The promoter activity in cells is determined by regulatory proteins (activators and repressors) that can recognize overlapping sequences specific for Eσ⁷⁰ RNA polymerase sites, and thereby mask the true promoter strength. In addition, almost 20% of E. coli RNA polymerase Eσ⁷⁰-specific promoters possess an extended -10 sequence that might compensate for the absence of a clear -35 site 35. Different prediction programs based on statistical and motif-searching approaches have been developed to detect a variety of binding sites in DNA, and both position-specific weight matrices 36 and hidden Markov models 37 have been used to improve the accuracy of the prediction of promoter sequences in bacterial genomes 383940. These programs usually detect hexanucleotide dyad patterns of RNA polymerase-promoter binding sites, such as -35 and -10 boxes, and none of them is free of false-positives, which correspond to similar, non-promoter sequences in bacterial genomes [for a review, see 41].

In this study, we exploited the strengths of the "triad pattern" approach to develop an algorithm able to detect strong promoters by matching three nucleotide sequences recognized by the σ⁷⁰ and α subunits of bacterial RNA polymerase. Theoretically, the presence of a UP element may not be essential for relatively strong promoter activity if two -35 and -10 boxes are well conserved and optimally distanced. Similarly, the presence of a well conserved UP element may compensate for a poor -35 box in some promoters. However, it seems very likely that the strongest promoters probably possess all three essential sequences. The specific interaction between the UP element and the α subunit significantly amplifies the association of RNA polymerase with promoter DNA 2728. Therefore, to improve the filter to exclude possible false-positive due to short hexanucleotide similar sequences scattered throughout the genome, our algorithm starts by first matching the UP element, and only then identifying the -35 and -10 boxes located further downstream.

We designed an algorithm able to detect the triad nucleotide patterns in bacterial genomes. The core of the algorithm is the FIND_TRIAD procedure, which given an input nucleotide string, s, returns the substring s' of s, which is the best approximation of a given triad pattern of the form (pat(1),L1)-(l1,l2)-(pat(2),L2)-(d1,d2)-(pat(3),L3), where each pat(i), i = 1,2,3, is a nucleotide string, Li is its length, l1 and l2 are the minimum and maximum distances respectively between the first and the second patterns, and d1 and d2 are the minimum and maximum distances respectively between the second and the third patterns. To avoid making a "bad" approximation, three scores Sc1, Sc2 and Sc3 are used as input parameters for the procedure. The resulting substring, s', can then be represented as (spat(1),Ls1)-ls1-(spat(2),Ls2)-ls2- (spat(3),Ls3), where each spat(i), i = 1,2,3, is a substring of s aligned to pat(i), Lsi is its length, ls1 is the distance between spat(1) and spat(2), and ls2 is the distance between spat(2) and spat(3). This result for s' satisfies the following conditions:

(1) for each i = 1,2,3 the similarity score (weight) Wi of the match or alignment of pat(i) and spat(i) is not less than Sci (or the number of "mismatches" does not exceed (Li - Sci));

(2) (l1 ≤ ls1 ≤ l2) and (d1 ≤ ls2 ≤ d2).

For each of the three patterns, one can either forbid insertions/deletions or allow them. In the former case, Lsi = Li and the weight = Wi are computed as the sum of matching pairwise symbols, whereas in the latter case, the difference |Lsi - Li| between spat(i) and pat(i) is bounded by a value Ri for the permissible deletions/insertions (gaps), an optimum alignment, and its weight, Wi, are computed by the standard dynamic programming method for global string alignment 42. In both cases, a symbol scoring matrix Mi(x,j) is used to define the weight of the symbol x in the position j, 1 ≤ j ≤ Lsi, of spat(i). If symbol x occurs in position j of pat(i), then Mi(x,j) = 1, otherwise Mi(x,j) ≤ 1. To choose the best approximation of the triad pattern from substrings satisfying conditions (i) and (ii), FIND_TRIAD uses a total score function with the form:

