Immobilization of Escherichia coli RNA Polymerase and Location of Binding Sites by Use of Chromatin Immunoprecipitation and Microarrays

The genome-wide location of RNA polymerase binding sites wasdetermined in Escherichia coli using chromatin immunoprecipitationand microarrays (chIP-chip). Cross-linked chromatin was isolatedin triplicate from rifampin-treated cells, and DNA bound toRNA polymerase was precipitated with an antibody specific forthe ß' subunit. The DNA was amplified and hybridizedto "tiled" oligonucleotide microarrays representing the wholegenome at 25-bp resolution. A total of 1,139 binding sites weredetected and evaluated by comparison to gene expression datafrom identical conditions and to 961 promoters previously identifiedby established methods. Of the detected binding sites, 418 werelocated within 1,000 bp of a known promoter, leaving 721 previouslyunknown RNA polymerase binding sites. Within 200 bp, we wereable to detect 51% (189/368) of the known ${sigma}$ 70-specific promotersoccurring upstream of an expressed open reading frame and 74%(273/368) within 1,000 bp. Conversely, many known promoterswere not detected by chIP-chip, leading to an estimated 26%negative-detection rate. Most of the detected binding sitescould be associated with expressed transcription units, but299 binding sites occurred near inactive transcription units.This map of RNA polymerase binding sites represents a foundationfor studies of transcription factors in E. coli and an importantevaluation of the chIP-chip technique.

INTRODUCTION

Top

Introduction

RNA polymerase (RNAP) in Escherichia coli is a key factor ingene expression and catalyzes the transcription of DNA to mRNAfor all genes (reviewed in references 33 and 37). The core enzymeis composed of four subunits: two ${alpha}$ , one ß, one ß',and one ${omega}$ . Core RNAP becomes transcriptionally active holoenzymewith the addition of a ${sigma}$ factor (Fig. 1A). E. coli possessesseven interchangeable ${sigma}$ factors, each with specificity for differentpromoters. Sigma factors function as global regulators of geneexpression and mediate the transcriptional response to conditionsunder which a large number of genes need to be turned on oroff, such as stationary phase or heat shock. Except for ${sigma}$ ⁵⁴,sigma factors do not bind to DNA except as part of the holoenzymecomplex (9).

View larger version (30K):
[in this window]
[in a new window]
FIG. 1. RNAP transcription cycle and the effect of rifampin. (A) Mechanisms of transcription initiation and rifampin binding. RNAP (blue) binds DNA to form a closed complex with affinities that differ among promoters by at least a factor of 100. The closed complex isomerizes to an open complex at rates that differ among promoters by at least a factor of 100. The open complex is stable at some promoters; at others it collapses back to the closed complex unless it binds initiating nucleoside triphosphate substrates. The competition between abortive initiation and promoter escape (initiation) is highly variable among promoters. Rifampin can bind to all species of RNAP except the elongation complex and traps RNAP in open complexes by forcing the release of 2- to 3-nt abortive transcripts. (B) Effect of rifampin on growth and RNAP location. E. coli cells were cultured, and the absorbance was determined at given time intervals. Rifampin (150 µg/ml) was added at 200 min to one culture ( ${blacksquare}$ ), while the other was allowed to continue growing as a control ( ${blacklozenge}$ ). After 20 min, the culture with rifampin added stopped growing, while the control continued in exponential growth. The inset shows the results of quantitative PCR using primers designed to amplify in the promoter, 16S, or 23S subunit genes from all seven rRNA operons. Cells were treated with rifampin for 0, 20, or 40 min followed by immunoprecipitation using antibody for RNAP ß' core subunit. The results of PCR performed on genomic DNA before immunoprecipitation are shown on the right end of the gel image, labeled "input." The quantified relative intensity of each 16S and 23S band is given below the gel image.

To map promoters in bacteria, we sought a way to force RNAPto reside only at promoters so that identifying DNA fragmentsbound to RNAP in vivo would report promoter locations. A varietyof small-molecule inhibitors of RNA polymerase were evaluatedfor the immobilization of RNAP, and rifampin was found to workbest (M. Raffaelle, E. Kanin, J. Vogt, R. R. Burgess, A. Z.Ansari, unpublished data). The antibiotic rifampin inhibitsbacterial growth by binding the ß subunit of RNAPjust upstream of the active site, blocking the synthesis ofRNAs longer than 2 to 3 nucleotides (nt) (11) (Fig. 1A). Rifampinhas no effect on RNAP promoter binding to form closed complexesor on open complex formation (30) and has no effect on RNAPin vitro when added after elongating RNAP has cleared the promoter(12, 38). Upon addition of rifampin to a growing culture, therifampin stops growth without killing cells by diffusion acrossbacterial membranes and tight binding to RNAP not engaged intranscription. RNAP holoenzyme complexed with rifampin can stillbind to promoters but is trapped in an abortive cycle and unableto extend RNAs beyond 2 to 3 nt (Fig. 1A). However, RNAP moleculesin elongation complexes with RNA and DNA are resistant to rifampinbinding. Once elongating RNAPs terminate transcription and releaseRNA and DNA, they become susceptible to rifampin binding, whichtraps them in newly formed open complexes.

