We applied stringent statistical criteria to identify the regions bound by each factor with either a 1% or 25% expected false discovery rate (FDR) 11. While there was considerable variation in the number of bound regions identified for each factor, there were typically around 1,000 bound regions at a 1% FDR and 5,000 at a 25% FDR (Table 2). We ranked bound regions for each factor based on the maximum array hybridization intensity within the 500-bp "peak" window of maximal binding within each region.

We carried out an extensive series of controls and analyses to validate the antibodies and array data, and to ensure that our array intensities could be interpreted as a quantitative measure of relative transcription factor occupancy on each genomic region, that is, as a measure of the average numbers of molecules of a particular factor occupying each region (see 11 for further details).

For all but three factors, antisera were affinity-purified against recombinant versions of the target protein from which all regions of significant homology to other Drosophila proteins were removed. Where practical, antisera were independently purified against non-overlapping portions of the factor. When this was done, the ChIP/chip data from these different antisera gave strikingly similar array intensity patterns (for example, Figure 1), strong overlap between the bound regions identified (mean overlap = 91%; Table 2; Additional data file 1), and high correlation between peak window intensity scores (mean r = 0.79; Table 2), all of which strongly indicates that the antibodies significantly immunoprecipitate only the specific factor and that our ChIP/chip assay is very quantitatively reproducible. The specificity of the antibodies used is further confirmed by immunostaining experiments that show that they recognize proteins with the proper spatial and temporal pattern of expression (Additional data file 1).

We used two different methods to estimate FDRs, one based on precipitation with non-specific IgG, and the other based on statistical properties of data from the specific antibody alone. These estimates broadly agree (Additional data file 2). Our previously published quantitative PCR analysis of immunoprecipitated chromatin for regions randomly selected from the rank list of bound regions and also control BAC DNA 'spike in' experiments support the FDR estimates, suggest that the false negative rate is very low for all but the most poorly bound regions, and indicate that the array intensity signals correlate with the relative amounts of genomic DNA brought down in the immunoprecipitation 11.

The enrichment of factor recognition DNA sequences in ChIP/chip peaks shows a modest positive correlation with peak array intensity score. Importantly, this is seen even in the upper portion of the rank list where the percentages of false positives are too few to significantly influence the analysis (Figure 2; Additional data files 3 and 4) 11. While the presence of predicted binding sites is neither a necessary nor sufficient determinant of binding, this correlation strongly suggests that the number of factor molecules bound to a DNA region in vivo significantly affects the amount of each DNA region crosslinked and immunoprecipitated in the assay.

Finally, the relative array intensity scores from our formaldehyde crosslinking ChIP/chip experiments broadly agree with the relative density of factor binding detected by earlier Southern blot-based in vivo UV crosslinking 3031 (Additional data file 5). For BCD, FTZ and PRD the Pearson correlation coefficients are 0.79, 0.67, and 0.48, respectively, comparing the data from these two assays on the same genomic regions. This agreement is important because it argues that the measured relative signals in both assays are not powerfully influenced by differences in crosslinking efficiency to various DNAs, indirect crosslinking of proteins to DNA via intermediary proteins (which should not be detected by UV crosslinking), or differences in epitope accessibility during immunoprecipitation (which again should be much lower for UV crosslinking). Instead, the correspondence indicates that both these methods provide a reasonable estimate of the relative number of factor molecules in direct contact with different genomic regions in vivo.

Like the 6 previously examined A-P factors, the 15 newly studied regulators are detectably bound to thousands of genomic regions widely spread throughout the genome (Figure 3; Table 2; Additional data files 2, 6 and 7). The median number of 1% FDR bound regions detected by the antibody giving the most efficient immunoprecipitation for each of the 21 factors is 1,591 and the median number detected at the 25% FDR level is 7,145. At a 1% FDR, 23 Mb of the euchromatic genome is covered by a bound region for at least one factor, and of this, 9.8 Mb is within 250 bp of a ChIP/chip peak. At a 25% FDR, 32.2 Mb of the genome is within 250 bp of a ChIP/chip peak, which is 27% of the 118.4 Mb euchromatic genome. This binding is so extensive that, for each factor, on average, the transcription start sites of 20% of Drosophila genes lie within 5,000 bp of its 1% FDR ChIP/chip peaks, and for its 25% FDR peaks the equivalent figure is 54% of genes (Table 3).

