarrays including 99% of all annotated open reading frames and the stable RNA genes. Slide preparation, reverse transcription with the Cy-dyes, hybridization, and image scanning were performed as previously described 135. The fluorescent probes were hybridized to an array at 65°C for 6 h. Intensities in both channels were smoothed using the Lowess method 136. Some biological replicate samples were split into technical replicates, on which dye-swap analyses were conducted. Known Lrp targets were taken from RegulonDB 7172, and are listed below.

aidB aroA b2659 dadA dadX fimA fimC fimD fimE

fimF fimG fimH fimI gabP gabT gcvH gcvP gcvT

gltB gltD gltF ilvA ilvD ilvE ilvG_1 ilvG_2 ilvH

ilvI ilvL ilvM kbl livF livG livH livJ livK

livM lrp lysU malT micF ompC ompF oppA oppB

oppC oppD oppF osmC osmY sdaA serA serC stpA

tdh yeiL ygaF

Dye- and array-specific noise was removed using the analysis of variance (ANOVA) error model 137. In pair-wise comparisons, differentially expressed genes were identified at an estimated false discovery rate of less than 5% using the two-class T-test in the SAM package 138. The NULL hypothesis was that gene-specific intensities in two classes have indistinguishable means.

Strains were grown to exponential phase in glucose minimal MOPS medium (Teknova). Samples were taken at 20 and 30-min intervals throughout the growth period. Levels of β-galactosidase were determined by o-nitrophenyl-β-D-galactoside (ONPG) hydrolysis 139. β-galactosidase levels were plotted against culture absorbance, and points were fitted via linear regression. The resulting slope yields the β-galactosidase activity.

The lrp genes (translational start to stop) from E. coli O157:H7, P. mirabilis HI4320 and V. cholerae El tor A1552 were PCR amplified from chromosomal DNA using Pfx DNA polymerase (Invitrogen). The upstream PCR primers (Table S1) contained a consensus E. coli ribosome binding site. Fragments were gel purified and cloned into the low-copy pCC1 blunt cloning vector (Epicentre), and transformed into E. coli EPI300 per the manufacturer's protocol. As a vector control, an irrelevant ~1360 bp DNA fragment (kanamycin resistance cassette provided by the manufacturer as a ligation control) was inserted into pCC1. Transformants were selected using chloramphenicol and sequence-confirmed. The recombinants pECLRP, pPMLRP and pVCLRP (Table 1) were isolated using Qiagen miniprep columns. The purified plasmids were then electroporated into E. coli BE3780 (Table 1) using a BioRad E. coli gene pulser and protocol. For experiments with P. mirabilis and V. cholerae Δlrp strains, which are already chloramphenicol resistant, these plasmids were digested with BsmI to remove the cat gene, and we inserted a kanamycin resistance gene PCR amplified from pACYC177.

The promoter regions of the lrp and gltB geneswere PCR amplified from E. coli O157:H7, Proteus mirabilis HI4320 and Vibrio cholerae El tor type N16969 chromosomal DNA using gene specific primers (Table S1) and Pfx DNA polymerase (Invitrogen). The PCR products were digested with BamHI and SalI and ligated into pBH403, which is a derivative of pKK232-8 and contains a promoterless lacZ gene between two bidirectional transcription terminators. The recombinant plasmids (Table 1) were electroporated into E. coli BE10.2 and PS2209;Proteus mirabilis U6450 and U6450Δlrp; and Vibrio cholerae El tor strain A1552 and A1552Δlrp.