We used a combination of literature search and direct experimentation to generate a list of qualified ER-binding sites. In this study we constrained ourselves to using only sites that have been validated for the modeling of functional EREs. We first extracted human ERE sequences that have been experimentally validated in the literature to either bind or not to bind ER. Klinge 4 and Bourdeau and coworkers 10 each described EREs that have been validated by electrophoretic mobility shift assays, transient transfection with reporter gene constructs, or ChIP assays.

Supplementing the list of confirmed EREs gleaned from the literature, we experimentally identified functional ER-binding sites using two whole-genome experimental strategies. The first strategy was to extract candidate ER-binding sites computationally from a list of putative direct ER target genes. Eighty-nine putative direct target genes were identified as genes expressed in MCF-7 cells that were responsive to estradiol treatment, sensitive to inhibition by Faslodex (ICI 182,780), and insensitive to cycloheximide 8. We then computationally surveyed 3.5 kb regions flanking the TSSs (-3 kb to +0.5 kb) of these 89 genes to identify proximate consensus EREs (allowing for deviations in up to two conserved positions of the consensus motif). Each site was then tested by ChIP assays and qPCR with site-specific primers to determine the true nature of ER binding. Eight EREs were found to be bound by ER, whereas 41 others were not found to be bound by ER.

In our second approach, we performed ChIP assays on estradiol-treated breast tumor cells and detected ER-binding sites using high-density oligonucleotide microarrays (NimbleGen, Madison, WI, USA) containing probes against proximal promoter regions (-1 kb to +0.2 kb from TSS; 12 probes per promoter) of over 30,000 human known gene and RefSeq transcripts annotated in the human genome sequence hg16 (July 2003), NCBI build 34 annotation of the UCSC genome browser. The ChIP-on-chip studies were performed using duplicate array experiments on the ChIP samples and on input control DNA. The promoters that appeared among the top 5% of the binding ratio range (ER antibody versus control) for both replicates, that had at least a 15% increase, and that were supported by consistent binding ratio enrichment across more than four probes or additional evidence of ER regulation from the microarray data were selected. Putative EREs (allowing for up to two mismatches from the consensus) were then identified in the selected promoters, and some were further validated by additional ChIP and qPCR (see Materials and methods, below, for more detail). Out of the total 28 sites tested, 13 were found to bind ER whereas 15 were not. From the literature sources and experiments described above, a total of 45 validated ER-binding sites and 58 validated non-ER-binding were identified, all of which bore close resemblance to the consensus ERE (Table 1). Each of the 45 binders and 58 non-binders was associated with a gene and most were located in the genes' upstream regulatory regions. This list of 103 genes were used as the training set to assess the significance of ancillary sequence signals beyond the core ERE that might better predict ER binding.

ER is known to interact with the 10 base pair (bp) long consensus ERE (hereafter referred to as the 'core ERE'). Presence of the consensus site (or its acceptable variants) is required for the direct binding of the ER dimer to the DNA. However, it is still unclear whether the core site alone is sufficient to signal activated ER for such binding or whether additional ER-binding signals in the sequences flanking the core can be used to distinguish binders from nonbinders. An in silico supervised learning experiment was devised to explore these possibilities.

We modeled the problem of finding additional signals for ER binding among the sequences surrounding the core ERE as a binary classification problem (binders versus nonbinders). The features were position-specific motifs surrounding the core ERE. In other words, we asked whether there is any motif (m) within a definitive distance (p) to the core ERE that could help distinguish the binders from nonbinders. The robust and versatile naïve Bayesian classification approach was employed, with binary tuple <m,p> as features, where m is a k-bp long motif and p is the distance between motif m and the core ERE. Two sets of experiments were set up. The first consisted of the core plus its flanking regions, whereas the second considered only the flanking regions of core ERE. The motif length k and the size of flanking regions were similarly varied in both setups. The goal was to learn whether motifs of certain length at particular distances from the core could contribute to the discrimination of binders from nonbinders. Although the results indicated that window size (k) of 1 bp generally outperformed the rest (Additional data file 1), the span of flanking regions did not appear to affect significantly the outcome of the two experiments.