A promoter is a DNA sequence to which RNAP binds and initiatesthe transcription of RNA. Knowledge of promoter locations isthe first step in the elucidation of the transcriptional regulatorynetwork. It allows the identification of which genes are cotranscribedand, from the DNA sequence, identification of regulatory motifsassociated with different regulators. It also provides a basisfor the interpretation of the binding site locations of varioustranscription factors. Promoters have been identified on anindividual basis by primer extension and DNA footprinting studies,by which 961 promoters have been identified (www.cifn.unam.mx/Computational_Genomics/regulondb/)(36). In addition, DNA binding motifs have been used to predictthe locations of 4,641 promoters throughout the genome (21,40). New experimental evidence of promoter locations may validateor refute these predictions and will assist in understandingthe function of uncharacterized genes and operons.

A powerful method combining chromatin immunoprecipitation andmicroarray analysis called chIP-chip (or genome-wide locationanalysis) has been developed for the global identification oftranscription factor binding sites (10, 23, 34). In E. coliit has been used to identify the locations of MelR binding sites(16), and in yeast it has been used to identify the targetsof over 200 transcription factors (18). Furthermore, chIP-chipwas used to identify the targets of yeast RNA polymerase III(Pol III) and to investigate the roles of the accessory factorsTFIIIB and TFIIIC (19, 31, 35).

These previous studies used microarrays spotted with PCR productsof intergenic DNA and sometimes coding sequence as well. Theresolution of the method depends on the size of the spottedPCR products in addition to the size of DNA fragments used inchromatin immunoprecipitation. Oligonucleotide microarrays havealso been used (13) and hold the promise of greater resolutionand the statistical confidence endowed by multiple signals fromeach region of the genome. However, the distribution of oligonucleotideprobes is highly skewed in E. coli Affymetrix arrays and manyregions of the genome are devoid of probes. Therefore, in thisstudy we used a custom-designed oligonucleotide array from Nimblegencontaining an evenly spaced "tiled" set of probes. Custom Nimblegenarrays have been used previously for chIP-chip of human transcriptionfactors (25), but to our knowledge, this is the first publicationdescribing the use of whole-genome, uniformly tiled arrays.

The accuracy and resolution of chIP-chip have been evaluatedin a number of previous studies. For yeast RNA Pol III, essentiallyall known targets were correctly identified (19, 31, 35). Acomparison of known transcription factor interactions to a largeset of chIP-chip data revealed a false-negative rate of 80%(41). In the present study, we used rifampin to trap RNAP atpromoters and compared promoter locations determined by chIP-chipto those determined by established methods. The existence ofhundreds of known promoter locations in E. coli provides a novelopportunity for the evaluation of this technique.

MATERIALS AND METHODS

Top

Materials and Methods

Chromatin immunoprecipitation. In preparatory studies performed at least three times, E. coliK-12 MG1655 cells were grown in 100 ml LB medium in a shakingflask at 37°C. Chromatin immunoprecipitation was performedsimilarly to the method described in detail below. Endpointquantitative PCR was performed for 30 cycles using primers designedto amplify from all seven rrn operons. For 16S the primers wereGGAGGAATACCGGTGGCGAAGG and GCGTTAGCTCCGGAAGCCACGCCTC. For 23Sthey were CCAGGATGTTGGCTTAGAAGCAGCC and ACAGAACGCTCCCCTACCCAAC.For the promoter, the reverse-direction primer was GTCTGATAAATTGTTAAAGAGCAGT,while the forward-direction primers were a mixture of CTCCCTATAATGCGCCACCACTG,CTCCCTATAATGCGCCTCCATC, and ATCCCTATAATGCGCCTCCGTT.

For chIP-chip, three biological replicates were performed (threeindividual cultures). For each culture, two cell pellets of50 ml culture were processed—one experimental sample asdescribed below and a second control sample identical to thefirst but with no antibody added to the immunoprecipitationreaction. The chIP DNA of the first biological replicate wasanalyzed by real-time PCR using primers in the promoter regionsof lpdA and fepB, which were measured to be present at levels35- and 7-fold higher than background, respectively. The experimentaland control PCR-amplified samples were labeled with Cy3 andCy5 by Nimblegen, swapping the cyanine dyes in one biologicalreplicate. The microarrays consisted of 371,034 oligonucleotidesfrom the E. coli genome (release m56 accession no. NC_000913)plus controls. The oligonucleotides were 50 nt long and spacedevery 25 nt on both top and bottom strands.