For each factor, the numbers of regions bound at progressively lower array intensity signals increases near exponentially. At an array intensity of only 3- to 4-fold less than that of the most highly bound 20 to 30 regions, typically several thousand regions are bound by a protein (Figure 4; Additional data file 8). Because DNA amplification and array hybridization and imaging methods compress the measured differences in the amounts of DNA in an immunoprecipitation, the actual differences in transcription factor occupancy will be approximately three times greater than the differences in ChIP/chip peak intensity scores 11. Nevertheless, many genes are bound over a surprisingly narrow range of transcription factor occupancies.

Our earlier analyses of the six maternal and gap A-P factors showed that although these proteins bind to large number of regions, the most highly bound regions clearly differ in many regards from the more poorly bound, many of which may not be functional targets. Parallel analyses of the other 15 factors demonstrate the same trends.

First, for those factors for which a significant number of target CRMs are known, the few hundred most highly bound regions are enriched for these targets. Transgenic promoter, genetic, in vitro DNA binding and other data have identified a set of 44 CRMs as direct targets of subsets of the A-P early factors and 16 CRMs as direct targets of particular combinations of D-V regulators 82532. Figure 4 and Additional data file 8 show that the 500-bp ChIP/chip peaks that overlap CRMs known to be targets of at least some members of a given regulatory class are bound by all members of that class, on average, at higher levels than the majority of genomic regions at which these proteins are detected.

Second, the most highly bound regions, on average, are closer to genes with developmental control functions, whereas poorly bound regions are frequently closer to metabolic enzymes and other 'house keeping' genes (Figure 5; Additional data files 4 and 9). For most of the 21 factors, this enrichment reduces significantly between the top of the rank list and the 1% FDR threshold, which, if our FDR estimates are good, rules out the possibility that the presence of false positives has influenced this result.

Third, for the majority of factors the more highly bound regions tend to be closest to genes that are transcribed at the blastoderm stage and whose spatial expression is patterned at this stage (Figure 6; Additional data files 4 and 10). Poorly bound regions, in contrast, are closest to genes that are transcriptionally inactive or not patterned at this stage. For a minority of factors this trend is not as pronounced. However, this is probably because the regions bound highly by these proteins are already further away from the transcription start site of their known or likely target genes than are those of other factors (for example, Runt (RUN) 1 in Figure 6; and Sloppy paired (SLP)1 in Additional data file 10).

Fourth, poorly bound regions for a subset of factors show a surprising preference to be located in protein coding regions. This is particularly striking for FTZ, Knirps (KNI), Mad (MAD), RUN and SNA, but a number of other factors show a less dramatic but similar trend (see regions between the 1% and 25% FDR thresholds in Figure 7 and Additional data file 11).

Fifth, for those bound regions in intergenic and intronic sequences (that is, in non-protein coding sequences) the more highly bound are significantly more conserved than those poorly bound (Figure 8; Additional data files 4 and 12). For most factors, however, their specific recognition sequences are not particularly more conserved than the remaining portion of the 500-bp peak windows (11 and our unpublished data). Thus, for most factors, it cannot be concluded from this analysis alone that recognition sequences are being conserved because they are functional targets. But it can be concluded that the more highly bound regions likely are, on average, more evolutionarily constrained function than poorly bound regions.

Taking all of these five analyses into account, the few hundred most highly bound regions have characteristics of likely functional targets of the early embryo network. Although some poorly bound regions are also likely to be functional targets at this time, including ones weakly modulating transcription of housekeeping genes (for example, 22), many do not appear to be classical CRMs that drive transcription in the blastoderm. A minority do become more highly bound in the later embryo and may be active then (our unpublished data), but the binding to many others we feel is likely to be non-functional, including that to most of those in protein coding regions.