tot_sc = C1*nsc1(L1,W1)+D12*nsc_dist12(11,l2,ls1) + C2*nsc2(L2,W2) + D23*nsc_dist23(d1,d2,ls2) + C3*nsc3(L3,W3),

where nsci(Li,Wi), i = 1,2,3, are normalized scores of matching (alignments) of pat(i) and spat(i), 0 ≤ nsci(Li,Wi) ≤ 1, and nsc_dist12(11,l2,ls1) and nsc_dist23(d1,d2,ls2) are the normalized scores of the distances between the first and the second, and the second and the third patterns, respectively, and 0 ≤ nsc_dist12(11,l2,ls1), nsc_dist23(d1d2,ls2) ≤ 1. The linear coefficients C1, C2, C3, D12, and D13 are chosen so that their sum is equal to 1. They indicate the relative importance of the corresponding sub-patterns of the triads; and the distances between them. So, the best value of tot_sc is 1.

Here we describe the main parameters of the FIND_TRIAD procedure used to detect strong promoter candidates in bacterial genomes. In this study, a bacterial promoter is assumed to be a nucleotide sequence, located upstream from genes encoding proteins, tRNAs or rRNAs that could be recognized by an RNA polymerase holoenzyme containing a major σ factor (using E. coli Eσ⁷⁰ RNAP as the reference). The triad patterns defined for strong promoter candidates include three specific nucleotide sub-regions: (i), a UP element, which is a 17-nt prefix of the strong promoter, and has the following consensus pattern: pat(1) = P_UP = aaaWWtWttttNNNaaa; (ii) the -35 site, which is located downstream of the UP element, and has the pattern pat(2) = P₃₅ = tc

ttgaca

To define the total score function, tot_sc (formula 1), we chose the following normalized scores for the three patterns and for the distance between the -35 site and -10 site (no information was available about the best values for the distance between the UP element and the -35 site):

nsc1(17,W1) = nsc_up = 1 - (17 - W1)/20,

nsc2(9,W2) = nsc_35 = 1 - (9 - W2)/10,

nsc3(6,W3) = nsc_10 = 1 - (6 - W3)²/10,

and the values of the normalized distance score, nsc_dist23(14,20,ls2) = nsc_dist, are defined as follows:

l s 2 : distance between the -35 and -10 sites, nt 17 16 , 18 15 , 19 14 , 20 n s c _ d i s t 1 0.95 0.85 0.7 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@892E@

We also chose linear coefficients C1 = 0.3, C2 = C3 = 0.25, D12 = 0, and D23 = 0.2. These coefficients indicate the relative importance of corresponding sub-regions for evaluating the total score of a candidate sequence. They were chosen empirically, after preliminary tests with several annotated genomes, assuming a higher significance of the UP element, equal significance of the -10 and -35 boxes, and lower significance of the distance between them. In this application, the value D12 = 0 means that we ignore the variations of the distance between a putative UP element and -35 box because a priory it is not known what value is the best in the interval 0–5 nt.

Formulas 2, 3 and 4 reflect the lack of exact matching for the different sub-regions. If the -10 box is highly conserved and is essential for initiation of transcription 22, then the penalty for its mismatches is higher than for those of the other parameters. For example, for 2 mismatches, the penalty is (6 - 4)²/10 = 0.4 for the -10 site, whereas it is (9 - 7)/10 = 0.2 for the -35 site, and (17 - 15)/20 = 0.1 for the UP element. The choice of the normalized score functions in equations 2, 3 and 4 is based on empirical observations, and on common sense, and may seem to be arbitrary. We want to stress that, in fact, the total score function tot_sc also has a further role: it does not significantly change the set of the best candidates identified by the algorithm. This set is defined by the three score bounds Sc1 = scup for UP element, Sc2 = sc35 for -35 site, and Sc3 = sc10 for -10 site. The total score affects only the ordering of these candidates amongst themselves.