The method used for chromatin immunoprecipitation was adaptedfrom Lin and Grossman (28). E. coli MG1655, obtained from theE. coli Genetic Stock Center (CGSC no. 7740), was inoculatedfrom frozen stocks into M9 minimal medium-0.2% glucose and grownovernight at 37°C. It was then diluted 1:50 into 200 mlM9-0.2% glucose in a 500 ml flask, and the mouth of the flaskwas sealed with foil and stirred for 5 to 6 h to an opticaldensity (OD) at 600 nm of $~$ 0.4. Rifampin dissolved in methanolwas added to a final concentration of 150 µg/ml and stirredfor 20 min. Cultures were monitored by OD to verify the inhibitoryeffects of rifampin. A slight increase in OD was observed butwith a greatly reduced growth rate. Samples (50 ml) were transferredto tubes containing formaldehyde and sodium phosphate (pH 7.6)such that the final concentrations were 1% and 10 mM, respectively.They were incubated at room temperature for 10 min with gentleagitation, and then glycine was added to 100 mM. They were gentlyagitated at 4°C for 30 min. Cells were centrifuged at 3,500x g and washed twice in chilled phosphate-buffered saline. Thecells were transferred to 1.5 ml tubes, centrifuged, and storedat –80°C.

Cells were resuspended in 0.5 ml of a solution containing 10mM Tris (pH 8), 50 mM NaCl, 10 mM EDTA, and 20% sucrose, andthen 0.2 µl Ready-lyse lysozyme (Epicenter) was added.They were incubated for 30 min at 37°C, and then 0.5 mlof a solution containing 200 mM Tris (pH 8), 600 mM NaCl, 4%Triton X-100, 1 mM phenylmethylsulfonyl fluoride, and 2 µg/mlRNase A was added. They were incubated at 37°C for 10 minand then chilled on ice. The lysates were sonicated four timesfor 20 s each time using a Sonics and Materials VC50 with a3 mm microtip. Cell debris was removed by centrifugation, anda sample was removed and extracted with phenol and phenol:chloroformfor DNA size analysis. DNAs ranged from 100 to 1,200 bp, withthe greatest intensity at $~$ 600 bp. Pan Mouse immunoglobulin Gmagnetic beads (Dynal catalog no. 110.22) (40 µl) wereresuspended and washed per the manufacturer's instructions andthen added to the lysates and rotated at 4°C for 3 h. Thebeads were removed using a magnet and discarded. A total of2 µl of antibody for RNA polymerase ß' (Neoclonecatalog no. W0001) was added, and the samples were rotated overnightat 4°C. Magnetic beads (50 µl) were resuspended andwashed as before and then added to the samples and rotated 1h at 4°C. The beads were recovered and then washed oncein 250 mM LiCl-100 mM Tris (pH 8)-2% Triton X-100, twice in100 mM Tris (pH 8)-600 mM NaCl-2% Triton X-100, twice in 100mM Tris (pH 8)-300 mM NaCl-2% Triton X-100, and twice in TE(10 mM Tris-HCl, 1 mM EDTA). DNA and protein were eluted fromthe beads by resuspending in 50 mM Tris (pH 8)-10 mM EDTA-1%sodium dodecyl sulfate and then incubated at 65°C for 20min. Beads were removed, cross-links were reversed by incubatingat 65°C for 4 h, and then DNA was purified using Qiaquick(QIAGEN).

ChIP DNA was amplified based on the Random DNA Amplificationmethod of June 2001 from http://www.microarrays.org/protocols.html.The following primers were designed specifically for E. coli:PF43 (TGGAAATCCGAGTGAGTNNNNNNNNN) and PF44 (TGGAAATCCGAGTGAGT).Round B amplification using exTaq (Takara) was performed for35 to 40 cycles. Round C amplification was not performed. PCRproducts were purified with Qiaquick and then ethanol precipitated.Pellets were resuspended in water, and a sample was dilutedin 1 mM Tris (pH 8) to measure the DNA concentration. Experimentalsamples (4 µg) and control samples (3 µg) were hybridizedto microarrays by Nimblegen.

Data analysis. For each array, the Cy5 signal values were multiplied by a scalingfactor in Microsoft Access so that the total signals for Cy3and Cy5 were equal. The log₂ of the ratio of experimental signalto control signal was then calculated. The microarray was designedfrom both the top- and bottom-strand sequences, which meansthat for every probe, its reverse complement was also representedon the array. To distinguish them in the resulting data, 12nt were added to the nominal genome position of probes fromthe bottom strand. Genome-scale data was visualized using Nimblegen'sSignalMap software, and most other analysis was performed withMatlab version 6.5 (see the program code in the supplementalmaterial). For each probe, the mean log₂ ratio from all threebiological replicates was calculated and then smoothed usinga moving average with a window size of 500 bp. Peaks were identifiedusing a reiterative process in which the probe with the highestlog₂ ratio was located upon each iteration. To prevent the samepeak from being reidentified each time, the probe with the highestlog₂ ratio and all data within 1,000 bp were set to zero. Thisprocess was reiterated 1,000 times, and the resulting peakswere manually edited using SignalMap for visualization. An additional139 peaks with a log₂ ratio greater than 0.1 were then identifiedby the same method and manually edited.