Our analysis contrasts with the predominant qualitative interpretation of in vivo crosslinking data by other groups studying animal regulators 323334353637383940414243444546. Many of these groups have also shown that factors bind to a large number of genomic regions. They have not, however, noted the many differences between highly bound and poorly bound regions shown in Figures 4 to 8. In addition, with only a few exceptions 434446, they have not seriously considered the possibility that some portion of the binding detected is non-functional. We suspect that similar correlations between levels of factor occupancy and likely function of bound regions will be found for other factors once quantitative differences amongst bound regions are considered.

Another striking feature of our in vivo DNA binding data is that there is considerable overlap in the genomic regions bound by the 21 factors (Figures 3), even though they belong to 11 DNA binding domain families and multiple regulatory classes, often act via distinct CRMs, and clearly specify distinct developmental fates. To quantify this overlap, we scored for each protein the percent of peaks that are overlapped by a 1% FDR region for each factor in turn (Figure 9a, b; Additional data file 13). This analysis shows, for example, that of the 300 peaks most highly bound by the A-P early regulator BCD, between 6% and 100% are co-bound by the other 20 factors, some of the highest overlap (>94%) being with the D-V regulators Medea (MED), Dorsal (DL) and TWI (Figure 9a, top row). Peaks bound more poorly are overlapped to a lesser degree, but there is still considerable cross-binding to these regions (Figure 9b; unpublished data).

To calculate the probability that this extensive co-binding occurs by chance, we used the Genome Structure Correction (GSC) statistic 43, which is a conservative measure that takes into account the complex and often tightly clustered organization of bound regions across the genome. For the great majority of the pair-wise co-binding shown in Figures 9a, b, these probabilities have Bonferroni corrected P-values < 0.05 (all instances with z scores ≥4 in Figure 9c, d) and, thus, the overlap is highly unlikely to have occurred by chance. With such extensive co-binding, it is not surprising that some regions are bound by many factors. Averaged over all regulators, 88% of their top 300 peak windows are bound by 8 or more factors and 40% are bound by 15 or more factors (Additional data file 13).

Several recent in vivo crosslinking studies have also noted significant overlap in binding between some sequence-specific factors in animals 3234374446. In these other cases, however, the overlapping factors are known to have related functions and, thus, the co-binding is less surprising. Work using the DamID method showed a high overlap in binding when transcription factors with different functions and specificities were ectopically expressed in tissue culture cells 47, and it was suggested that these binding 'hotspots' were non-functional storage sites. In contrast to these other studies, we have found overlapping binding for a larger number of regulators, many of which are well characterized as having distinct biological and transcriptional regulatory specificities. The binding we have measured is for endogenous factors, and the greatest overlap in binding is at known and probable functional targets. Thus, it does not seem that overlapping patterns of binding reflect either shared functions or a lack of function. Instead, we must ask how the undoubtedly distinct specificities of the blastoderm factors arise despite the overlap.

To address this question, we first looked in detail at the pattern of binding on the CRMs of two well-studied target genes. The eve gene is expressed in a seven stripe pair-rule pattern along the A-P axis and contains four stripe CRMs that are known targets of the A-P early factors (Figure 10, S3/7, S2, S4/6 and S1/5) and a later activated autoregulatory CRM thought to be a target of the A-P pair rule factors EVE and PRD (Figure 10, Auto) 242829. The sna gene is expressed in a ventral stripe of expression and has two known CRMs that are targets of the D-V regulators TWI or DL (Figure 10, AE and VA) 48. Consistent with the analysis in Figure 9, there is a high co-binding of members of all three major regulatory classes to each of these CRMs at a 1% FDR (Figure 10; Additional data file 14), and even more extensive co-binding is seen when lower level interactions detected at a 25% FDR and in in vivo UV crosslinking experiments are taken into account 31 and our unpublished data). However, the factors show quantitative preferences in binding to the CRMs that broadly correlates with their expected function: A-P early factors most strongly occupy the four eve stripe CRMs, A-P pair rule factors most strongly occupy the eve autoregulatory element, and the D-V factors TWI and SNA most strongly occupy the two sna CRMs (Figure 10). Thus, differences in the levels of occupancy on common genomic regions could be significant determinants of regulatory specificity.