These observations suggested that additional signal for ER-binding might lie in the distribution of single nucleotides adjacent to the core ERE. This hypothesis was initially investigated by visually inspecting the sequence logo 11 constructed from the binders, including their flanking sequences. Shown in Figures 2a (for ER binders) and 2b (for nonbinders) are the logos for up to 10 flanking nucleotides. Comparison between the binders and nonbinders revealed that additional binding signals potentially came from adjacent nucleotides, specifically those up to 3 bp flanking the core ERE, which extended the consensus palindrome. A series of Monte Carlo runs, performed to estimate the probability that observing such additional signals could happen by chance alone, showed that the signals are statistically significant at 3 bp away from the core motif (Monte Carlo P value = 0.002 and P value < 0.001; see Materials and methods and Additional data file 3).

To determine the functionality for the conserved cytosine and guanine three bases upstream of the first ERE half-site and downstream of the second ERE half-site, respectively, we examined the interactions between ER and wild-type and mutant binding sites using surface plasmon resonance (SPR) spectroscopy. Purified ER was incubated with either the previously validated ERE (wild-type) adjacent to the GREB1 gene or mutants containing substitutions in the conserved guanine (mutant 1), the canonical half-sites (mutant 2), in the conserved guanine and the cytosine in the symmetrical position upstream of the first ERE half-site (mutant 3; see Figure 3a), and at the sixth bases upstream of the core ERE (mutant 4; see Figure 3a) as the negative control. Substitution of the conserved guanine (mutant 1) disrupted ER binding by about 40%, and, as expected, mutations in the consensus half-sites reduced binding significantly (see Figure 3b). Interestingly, substitution of the cytosine three bases upstream of the first half-site with an adenine (Figure 3b, mutant 3), in addition to the substitution in the conserved guanine adjacent to the second half-site, further diminished binding. As was also expected, the substitution outside the three bases flanking the ERE did not perturb the binding significantly. These results indicate that the conserved guanine outside of the canonical ERE, discovered by modeling novel ER binding site, is involved in mediating ER binding to the ERE.

The model we propose, h-ERE, exploits the above observation and consists of two PWMs representing the models for binders and nonbinders. The model relies on a decision tree for classifying sites into binders or nonbinders, based on the scores obtained from the individual PWMs. Two sets of 19 bp sequences, one for binders and the other for nonbinders, were formed from the core sites plus three adjacent nucleotides. We further optimized the binding EREs by minimizing the total entropy of the aligned sites (see Materials and methods), while augmenting the nonbinding EREs by taking both strands of the validated nonbinding loci when constructing the weight matrix.

With this information we constructed a decision tree for the selection of high-likelihood binding EREs versus nonbinding EREs. Each matrix was used to calculate the log-likelihood of a given 19 bp site to be a binder or a non-binder. For each site two scores can be calculated, the binding score (S_B) and nonbinding score (S_NB). Complementing the matrices, a decision tree for distinguishing binders and nonbinders based on S_B and S_NB was constructed from all of the training dataset using the CART algorithm 12 implemented in R, with 100 cross-validation runs. Figure 4 depicts the resultant tree. Putative binders are further subcategorized into three groups, from weak binding (group 1) to strong binding (group 3). Apart from these groupings, sites whose raw log-likelihood binding score (S_B) is greater than its nonbinding (S_NB) scores are potentially functional sites. Additionally, to reflect the nature of the validated sites, the model considers sequences whose core EREs have more than 4 bp mismatches with the consensus ERE, GGTCAnnnTGACC, to be non-binding.

In all, given a 19 bp sequence, the proposed h-ERE first checks whether the core 13 bp nucleotides contains at most four mismatches to the consensus ERE. Next, based on the computed PWM scores, predictions can be made based on four stringency levels: stringent (considers only sites in group 3 to be binders), medium (predicts sites in group 3 and group 2 to be as binders), relaxed (considers sites of groups 1-3 to be binders), and loose (defines sites whose S_B > S_NB as binders).

In previously described studies conducted to identify EREs, the analyses have largely focused on the 5' cis-regulatory regions of direct target genes. However, ChIP analysis of predicted EREs in the extended promoters of 89 putative direct target genes defined by hormone and inhibitor treatments and microarray expression data 8 indicated ER binding in only 9% of the promoter regions from genes apparently directly regulated by ER. These results suggest that ER may target binding sites outside of the canonical 5' promoter regions. Therefore, to discover additional EREs in an unbiased manner and to generate a dataset for testing model performance, we employed the 'ChIP-and-clone' strategy of cloning precipitated DNA fragments into a bacterial plasmid vector, followed by direct sequencing of the inserts to identify ER binding sites. This approach has the potential to sample any region of the genome, as opposed to PCR-based or microarray-based directed strategies, which target specific sites or functional regions, respectively. Anti-ER ChIP was performed on nuclear lysates from estradiol-treated MCF-7 cells, followed by cloning of precipitated binding sites into the pCR-Blunt (Invitrogen, Carlsbard, CA, USA) vector. From the ChIP library, a total of 1006 clones were successfully sequenced and specifically mapped to the human genome. Based on the presence of ERE-like sequences or supporting microarray expression data for ER regulation of the adjacent transcript, 33 clones were selected for subsequent validation by ChIP and site-specific qPCR. An additional 75 clones were randomly selected from those that have neither EREs nor adjacent transcript expression data for further validation (data not shown). Thus, a total of 108 clones were validated (five contained EREs and are supported by microarray expression data, 23 with only EREs and no supporting expression data, five supported by microarray but no EREs, and 75 with neither EREs nor expression data).