RESULTS AND DISCUSSION

Top

Results and Discussion

Effects of rifampin and experimental design. We first determined the time required for rifampin to localizeRNAP to promoters when added to mid-log E. coli cultures. Theouter membrane of gram-negative bacteria like E. coli posesa significant barrier to rifampin, requiring much higher concentrationsfor inhibition of RNAP in vivo than in vitro (4). When 150 µg/mlrifampin was added, growth ceased almost immediately (Fig. 1B).However, chIP signals in the rRNA operon persisted much longer,indicating a lack of coupling between rRNA transcription andcell growth. We found that 20 min of incubation with rifampinwas required for chIP signals in the rRNA operon to reach near-backgroundlevels. Based on this result, we chose 20 min as a suitabletime for rifampin treatment. This amount of time should allowdetection of even the weakest promoters; when the occupancyrate of a promoter by RNAP is only once every 20 min, then achIP-chip signal will be detectable, though slightly diminished.

Binding sites for RNA polymerase in the genome of E. coli strainMG1655 were identified using chIP-chip. Briefly, cells weregrown to mid-log phase in minimal glucose medium and then treatedfor 20 min with rifampin. DNA and protein were cross-linkedwith formaldehyde, and cells were lysed. DNA was sheared to $~$ 600 bp with sonication, and RNAP was immunoprecipitated usingan antibody specific for the ß' subunit. DNA boundby RNAP and from a no-antibody mock immunoprecipitation controlwas amplified using random primers and then labeled with Cy3and Cy5 fluorescent dyes and hybridized to a custom microarrayconsisting of 371,034 E. coli oligonucleotides spaced 25 bpapart across the whole genome.

Reproducibility. This experiment was repeated three times. The reproducibilityof the log₂ ratios from each array is shown in Fig. 2. Highercorrelation was seen when comparing just the experimental channels(coefficients of 0.71, 0.63, and 0.84) or just the controls(0.66, 0.61, and 0.87). This level of correlation is lower thantypically observed in microarray analysis of gene expression,but important differences should be considered. In this experiment,the whole genome was represented with oligonucleotide probes,and yet only a fraction of the genome should be present in theDNA that was hybridized to it. Thus, the majority of oligonucleotideprobes on the array are not hybridized, leading to an increasedamount of noise. The log₂ ratios of chIP-chip experiments tendto have less dynamic range and greater noise than those fromexpression analysis (R. Green, Nimblegen, personal communication).

View larger version (31K):
[in this window]
[in a new window]
FIG. 2. Reproducibility of RNA polymerase chIP-chip data. Log₂ ratios of immunoprecipitated versus no-antibody control DNAs from three hybridizations are plotted against each other. A cluster of outliers in biological replicate 1 consists almost entirely of probes from rRNA gene operons. corr. coef., correlation coefficient.

Unlike a typical gene expression experiment, the absolute log₂ratio is not as important as the shape and location of peakswithin the log₂ ratio data, as visible in Fig. 3. As expectedfrom the high density of evenly spaced probes, the log₂ ratioshowed a clear pattern of peaks and valleys. When peaks wereidentified from each biological replicate individually, themean standard deviation of peak location was 206 bp. The maximumlog₂ ratio in each biological replicate ranged from 4.2 to 5.9,and for all further analysis, replicates were combined by calculatingthe mean. To control for differences in hybridization efficiencyof the oligonucleotides, the log₂ ratios were smoothed witha moving average. In Fig. 3B it can be seen that smoothing preservedthe shape of peaks yet eliminated much of the noise.

View larger version (24K):
[in this window]
[in a new window]
FIG. 3. A sample of RNA polymerase chIP-chip data. In both panels, genome position is indicated at the top and ORFs are shown at the bottom, with those above the line being in the forward orientation and those on bottom being the reverse. (A) rRNA-encoding genes showed a different pattern of distribution in biological replicate 1. The smoothed log₂ ratios from the three individual biological replicates are shown, along with the peaks identified in this region. (B) A sample from another region shows the combined log₂ ratios from all biological replicates before and after smoothing with a moving average over 500 bp.

A large cluster of outliers is visible in the top two panelsof Fig. 2. These consist almost entirely of probes from biologicalreplicate 1 located in genes encoding rRNA and ribosomal proteins,which are the most highly transcribed promoters in the genome.Almost 96% of probes with differences in log₂ ratios > 2between biological replicates 1 and 3 were located in rRNA genesor ribosomal protein operons. The variation at the rrnC operonin the three biological replicates can be seen in Fig. 3A. Itappears that RNA polymerase was distributed throughout the operonin biological replicate 1 but was confined to two promoter regionsin biological replicates 2 and 3. The rifampin stock used inbiological replicate 1 was filter sterilized, which probablyremoved some rifampin and lowered the actual concentration.For biological replicates 2 and 3 the rifampin was not filtered.It seems likely that in biological replicate 1 the rifampinconcentration was lower, and RNAP was able to escape from thesevery strong promoters. It should be noted that two peaks wereobserved at all of the seven rRNA gene operons. The first peakcorresponds to the P1 and P2 promoters, indistinguishable atthis scale. The other peak is probably a conserved promoterof unknown significance noted previously (2).