The fact that the higher levels of binding better reflect expectations based on earlier molecular genetic experiments, however, does not necessarily indicate that only these interactions are functional. For example, recent studies using image analysis of three-dimensional cellular resolution data have shown that there are modest quantitative affects of D-V regulators on eve expression and of A-P regulators on sna expression (Figure 11) 91011, which could be due to the low-level occupancy of D-V regulators on eve and of A-P regulators on sna. Indeed, these quantitative methods show that the expression patterns of most genes in the blastoderm are much more complex than early low-resolution expression data implied 91012. Thus, the regulation of blastoderm genes may involve input from a broader range of factors than first assumed, with the degree of transcriptional regulation correlating with the degree of factor binding.

Other work, however, cautions against assuming that all of the lower level interactions shown in Figure 10 result in transcriptional regulation. In the case of the binding of A-P gap factors to the eve autoregulatory element, transgenic promoter analysis indicates that this binding is not sufficient to detectably activate this CRM in early stage 5 embryos 24. A similar argument can be made for binding of A-P pair rule factors to the eve stripe CRMs 24272849. In these cases, either this lower level binding is non-functional or it plays an augmentary role only in the context of multiple promoter elements. It is not sufficient for regulation on its own.

To more fully explore if there is a correlation between the level of factor occupancy on common sequences and functional specificity, we next compared the binding of all 21 factors on the 44 A-P early CRMs and 16 D-V CRMs described earlier. (There are too few known A-P pair rule CRMs to analyze in this way.) While most of these CRMs are each bound above the 1% FDR threshold by members of all three of the major regulatory classes (Figure 12a), the normalized levels of factor occupancy can be seen to broadly meet expectations (Figure 12b): the A-P early factors bind more highly to A-P early CRMs, the D-V factors mostly bind more highly to D-V CRMs, and the pair rule factors bind at lower levels to all of these CRMs than they do to other regions of the genome. There are a few instances where relatively high levels of binding are found to CRMs initially identified as targets of another regulatory class, but these likely reflect the fact that some of these CRMs show strong patterning along both the A-P and D-V axes (for example, 32). Averaged over all interactions for members of each regulatory class, the levels of binding of each class match the general expectations for their specificity (Bonferroni corrected Mann Whitney test P-values all < 1 × 10^-8; Additional data file 13).

Informative as the above analyses are, however, they are restricted to previously identified CRMs. These CRMs were identified experimentally using criteria that could well have excluded some types of functional targets. To explore if levels of occupancy correlate with functional specificity more widely, therefore, we examined binding to genomic regions without regard to any published information on which sequences different factors might act on. In the absence of any prior knowledge, we exploited our observation that, on known CRMs, members of a regulatory class have more similar specificities than members of different classes and used this to provide an expectation of specificity elsewhere in the genome. We used two measures to compare binding for factors within and between classes (Figure 13).

First, we used the previously described GSC statistic for the likelihood that two factors bind the same regions more frequently than expected by chance, but this time focusing only on the overlap between highly bound regions. All 441 pair-wise comparisons of overlap were computed between the 300 regions bound most highly by each factor (Figure 13a) and separately between the 300 next most highly bound regions (Figure 13b). For both cohorts, not surprisingly given our earlier analysis, co binding between most pair-wise combinations of factors occurs far more frequently than expected by chance, even where the proteins belong to different regulatory classes (z scores ≥4 in Figure 13a, b). However, for the top 300 bound regions, there is an obvious further preferential overlap among A-P early regulators as well as a moderate preference among the A-P pair rule factors and the D-V factors (Bonferroni corrected Mann Whitney tests suggest that, taken collectively, the preferential co-binding among A-P early regulators is highly significant (P < 9 × 10^-15), while that among the A-P pair rule factors and the D-V factors is moderately significant (P < 2 × 10^-3) for both; Additional data file 13). The next most highly bound cohort shows reduced preferential co-binding within regulatory classes, with only that among A-P early regulators being significant (P = 7 × 10^-9; Figure 13b; Additional data file 13).