The general scheme of the algorithm is as follows. It has the following input: (i) the name of a genome file in GenBank format; (ii) three parameters of scores: scup, sc35 and sc10, determining the minimum acceptable similarity between candidate sequences of the UP element, the -35 box, and the -10 box, respectively, and the E. coli consensus patterns. For each gene in the genome input file that is not inside an operon, the algorithm runs in two steps:

(i) it extracts a 300-bp DNA region, s, upstream of the annotated coding sequences for tRNA, rRNA or proteins (we limited the search to 300 bp, since most E. coli promoters fall within this length inter-gene space 4145);

(ii) then it uses the FIND_TRIAD procedure to identify the best strong promoter candidate within s that satisfies conditions (1) and (2) above. If such a candidate is found, it is added to the output list of strong promoters.

We recommend to read attentively the "ReadMe" information [see Additional file 1] before to start proceeding the "strong_promoters.doc" software [see Additional file 2]. The algorithm is implemented by a program that produces the results in two forms: (i) a Text-format table which lists all strong promoter candidates in a genome, and provides additional information about the operon organization of genes located downstream (for example, see Fig. 1); (ii) a Word-format table which lists strong promoter candidate sequences. A 20-nt sequence preceding a possible initiation codon of each ORF is also included in the annotation, as this could be useful for the visual examination of the translation signals of the corresponding genes. Lastly, the user can select a convenient score for each sequence-specific motif taking into consideration the promoter features of the annotated genome if they differ from the E. coli-specific patterns used to create the algorithm (for example, a weakly conserved -35 or -10 box).

Putative promoter regions in the T. maritima genome, identified by the algorithm described above, were amplified by PCR using appropriate oligonucleotide primers connected to the previously-described G. stearothermophilus argC gene 46. This reporter gene encodes N-acetyl glutamylphosphate reductase, a thermostable and soluble protein that is easily detectable after exposing E. coli cleared lysates to 65°C. In order to increase protein yield, the ribosome-binding site of G. stearothermophilus argC was modified to the sequence

GGAGG

ATG

Two well-characterized strong promoters, Ptac and PargC, were used as references to compare the strength of the putative promoters of T. maritima. The strong promoter Ptac contains an AT-rich nucleotide sequence upstream of a -35 site 48, which has no defined UP element; it was obtained from the vector pBTac2 (purchased from Boehringer Mannheim). PargC, a strong promoter of G. stearothermophilus, contains the UP element, as demonstrated both in vivo and in vitro, and was amplified from the plasmid pHAV2 32.

PCR-generated linear DNA fragments carrying a promoter region fused to the argC reporter gene were used to evaluate the promoter strength in a coupled transcription-translation system, as described previously 49. The cell-free extracts were prepared from the E. coli strain BL21 (DE3) Star recBCD (our laboratory construction) as described by Pratt 50. Protein synthesis was carried in the presence of pyruvate oxidase to generate ATP 51. Typically, 50 ng of PCR-amplified DNA was added to a pre-mix containing all necessary compounds and 10 μCi of [α³⁵S]-L-methionine (specific activity 1000 Ci/mmol, 37 TBq/mmol, Amersham-Pharmacia Biotech), and E. coli S30 cell-free extracts. The reaction mixture was incubated at 37°C for 90 min, and heated to 65°C for 10 min. After centrifuging, the supernatant was precipitated with acetone, and then protein samples were separated by SDS-PAGE and bound to 3 MM paper. The ArgC protein synthesized in vitro was quantified by counting the radioactivity of the corresponding band with a PhosphorImager 445 SI (Molecular Dynamics).

The bacterial genome sequences were extracted from available data banks. The logo of T. maritima promoter consensus sequences was generated at the WebLogo site as described in 5253.