Binding site locations. From the combined smooth data, 1,139 peaks were identified,representing binding sites for RNAP. They may also representpromoters, if RNA is transcribed from them. Binding sites lessthan a few hundred base pairs apart or on opposite strands areexpected to appear as a single peak. Binding and initiationby RNAP is strongly dependent on other transcription factorsand the conditions tested. Presumably, the sites we detectedare a subset of the total binding sites in E. coli. For instance,peaks were not observed at the sites of either the araBAD orlacZYA promoters, probably because those binding sites are occludedby repressors in glucose medium.

There are $~$ 2,800 RNAP molecules in E. coli in minimal medium(8) and 2,428 predicted transcription units (TUs) in the genome,not including $~$ 40 TUs containing recently annotated small regulatoryRNAs (TU predictions are from www.biostat.wisc.edu/gene-regulation/)(7). Of these TUs, 1,552 contain an open reading frame (ORF)that appears to be expressed under these conditions (log₂ signal> 8.0 in gene expression data generated with Affymetrix arrays)(data from reference 14). Of these, 1,141 have 5' ends thatare more than 500 bp apart, providing an estimate of the numberof binding sites that we might expect in this experiment. Thisnumber agrees remarkably well with our results. Of the 1,139sites we detected, 958 occurred within 1,000 bp of the startsite of a TU. For 659 of these, the TU contained an ORF witha log₂ signal > 8.0 in gene expression data from the samegrowth conditions. Interestingly, this indicates 299 bindingsites that are not associated with active transcription. Inthese cases, RNAP may bind to the chromosome but not activatetranscription until a repressor or activator is triggered inresponse to environmental conditions, as is generally the casefor transcription from ${sigma}$ ⁵⁴-specific promoters. It should be noted,though, that after treatment with rifampin, all active promotersmay become saturated by RNAP, forcing binding to sites thatare not normally bound. Additionally, some promoters may bindRNAP weakly and yet still be active by having high rates ofopen complex formation or promoter clearance.

Of the 1,139 binding sites, 501 occurred in intergenic regionsand 638 occurred in ORFs. Those in ORFs can be subdivided intothree groups on the basis of the distance from the site to thenearest 5' end of an ORF (median distance = 158 bp). The firstgroup consists of binding sites that are probably located upstreamof the ORF but mistakenly identified inside the ORF due to poorresolution of the chIP-chip technique. In this group are 372sites for which the nearest 5' end of an ORF is that of theORF in which they occur and is <500 bp away. The second groupconsists of sites that are probably promoters for a downstreamORF. In this group are 191 sites for which the nearest 5' endof an ORF is an ORF that is different from the one in whichit occurs and is <500 bp away. The third group consists ofsites that occur >500 bp away from the 5' end of any ORF(see Table S3 in the supplemental material). Of these 75 sites,many have small log₂ ratios and occur in insertion elementsor in genes of unknown function. The others occur either inthe middle of an ORF (sites in sapA, rfbX, evgS, emrA, nlpD,pyrG, xylG, and fimD) or near the intergenic region betweentwo convergent genes (sites in pinQ, intQ, and otsA). Thesesites may play a regulatory role, possibly as RNAP pause sitesor promoters of antisense transcripts, or they may be crypticpromoter-like elements ordinarily masked by transcribing RNAP.

Correlation to other factors. There was very little correlation between the chIP-chip log₂ratio and the similarity of known promoters to a ${sigma}$ ⁷⁰ consensusbinding matrix (Fig. 4A). This result is surprising given that(i) the relative efficiency of chIP, often measured as "occupancy,"has been observed to correlate with in vitro binding affinity(39) and (ii) RNAP has high affinity for promoters containingsequences close to consensus (26). A possible explanation forthis lack of correlation may be the excess of RNAP moleculesover the number of potential binding sites in the presence ofrifampin. In excess and "trapped" at promoters, RNAP may bindto all sites equally. We also found very little correlationbetween the log₂ ratio of chIP-chip peaks and the expressionof the nearest downstream ORF (Fig. 4B), which is not surprisingconsidering that the log₂ ratio correlates poorly with the consensusmatrix. This observation differs from studies of yeast Pol IIIin which the log₂ ratio was directly correlated with transcription(35). The discrepancy may be due to biological differences betweeneukaryotic and prokaryotic RNAP, such as the ratio of enzymemolecules to potential binding sites or to disparate mechanismsof gene regulation. More likely, it is due to experimental differences;in the study performed by Roberts et al. (35) rifampin was notused to trap RNAP at the promoters and occupancy was measuredin a transcribed region.