Second, because the above measure only partially takes into account the different levels of occupancy on each bound region, we sought a measure that better captures this information. Scatter plots show that while ChIP/chip scores from experiments using antibodies to distinct portions of the same factor are highly correlated, pair-wise comparisons between factors reveal marked differences in scores, suggesting that correlation coefficients calculated in this way would be a useful measure of binding specificity (Figure 14). Therefore, we computed Pearson correlation coefficients for all pair-wise comparisons between factors for the most highly bound 300 regions and separately for the next most highly bound 300 regions (Figure 13c, d). Visual inspection shows that A-P early and D-V regulators generally show higher similarity in binding with members of their own regulatory class than they do with other factors. Similarly, the highest correlations for the A-P pair rule factors FTZ, HRY, RUN and SLP1 are with other pair rule proteins, though in this case preferences are shared with specific proteins rather than class-wide. Bonferroni corrected Mann Whitney tests indicate that correlation coefficients generally show more significant distinctions in binding preferences between the three regulatory classes than the z score measure, both taken collectively (A-P early P < 10^-15, A-P pair rule P = 1 × 10^-6, D-V P = 1 × 10^-9), and on a per factor basis (Additional data file 13). They even detect moderate discrimination between the classes in the 301 to 600 cohort (A-P early P = 1 × 10^-3, A-P pair rule P = 1 × 10^-3, D-V P = 2 × 10^-2).

Thus, across a broad array of mostly uncharacterized genomic regions the levels of binding of transcription factors correlate with the expectation that factors with more similar functions show more similar binding specificity. Consistent with our previous observation that highly bound regions appear more functionally significant, the distinctions in binding preferences between regulatory classes is larger on the most highly bound regions. Just as on the known CRMs, however, the distinctions between the different classes are relatively modest, suggesting that the regulatory specificity of transcription factors in general may be fuzzier than widely realized and perhaps also suggesting a role for post-DNA-binding events to increase the distinctions between factors.

All of the preceding analyses consider binding to short genomic regions. The target genes of blastoderm factors, however, are often found associated with several such regions (for example, Figure 10). Thus, while the above analyses establish that regulators show quantitative preferences for binding to individual genomic regions, they do not establish if they exhibit preferences for different genes.

To determine whether these factors are targeting distinct sets of genes, for each bound region we identified the Gene Ontology (GO) term associated with the gene whose transcription start site is closest to the peak of binding. The enrichment of the GO terms associated with the 300 most highly bound peaks and the next most highly bound cohorts of peaks were then plotted as heat maps (Figure 15a, b). This shows that there are clear differences between factors as to which GO terms are associated with their top 300 peaks. In addition, there is a broad within regulatory class preference for which types of gene transcription factors bind to most strongly. More fine-grained similarities between subsets of factors within the major regulatory classes are also apparent. Thus, the quantitative preferences apparent at the level of individual bound regions must extend to some degree to the level of the clusters of regions associated with each gene and with different gene types.

The GO terms associated with the 301 to 600 ranked peaks show much less difference between each factor and regulatory class, consistent again with these less highly bound regions playing a lesser role in determining biological specificity and function.

Some of the types of gene preferentially bound by the regulatory classes readily fit expectations; for example, the strong association of genes involved in A-P axis specification, patterning by pair rule genes, and trunk and head segmentation with A-P early and A-P pair rule factors. Others are unexpected, such as the preference of D-V regulators for a series of GO terms related to eye development. Most likely the differences between factors revealed in these heat maps reflect differences due to target genes that are strongly patterned along the A-P axis versus those strongly patterned along the D-V axis. Because important effectors of blastoderm regulators' functions are patterned along both body axes, because the early factors both activate or repress target genes, and because GO terms imperfectly capture and categorize the biological function of each gene, this analysis does not provide a complete description of the different specificities of each factor at the target gene level.

What mechanism, though, drives the extraordinarily extensive, overlapping pattern of binding? We speculate that the pattern is a natural consequence of these factors' intrinsic DNA binding specificities (as measured in vitro), the relatively high concentrations at which they are expressed in nuclei in which they are active, chromatin structure, and the law of mass action.