The validation results indicate that ERE-like sequences remain the predominant feature of functional ER-binding sites. In the five clones with EREs and supporting microarray expression data for ER regulation, the validation rate was 100%; for the 23 clones that encode EREs but lack supporting expression data, the validation rate was 57% (13/23). In contrast, clones for which no ERE-like sequences were detected, the validation rates were 40% (2/5) and 9% (7/75), respectively, for those with and without supporting expression data for the adjacent gene. A total of 19 EREs were found in the 18 empirically verified ER-bound clones. Interestingly, the five validated clones that contain EREs and are adjacent to genes that were shown to be hormone regulated map to intronic regions of the target genes. This is consistent with our hypothesis that ER may bind outside of the 5' cis-regulatory regions of target genes. Moreover, when we tested ERE-like sequences in the promoter region of one of the target genes, SIAH2, we did not detect ER binding, suggesting that the intronic ERE is the functional ER binding site (data not shown) for this particular target gene. From this analysis, all EREs that bind ER and did not bind ER in the validation experiments were then used to test model performance (Table 2).

Currently, three other models have been widely used to predict functional EREs: consensus sequence search (allowing for certain mismatches), TRANSFAC matrices using MATCH 13 search algorithm, and Dragon ERE finder 6. The performance of these models (under different settings) is compared with h-ERE in Table 3. Although h-ERE was not the most sensitive or the most specific, it offered the best balance between the two criteria. With the interest of having a single performance measure that captures the balance between sensitivity and specificity, harmonic means of the two were computed (see van Rijsbergen 14 and Materials and methods). By this measure, h-ERE offers the best balance in performance, even under different stringency settings.

In order to assign specific ERE predictions, we constructed a decision tree using binding and nonbinding scores from the PWMs (see Materials and methods). The parameters were selected to minimize error on the classification of the training set. We scanned the human genome (UCSC hg17) using the h-ERE decision tree and detected 38,024 putative sites under the 'stringent' criteria, including 3607 EREs encoded by Alu repeats. To assess further the performance of our predictive algorithm, we randomly selected 60 sites predicted to be ER binders by h-ERE (group 3 sites) and 60 nonbinders (group 0 sites) for further experimental validation by ChIP and qPCR. Of the 120 sites, specific primers for qPCR could be designed for only 64 sites, 44 of which are binders whereas 20 are nonbinders. Fourteen per cent (6/44) of the predicted binding sites were shown to bind ER (more than twofold enrichment over control) whereas no binding was detected in any of the sites classified as nonbinders (0/20), suggesting that the false-negative rate is less than 5%. The low rate of false negatives allows us to demarcate in the human genome the global set of EREs that contain the universe of putative true binding motifs. This suggests that, taking into account the 14% validation rate, there would be 5363 validated ER-binding sites within the global optimized ERE set for the MCF-7 cells, under conditions similar to our experimental setup.

We then considered how much of the predictions could be attributed to random occurrences simply by chance alone. A series of Monte Carlo simulations were carried out to estimate the false positive rate of h-ERE. One thousand nucleotide sequences 1 megabase (Mbp) long were generated randomly, governed by the empirical single nucleotide distribution of the human genome (UCSC hg17), and were run through h-ERE. The numbers of predicted binders divided by 1 Mbp was reported as the h-ERE false discovery rate per base pair. Taking a conservative estimate of the noise and extrapolating it, for the human genome (about 3 gigabases [Gbp]) about 33,000 (approximately 86%) were estimated to be false positives, and hence approximately 5000 ER-binding sites are present in the human genome.

Taken together, the convergence of these two analyses suggest that binding site motifs will be subject to statistical noise from random motif generation, but that a consistent number of bona fide binding sites, for the MCF-7 cells and under similar conditions as our experimentations, is likely to exist (about 5000).