View larger version (28K):
[in this window]
[in a new window]
FIG. 4. Poor correlation between the magnitude of the RNAP chIP-chip log₂ ratio, gene expression, and promoter similarity to the consensus matrix. (A) The RNAP binding site log₂ ratio was plotted against the similarity of the promoter to consensus. Of 681 known ${sigma}$ 70 promoters from RegulonDB, 497 were located within 1,000 bp of an RNAP binding site in chIP-chip data. For 413 of these, a score was available measuring the similarity of the promoter –10 and –35 boxes to consensus, plotted on the vertical axis. The score was calculated by RegulonDB as described previously (21) by establishing a consensus from 584 known promoters and then scoring each promoter by the use of this consensus matrix. (B) For each of the 1,139 RNAP binding sites (peaks) detected by chIP-chip, the log₂ ratio was plotted against gene expression of the nearest downstream ORF. Gene expression data from reference 14 were generated from three biological replicates with Affymetrix arrays using the same growth conditions as for chIP-chip (M9 minimal medium plus glucose at mid-log phase).

We examined the distribution and periodicity of binding sitesin the E. coli genome. The densities of binding sites differedthroughout the genome, ranging from 13 to 36 sites per 100 kb.There was no correlation between the number of binding sitesper 100 kb and the average predicted TU size in each 100-kbregion (correlation coefficient = –0.11; TU predictionsfrom Bockhorst et al.) (7). We did not observe any significantperiodicity in the location of binding sites, but we did detectperiodicity when both the location and log₂ ratio of the peakswere considered (Fig. 5). A wavelet transform (6, 29) revealedthat the average log₂ ratio ranged in a region from $~$ 1,900 to3,400 kb with periodicity of $~$ 660 kb, which is significant relativeto the periodicity found in randomized data. The regions ofhigh and low average log₂ ratios correlated with regions ofhigh and low average gene expression levels (Fig. 5C). Periodicityof this sort has been observed for many different genomic parametersacross the evolutionary spectrum, though the meaning, whetherrelated to the physical organization of the genome or some otherfactor, is unknown (1, 24) (T. E. Allen, N. D. Price, A. R.Joyce, B. Ø. Palsson, unpublished data). It is unusualthat only a fraction of the genome shows periodicity for RNAPbinding sites. It is also interesting that a correlation betweenlog₂ ratio and gene expression was only observed when the log₂ratios were smoothed over 330 kb. On a gene-by-gene basis theremay have been too much noise to see this trend.

View larger version (40K):
[in this window]
[in a new window]
FIG. 5. Density and periodicity of RNA polymerase binding sites in glucose minimal medium as detected by chIP-chip. (A) The log₂ ratios of 1,139 binding sites were averaged in each 1-kb section of the genome. Those sections lacking a binding site were assigned a value of zero. These data were examined for periodicity by wavelet analysis, which smoothes the data over a range of window sizes to reveal the wavelengths at which the data fluctuates most consistently. The period (wavelength) was plotted against genome position, with colors indicating the magnitude of average log₂ ratio on a scale from white to yellow to red to black. The data were randomized and reanalyzed 1,000 times, allowing the estimation of significance of the original plot. Areas of the wavelet plot where the false-discovery rate was <10% are highlighted, while other areas are faded (5). (B) Profile of 660-kb periodicity. The average log₂ ratios per kilobase were smoothed with a moving average window size of 330 kb. The vertical axis indicates the number of standard deviations by which the data differed from the mean (mean = 0.177; standard deviation = 0.027). Regions marked in red and green indicate positive and negative phases of periodicity at wavelength of 660 kb from panel A above. (C) The same as panel B above, but instead of average log₂ ratios from chIP-chip, the log₂ signal from gene expression measurements in glucose minimal medium was smoothed and plotted (14) (mean = 8.38, standard deviation = 0.32). D) The same as B above, except the number of detected RNAP binding sites per kb was smoothed and plotted (mean = 0.246; standard deviation = 0.028).

Comparison to locations of known promoters. To validate the RNAP binding sites, we examined the abilityof chIP-chip to detect 961 known promoters compiled in RegulonDB(36). For each known promoter, the nearest peak was locatedin the RNAP chIP-chip data (Table 1). A total of 418 RNAP bindingsites were located within 1,000 bp of a known promoter, leaving721 binding sites identified by chIP-chip that have not beenpreviously characterized. Of the known promoters detected, mostwere specific to ${sigma}$ ⁷⁰, but high percentages of ${sigma}$ ²⁴-, ${sigma}$ ²⁸-, and ${sigma}$ ³²-specificpromoters were also detected. These sigma factors are associatedwith the transcription of genes for flagella biosynthesis andstress response. Gene expression measurements indicate thatthe genes downstream from ${sigma}$ ²⁸-specific promoters are probablynot expressed under these conditions (average log₂ signal =7.13), while those downstream of ${sigma}$ ²⁴- and ${sigma}$ ³²-specific promotersprobably are (average log₂ signals of 9.62 and 10.71, respectively).The expression of stress response genes does not necessarilyindicate the presence of stress; heat-shock proteins are normallyexpressed at 37°C (20), and ${sigma}$ ³² is required for growth above20°C (42). The log₂ ratio of chIP-chip peaks near ${sigma}$ ²⁴- and ${sigma}$ ³²-specific promoters showed poor correlation to gene expression(r = 0.17). Few ${sigma}$ ³⁸- and ${sigma}$ ⁵⁴- specific promoters were detectedby chIP-chip. These sigma factors are associated with the transcriptionalresponse to stationary phase and nitrogen assimilation, respectively.