Most animal transcription factors recognize short degenerate DNA sequences that occur frequently throughout the length of most genes 156. It has long been proposed on thermodynamic grounds that the majority of transcription factor molecules would be bound to DNA in the nucleus, rather than be free in solution 505152. In eukaryotes, only a subset of the genome is fully accessible to sequence-specific DNA binding factors because of the presence of nucleosomes 5354555657585960. Any several hundred base-pair segment of such accessible DNA will likely contain moderate to high affinity recognition sequences for a large proportion of transcription factors. Since many of the blastoderm factors are present at concentrations of many tens of thousand of molecules per cell 3061, they may well be able to significantly occupy these sites, generating a highly overlapping pattern of binding focused at open chromatin regions.

In addition to the independent interactions of transcription factors with their target DNA sequences in open chromatin, some of the overlap in binding may likely arises from protein-protein interactions in which a factor associates with an accessible region as a result of direct interactions with protein molecules bound to the region. Such indirect binding - whether between transcription factors, or mediated by the large numbers of co-factors associated with CRMs - could explain the frequent absence of high affinity DNA recognition sequences for proteins bound to a given region. Like protein-DNA interactions, protein-protein interactions are enhanced when protein concentrations are high and, thus, in vivo could also mediate low level binding at non-functional sites.

How is it possible to have so much binding that has no or little effect on transcription? Natural selection clearly acts on CRMs to preserve the proper number, arrangement and affinity of recognition sequences for whichever factors are needed for its activity. There is also evidence that selection acts against sites that might interfere with activity 62. Purifying selection will remove any 'spurious' binding that interferes with the proper expression of a gene. But weak binding that has only a small or no affect on transcription could well be tolerated in many cases. Just as there is a quantitative continuum of binding, there may also be a continuum of effects on transcription, and ultimately on phenotype.

Embryos were collected in population cages for 1 hour, and then allowed to develop to the required stage before being harvested and fixed with formaldehyde 11. Embryo aging times were determined based on the transcription factor analyzed: for A-P maternal, gap, and terminal factors, as well as D-V maternal and a subset of D-V zygotic factors, including SNA and TWI, the embryos used were 2 to 3 hours old (mainly between late stage 4 and early stage 5), while for A-P pair rule, and the D-V zygotic factors, MED, MAD, and Schnurri (SHN), the embryos were 2.5 to 3.5 hours old (mainly at mid- to late stage 5). The chromatin used for immunoprecipation was isolated from the fixed embryos by CsCl gradient ultracentrifugation and then fragmented to an average size of about 700 bp.

All of the antibodies used were immunoaffinity purified from rabbit antiserum. The two anti-PRD antibodies, PRD 1 and PRD 2, were available from a previous study 31. The MAD and MED antisera were a generous gift from L Raftery 63, the RUN antiserum from E Wieschaus, and the TFIIB antibody from R Tjian. For other factors, antibodies were produced in rabbits immunized with recombinant His-tagged fusion proteins expressed and purified in Escherichia coli using the Invitrogen Gateway system. Rabitts were immunized with either the full length protein (Dichaete (D), HRY, SLP1, Daughterless (DA), DL, SNA, and TWI) or portions of the protein (TLL amino acids 110 to 259, SHN amino acids 1,617 to 1,750, and SHN amino acids 2115 to 2,279). Immunoaffinity purifications were performed using E. coli-expressed purified recombinant His-tagged proteins. The amino acid sequences used (listed in Table 2) were chosen to exclude regions with any significant homology to other Drosophila proteins, as previously described 11. Additional results demonstrating the specificity of the antibodies are provided in Additional data file 1.

Chromatin was immunoprecipitated and the resulting DNA was amplified and hybridized to Affymetrix Drosophila Genomic Tiling Arrays as previously described 11. For each antibody, duplicate immunoprecipitations were performed along with duplicate control IgG immunoprecipitations. These were each hybridized to separate arrays as were duplicate input DNA samples. All raw microarray data (CEL files) have been deposited at Array Express [E-TABM-736] 64. In addition, these and more processed forms of the data are available from the BDTNP's public web site, together with more detailed information about antibodies used and so on 65.

The data from the complete set of six arrays from each ChIP/chip experiment were processed using TiMAT 66 as described previously 11 to derive peak window locations, bound regions, 1% and 25% FDR thresholds for both the IgG and Symmetric null tests, and so on. Bound regions were associated with the gene (from release 4.3 of the D. melanogaster genome) whose 5' end was closest to the array intensity peak in the bound region. To identify the closest transcribed gene, the subset of release 4.3 annotations that completely overlap regions bound by RNA polymerase II in our ChIP-chip experiments was used.