In our algorithm, 26 of the 32 symbols used to evaluate matches in the three promoter-specific patterns, namely in the UP element and the -35 and -10 boxes, are a and t. One could expect the number of genes transcribed from potential strong promoters to depend on the A+T content of a given genome. To find out whether this is indeed the case, we compared the frequency of candidates in 300-bp regions located upstream of genes of annotated bacterial genomes and in random sequences of the same regions generated by computing. First, we calculated the (A+T)% in all 300-bp regions preceding each gene or operon in the annotated genomes (Table 1). The A+T content in these DNA regions was found to be slightly higher than that of the entire genomes of almost all bacteria that have been analyzed. Next, we generated 10.000 random sequences with the same A+T content for all the 300-bp regions of each genome. The algorithm was applied to detect strong promoter candidates in the 300-bp real genomic and random-generated regions of 43 bacterial genomes.

We tested different matching stringencies and empirically found that the score parameters sUP = 13, s35 = 5.5 and s10 = 4.5 satisfied the criteria required for scaled comparative analysis without grossly exaggerating the number of candidate sequences identified in the various genomes. This analysis revealed that the real genomes with an A+T content of less than 50% contained many more potential strong promoters than their simulated counterparts (see Table 1). The percentage of candidate sequences was very low in the bacterial genomes with an A+T content of between 33% and 47%, and these sequences were completely absent in the corresponding 300-bp, random-generated sequences. When the A+T content increased from 47% to 78%, the percentage of strong promoter candidates increased dramatically, whereas the difference between the real and random sequences decreased, and virtually disappeared when the A+T content exceeded 62%. There were two exceptions where the genomes analyzed did not display this pattern at an A+T content of less than 62%. One was M. pneumoniae, the genome of which had an A+T content of about 60%, and in which the promoters had no -35 consensus 54. The other example is the hyperthermophilic species A. aeolicus (~58% AT-rich genome). This species is very close to the Archaea, and occupies a unique position in the bacterial kingdom 55.

Our data show that the number N(A+T) of strong promoter candidates in 300-bp random-generated sequences corresponding to upstream regions of bacterial genes satisfies the "exponential low" of the form N(A+T) = exp [c₁ (A+T) + c₂]. The distribution of strong promoter candidates in real genomes indicates that the critical point of the A+T content is close to 62% (Fig. 3). Above this level, the number of random sequences reminiscent of strong promoter patterns increases markedly.

Another important aspect of the quality of detection is the location of candidate sequences with regard to coding regions in the genome analyzed. We compared the frequencies of strong promoter-like patterns identified upstream and downstream of the initiation codon in all the genomes. The frequency of candidate sequences was clearly greater in the upstream region of ORFs in most of the genomes with an A+T content of less than 62% (Table 2). No difference was detected in T. pallidum (~47% AT-rich genome), which belongs to a distinct phylum of Spirochetes that appear to use different DNA patterns for the promotion and regulation of transcription 56.

The fact that more candidate sequences were identified upstream of ORFs highlights the fact that they are not randomly distributed in bacterial genomes, which suggests that the detection of strong promoter candidates in genomes with an A+T content of less than 62% is fairly reliable.

Taking our cue from the results of the virtual prediction, we sought to find out whether, and if so, to what extent the putative promoters are functional in a biological context. To do this we used reporter-gene technology, which relies on the fusion of an assayable sequence with a promoter being investigated, and the subsequent evaluation of promoter strength in a cell-free system (see Fig. 2). The genome of the hyperthermophilic bacterium T. maritima 57 was used to evaluate the feasibility of the algorithm experimentally.

63 candidate sequences were detected in the T. maritima genome using the matching scores described above. We increased the penalty for mismatching of -35 and -10 boxes by raising the scores of s35 and s10 to 6 and 5, respectively. This reduced the number of candidate sequences to 34 (Table 3). In this shorter list, 28 T. maritima strong promoter candidates possessed a total score higher than the 0.8475 calculated for the reference strong promoter, Ptac, that does not have a typical UP element 48. 15 of these candidates had a total score higher than 0.8775, as estimated for PargC, another reference strong promoter that has a well defined UP element 3249. It is worth mentioning that 6 candidate DNA regions in T. maritima had a total score higher than 0.91, a value estimated for E. coli promoters that govern the transcription of 16S ribosomal RNA, and which were used as models for studying the stimulating effect of the UP element on gene expression 58.