View this table:
[in this window]
[in a new window]
TABLE 1. Proximity of known promoters to RNAP binding sites detected in this study

These differences in the proportions of known promoters detecteddo not appear to match the relative abundances of the sigmafactors. During logarithmic growth in LB medium, protein measurementsshowed that ${sigma}$ ⁷⁰ is predominant, followed in abundance by ${sigma}$ ²⁸ and ${sigma}$ ⁵⁴, while the others were undetectable (22). Under the conditionsused here for chIP-chip, the mRNAs for all of the sigma factorsappeared to be expressed, though ${sigma}$ ¹⁹ and ${sigma}$ ²⁸ were less abundant(Table 1). In the presence of rifampin, the alternate sigmafactors may have an increased ability to form complexes withcore RNAP due to an increased level of free core RNAP and thedepletion of ${sigma}$ ⁷⁰ through binding with RNAP at promoters. The ${sigma}$ ³⁸- and ${sigma}$ ⁵⁴-specific promoters may be less often detected becauseof the activity of other transcription factors precluding thebinding of RNAP to ${sigma}$ ³⁸- and ${sigma}$ ⁵⁴-specific promoters under theseconditions. In any case, the detection of binding sites forthe alternative sigma factors indicates that either those sigmafactors are present at some level under these conditions orthere is degeneracy in the binding specificity of ${sigma}$ ⁷⁰-holoenzymein vivo.

The distribution of distances between known ${sigma}$ ⁷⁰-specific promotersand chIP-chip RNAP binding sites is shown in Fig. 6. We foundthat 49% of known ${sigma}$ ⁷⁰-promoters correlated to a chIP-chip peakwithin 200 bp, and 73% correlated within 1,000 bp. Of 681 known ${sigma}$ ⁷⁰-promoters, 532 occur upstream of an apparently expressedORF (the nearest downstream ORF had a log₂ signal > 8.0).Of these, 368 were separated by at least 500 bp, and out of368 sites, 273 could be linked to a chIP-chip peak within 1,000bp. This leaves 95 known expressed promoters that we were unableto detect, resulting in a 26% (95/368) false-negative rate.This compared favorably to chIP-chip experiments in yeast, forwhich a false-negative rate of 80% has been estimated (41).It falls short, though, of the nearly perfect detection of RNAPol III binding sites (19, 31, 35). The false-positive rateis impossible to estimate, because binding does not always resultin transcription and because of the lack of any other viablemethods for measuring protein binding in vivo.

View larger version (29K):
[in this window]
[in a new window]
FIG. 6. Histogram of the distance from 681 known ${sigma}$ ⁷⁰ promoters to the nearest RNA polymerase binding site (peak) observed in chIP-chip data.

False negatives. A variety of factors could explain the false negatives. Oneexplanation may be failure to amplify some parts of the chromosomeby random PCR or low efficiency of formaldehyde cross-linkingat some promoters (17). As chIP-chip data becomes increasinglycommon, characteristics of poorly cross-linked proteins or DNAsequences may emerge. It is also possible that the epitope forthe antibody we used was occluded by other transcription factorsat some promoters. Antibodies specific to other RNAP subunitsare available, and it may be found that their use detects anoverlapping set of RNAP binding sites. Results obtained by performinga single replicate of chIP-chip with antibodies for ${sigma}$ ⁷⁰ and ß'with Affymetrix arrays showed that the ${sigma}$ ⁷⁰ sites were a subsetof the sites detected using antibody for ß', as expected(see Table S4 and Fig. S2 in the supplemental material). Of13,499 probes with ${sigma}$ ⁷⁰ log₂ ratios > 1, 97% had similar ß'log₂ ratios (difference < 1). On the other hand, of 15,032probes with ß' log₂ ratios > 1, only 74% had similar ${sigma}$ ⁷⁰ log₂ ratios. The log₂ ratios from the ß' and ${sigma}$ ⁷⁰Affymetrix chIP-chip experiments correlated with a coefficientof 0.72. The log₂ ratios from the Affymetrix chips were comparedto those from the Nimblegen chips. For each Affymetrix probe,the nearest Nimblegen probe in the same genomic orientationwas identified, correcting for the different versions of theE. coli genome sequence that were used in chip design. The correlationcoefficients were 0.60 for comparisons of Affymetrix ß'to Nimblegen ß' and 0.64 for comparisons of Affymetrix ${sigma}$ ⁷⁰ to Nimblegen ß'. Affymetrix arrays were not usedin any further experiments because of the incomplete coverageof the E. coli genome provided by Affymetrix expression arrays(see Fig. S2 in the supplemental material).