Relative percentages of UV crosslinking to defined restriction fragments 31 and the corresponding mean oligo ChIP/chip scores of the same genomic regions are plotted as scatter plots in Additional data file 5. Pearson correlation coefficients were calculated for each plot.

Enrichment of recognition sequences, GO terms, distance to transcribed genes, genomic locations and phastcons scores were determined essentially as described in 11. A statistical analysis of the significance of these plots is presented in Additional data file 4.

In Figure 4, peaks were distributed by the mean ChIP/chip peak scores in the 500-bp peak window. For A-P early factors, a peak was associated with A-P early CRMs if the peak single nucleotide position was contained within one of the CRMs extended by 250-bp flanking regions. For D-V factors, a peak was associated with D-V CRMs in the same way.

Overlap of bound regions between two transcription factors in Figure 9 was measured by the percentage of single nucleotide peak locations of one factor contained in 1% FDR bound regions of the other factor. The top 300 peaks (1-300) and separately peaks 301 to 600 of each factor were used in the analysis. Overlap of one factor by multiple factors was measured by the percentage of peaks of that factor contained in 1% FDR bound regions of a defined number of other factors (Additional data file 13).

To calculate the liklihood z score that overlap occurs by chance (Figures 9c, d and 13a, b), z-scores were computed using the GSC statistics 43. A null distribution of feature-feature overlap was computed by selecting pair-wise block samples from the genome, and in each block in the pair the annotations of one of the two features of interest were swapped to yield artificial overlaps. The resulting null distribution is more realistic than that derived from other methods in that the complex and often tightly clustered organization of each feature across the genome is preserved, resulting in a much larger (conservative) estimate of standard deviation than derived via other methods. Like most methods, including feature start-site randomization, this null is Gaussian, and hence after centering, the only quantity that needs to be estimated is precisely the standard deviation.

In Figure 12a a CRM is defined as being bound by a transcription factor if it was overlapped by at least 300 bp by one of the factor's 1% FDR bound regions, or for CRMs less than 300 bp long, if the CRM was completely overlapped by a 1% FDR region. In Figure 12b, the binding intensity of a transcription factor to a CRM is defined by the highest 675-bp smoothed window score in the CRM for that factor, without regard to FDR threshold. The window scores for each factor were placed on the same scale by setting the highest 675-bp window contained on the whole array data to 10.

In Figure 13c, d, for each transcription factor, the score associated with each 500-bp peak window was derived from the mean score of oligos in the window. The scores for the equivalent 500-bp windows for each of the other 20 factors were then derived from the mean oligo scores from those datasets, without regard to any FDR threshold. Scores for each pair-wise comparison of factors were used to calculate the Pearson correlation between the top 300 bound regions (1-300) and separately for regions from 301 to 600 on the ChIP/chip rank list. Because the original data for the PRD 1 antibody was derived from a different array scanner than that used for the other factors and because we found that a subtle scaling difference between the two scanners affected the correlation coefficients, the PRD 1 data used in all of Figure 13 were from a replica set (PRD 1*) that used the same Affymetrix G7 scanner used to derive data for the other factors.

Mann-Whitney tests were applied to the binding intensity data of transcription factors to CRMs (Figure 12b), overlap GSC Z scores between factors (Figure 13a, b), and Pearson correlation of intensity scores of peak windows between factors (Figure 13c, d) and are reported in Additional data file 13. Each data set was divided into two categories by factor regulatory classes. The Mann-Whitney test was one-sided, with the null hypothesis that the two categories of data followed the same distribution. Bonferonni corrected values are provided where stated.

In Figure 15, each bound region is associated with the 'biological process' GO term for the gene whose transcription start site was closest to the array intensity peak in the bound region. The non-redundant set of the 7 most enriched GO terms associated with the top 300 bound regions of each factor were used in the analysis. Negative logged probabilities from a hypergeometric distribution were used to measure the association of the top 1 to 300 and 301 to 600 bound regions of each factor with a GO term. The scores of different factors were put on the same scale by setting the most enriched value to 10.