Another likely explanation for failure to detect RNAP at known ${sigma}$ ⁷⁰ or other sigma factor promoters is the high variability ofthe rates of open complex formation and promoter escape andof the affinity of promoters for initial binding of RNAP holoenzyme(33). These same parameters may explain why it takes so longto clear RNAP from the rRNA operons. At some E. coli promoters,open complexes are so unstable that they initiate immediatelyupon binding nucleoside triphosphate without the slow promoterescape and high probability of abortive initiation observedat many promoters. The rrn P1 promoter is a classic exampleof such a promoter. It exhibits no abortive initiation and hasthe added features of highly efficient initial RNAP bindingand the ability to sequester up to 80% of transcribing RNAPinto elongation complexes on the seven rrn operons in E. coli(15, 32). RNAP released at a terminator may rebind an rRNA promoterand reform an rrn elongation complex more rapidly than rifampincan bind to it ( ${sigma}$ ⁷⁰ can associate with RNAP during termination)(3). Thus, RNAP may become trapped at rRNA promoters only aftermultiple rounds of transcription and when the rifampin concentrationinside the cell accumulates to higher concentration, accountingfor the slow clearance of RNAP from the rrn operon (Fig. 3A).Other promoters, which form unstable open complexes but lackthe avid initial RNAP binding and rapid promoter escape propertiesof rrn promoters, may never accumulate significant RNAP, becauseRNAP will equilibrate away to promoters able to form stableopen complexes. Thus, certain promoters that form unstable opencomplexes may escape detection. Consideration of the kineticproperties of promoters will be an important issue in the developmentof promoter mapping by the rifampin-trapping chIP-chip method.

It should also be noted that many of the sites we identifiedare likely to be specific to the minimal glucose conditionsused here. Expansion of this experiment to other culture conditionswill likely reduce the number of undetected known promotersand reveal differences in RNAP binding resulting from the activityof the regulatory network. Gene expression data also show changingpatterns under different conditions but are often difficultto interpret. By using chIP-chip, direct regulatory interactionscan be distinguished from indirect regulatory interactions (27,34), but care should be practiced given the false-negative rateof 26% that we observed.

Closing remarks. Transcriptional regulation can be seen as a combinatoriallybased molecular "computation." In this analogy, the bindingand interaction of various transcription factors with RNAP arethe input variables and transcription level is the output. Theidentification of transcription factors involved in each "computation"is a fundamental step in understanding observations of geneexpression. As each new factor is identified by chIP-chip orsome other technique, its role will be defined by its interactionand colocation with RNAP. The work presented here lays the foundationfor future work in elucidating the regulatory network of E.coli. The location data can also be used for validation of computationalpredictions of promoters, and the log₂ values may provide insightinto some characteristic of RNAP binding or activity. Whetherthe observed false-negative detection of known promoters revealsan inherent limitation of the technique or is merely a peculiarityof the experimental conditions is yet to be seen. A comparisonof experiments with or without rifampin added, under differentgrowth conditions, and with antibodies for various sigma factorswill shed light on this question.

ADDENDUM IN PROOF

Top

Addendum in proof

After this article was accepted, a genome-wide map of humanRNA polymerase II preinitiation complexes made using chIP-chipand Nimblegen microarrays was published (T. H. Kim, L. O. Barrera,M. Zheng, C. Qu, M. A. Singer, T. A. Richmond, Y. Wu, R. D.Green, and B. Ren. Nature doi:10.1038/nature03877 [29 June 2005]).A large number of false negatives were encountered in that studyas well, which the authors attribute to lack of sensitivityin microarray detection of immunoprecipitated DNA.

ACKNOWLEDGMENTS

The analysis of known promoters would not have been possiblewithout the data provided by Araceli Huerta-Moreno, HeladiaSalgado-Osorio, and Julio Collado-Vides. We thank B. Ren, B.K. Cho, and A. Joyce as well as H. Holster, K. Nuwaysir, andM. Hogan at Nimblegen for their help.

This work was initiated at the University of Wisconsin Madisonwith the support of University of Wisconsin I&EDR funds,Hatch McIntyre Stennis (U.S. Department of Agriculture), andUniversity of Wisconsin trust funds (to A.Z.A.) and a grantfrom the National Institutes of Health (GM38660 to R.L.) andcompleted at the University of California San Diego with supportfrom the National Institutes of Health (GM62791 to B.Ø.P.).C.D.H. was supported in part by NLM 5T15LM007359.

B.Ø.P. serves on the Scientific Advisory Board of GenomaticaInc., a spin-off company from the University of California SanDiego.

FOOTNOTES

* Corresponding author. Mailing address for Bernhard Ø. Palsson: UC San Diego Bioengineering, 9500 Gilman Drive, Dept 0412, La Jolla, CA 92093-0412. Phone: (858) 534-5668. Fax: (858) 822-3120. E-mail: bpalsson{at}ucsd.edu. Mailing address for Aseem Z. Ansari: Department of Biochemistry, 433 Babcock Drive, University of Wisconsin-Madison, Madison, WI 53706. Phone: (608) 265-4690. Fax: (608) 262-3453. E-mail: ansari{at}biochem.wisc.edu.

Supplemental material for this article may be found at http://jb.asm.org/.

REFERENCES

Top